Theo spends a full day inside Claude Opus 5, pits it against Fable 5 and GPT-5.6-Sol on benchmarks and real coding tasks, and argues the cheaper, weirder model just won his default slot.
Claude Opus 5 beats Fable 5 and GPT-5.6-Sol on most published coding benchmarks while listed at roughly half Fable's price, and after a day of real use its middle-ground diligence and instruction-following — not the benchmarks — is the actual reason to make it a daily default.
Who This Is For
Read if. Skip if.
READ IF YOU ARE…
You pay API or subscription prices for Claude models and want to know if switching your default from Fable 5 to Opus 5 saves real money, not just sticker-price money.
You regularly burn through a weekly usage limit on a Claude subscription and want a model that stretches further without losing capability.
You're deciding between Anthropic-, OpenAI-, and other frontier-style coding models and want a working mental model of how each one actually behaves, not just a benchmark table.
You're on an enterprise plan that legally can't use a model with default request logging and need to know which frontier model actually clears that bar.
SKIP IF…
You've never used an AI coding model and want a plain feature list, not a model-personality comparison.
You're looking for a reproducible benchmark study — this is one developer's day-one impressions and anecdotes, not a controlled evaluation.
TL;DR
The full version, fast.
Opus 5 leads most of Anthropic's own coding and knowledge-work benchmarks against Fable 5 and GPT-5.6-Sol, and its list price is about half of Fable's. But Opus 5 burns more tokens per task than Fable, so the real-world savings land closer to 20-25%, not 50% — except on Claude subscriptions, where Opus can use 100% of the weekly limit versus Fable's 50%, which is the actual budget unlock. After a day of testing plans, reviews, and real PRs, the video's case for Opus rests on three things: subscription economics, zero data retention (unlike Fable/Mythos, which Anthropic logs on every request), and a genuine middle-ground personality — more diligent and instruction-following than GPT-5.6-Sol's 'do exactly what's asked and nothing more,' less prone than Fable to quietly working around a hard problem instead of solving it.
Free for members
Chat with this breakdown — free.
Sign in and you get 23 free chat messages on us — ask for the hook, quote a framework, find the exact transcript moment, generate a markdown action plan. Bring your own key when you want unlimited.
Opus 5 leads most benchmarks Theo could find against Fable 5 and GPT-5.6-Sol, and it's cheaper than both — a result he calls confusing and spoils the ending on immediately.
00:50 – 03:37
02 · Sponsor: Sent DM
Ad read for Sent, a platform that routes messages to users over SMS/RCS/WhatsApp based on their preference, with an MCP server so agents can send messages and look up numbers directly.
03:37 – 06:42
03 · The official Anthropic announcement
Reads Anthropic's launch post: Opus 5 is 'designed to be used every day,' is now the default model on Claude Max and the strongest on Claude Pro (with Fable removed from the $20 tier), and posts strong early benchmark numbers.
06:42 – 10:50
04 · Benchmark deep dive
Walks Frontier Code, GDPval, ARC-AGI-3, BrowseComp, Humanity's Last Exam, OS World, DeepSWE, and Business/Legal/Health/Bio benchmarks — Opus 5 wins most coding and business evals, trails Fable 5 on pure-knowledge and Health Bench, and trails GPT-5.6-Sol on DeepSWE.
10:50 – 14:44
05 · Real cost vs. the sticker price
Opus 5's $5/$25-per-million pricing looks like half of Fable's $10/$50, but because Opus uses more tokens per task, the real per-task cost gap is closer to 20-25%, not 50%.
14:44 – 18:36
06 · Models judging each other's plans
Theo has Opus 5 and Fable 5 plan the same feature and blind-review each other's plan — both rate the other's plan higher than their own — then has GPT-5.6-Sol independently judge both plans without knowing which was which, scoring Opus's plan well ahead of Fable's.
18:36 – 22:31
07 · Alignment, distillation, and the Ars Technica take-down
Covers Opus 5's alignment scores, explains it was distilled from Anthropic's larger Mythos model via the 'bowl of food' metaphor, then corrects an Ars Technica article's claim about token efficiency and model routers.
22:31 – 26:25
08 · Vulnerabilities and usage accounting
Opus 5 can still find security vulnerabilities but is much worse at exploiting them than Mythos/Fable; screenshots of real Claude account quota dashboards set up the usage-limit argument that follows.
26:25 – 30:18
09 · Reason #1: cost and limits
Claude subscriptions only allot Fable 5 half the weekly usage limit, but Opus 5 gets the full 100% — in practice, a full day's coding work used about 12% of a weekly Opus quota versus roughly 1.5 weekly Fable quotas burned in a comparable stretch.
30:18 – 34:12
10 · Sponsor: Convex, plus Reason #2 (ZDR)
Live demo of Convex's Claude Code plugin building a synced Kanban app, then the second reason to use Opus: it isn't subject to the request logging that rules Fable/Mythos out for enterprise and legally-restricted use cases.
34:12 – 38:05
11 · Reason #3: diligence and instruction-following
Opus 5 is described as the first Anthropic model that reliably does what it's told without over-guessing intent, landing between GPT-5.6-Sol's blunt literalism and Fable 5's tendency to cleverly work around hard problems — illustrated by a live incident where Opus repeatedly opens a browser nobody asked for.
38:05 – 41:26
12 · Hallucination scores and the full comparison
Opus 5's AA Omniscience score shows real improvement over Opus 4.8, though Fable 5 still hallucinates least of the three; builds out a full whiteboard comparing GPT-5.6-Sol, Fable 5, and Opus 5 side by side.
41:26 – 44:29
13 · Final verdict and sign-off
Closes the browser-incident story, delivers the final recommendation (default to Opus 5, keep GPT-5.6-Sol for throwaway work, reach for Fable 5 for frontend/niche knowledge), and asks viewers to test the models themselves rather than take his word for it.
Atomic Insights
Lines worth screenshotting.
Opus 5 is listed at $5/$25 per million input/output tokens versus Fable 5's $10/$50 — literally half price on paper.
Because Opus 5 uses more tokens per task (~37k) than Fable 5 (~33k) on Artificial Analysis, the real-world cost gap is closer to 20-25%, not 50%.
Claude subscription plans cap Fable 5 at 50% of the weekly usage limit but let Opus 5 use 100% of it — switching alone roughly doubles how much work a subscription covers.
On Frontier Code, which grades a patch's mergeability and adherence to a codebase's own conventions rather than passing unit tests, Opus 5 scored 43.3% against a prior high of 34.4%.
ARC-AGI-3 penalizes models for needing more reasoning steps than a human would; no model scored above 1% at its launch, and Opus 5 is already near 30%.
In a blind cross-review test, GPT-5.6-Sol independently scored Opus 5's plan 8.3 out of 10 versus Fable 5's plan at 6.0.
Fable 5 and Mythos are audited and logged by Anthropic on every request even on enterprise plans; Opus 5 doesn't carry that same zero-data-retention restriction.
Opus 5 is still good at finding vulnerabilities in code but meaningfully worse than Mythos/Fable at actually exploiting them.
Opus 5 scores 31 on the AA Omniscience bench, an improvement over Opus 4.8's tendency to confidently make up answers, though Fable 5 still hallucinates the least of the three.
Anthropic's own team published new 'context engineering' guidance after Opus 5 shipped, saying users no longer need to over-specify instructions in system prompts and CLAUDE.md files the way earlier Claude models required.
In one real test, a full day of coding work used about 12% of a weekly Opus 5 quota, versus roughly 1.5x the weekly Fable 5 quota burned in a comparable day.
Even in 'auto' permissions mode, Opus 5 opened a non-default browser three separate times uninvited to run a visual check nobody asked for.
Takeaway
Three concrete reasons carry the case for Opus 5, and none of them is the benchmark score.
WHAT TO LEARN
Opus 5's benchmark wins matter less than its subscription economics, its lack of request logging, and a genuine middle-ground behavior between an overly literal model and an overly clever one.
01Cold open
Opus 5 tops most published benchmarks against both Fable 5 and GPT-5.6-Sol while costing under half of Fable's list price.
The video's core claim, stated up front: Opus 5 is likely the only model most people need for daily coding work.
03The official Anthropic announcement
On Frontier Code (called 'Frontier Bench' in the video before a mid-recording correction), which grades a patch's mergeability and adherence to a codebase's own conventions rather than just passing unit tests, Opus 5 scored 43.3% versus a prior high of 34.4%.
ARC-AGI-3, redesigned to penalize models that need more reasoning steps than a human would, went from no model scoring above 1% at launch to Opus 5 already near 30%.
04Benchmark deep dive
Opus 5 essentially ties GPT-5.6-Sol on agentic search (BrowseComp: 90.8 vs 90.4) but trails Fable 5 when a task is pure knowledge recall without tool use.
On coding-adjacent evals Opus 5 matches or edges out Fable 5, but results vary wildly outside coding — it's behind Fable on Health Bench, ahead on Bio Bench, and behind GPT-5.6-Sol on the more realistic DeepSWE benchmark.
OS World's claim that Opus is the best computer-use model shouldn't be trusted — GPT-5.6-Sol is meaningfully better at computer use in practice.
05Real cost vs. the sticker price
Opus 5 is listed at $5/$25 per million input/output tokens versus Fable 5's $10/$50 — literally half price on the sheet.
But Opus 5 uses more tokens per task than Fable (about 37k vs Fable's 33k on Artificial Analysis), so the real-world discount lands closer to 20-25% (about $2.03 vs $2.75 per task), not the advertised 50%.
06Models judging each other's plans
In a same-prompt test, Opus 5 and Fable 5 each planned the same feature and blind-reviewed each other's plan — both rated the other's plan higher than their own, with Fable perceiving the bigger gap in Opus's favor.
A third model, GPT-5.6-Sol, independently reviewed both plans without knowing which was which and scored Opus's plan 8.3 versus Fable's 6.0.
Peter Gostev's framing, adopted as the video's throughline: Fable is a 'wise owl' — thoughtful and well-spoken but sometimes evasive — while GPT-5.6-Sol is a 'rottweiler' that grabs a problem and doesn't let go; Opus 5 sits in between.
07Alignment, distillation, and the Ars Technica take-down
Anthropic distilled Opus 5 from Mythos, its more capable but riskier internal model, by filtering capabilities down to a smaller model — like emptying a bowl of mixed food down to just the parts you actually wanted.
Opus 5 is described as Anthropic's most aligned model yet: lowest rate of deceptive behavior, least susceptible to being tricked into misuse, and it does not advance the frontier on risky dual-use capabilities.
An Ars Technica piece implied cheaper models save tokens via routing, but the smartest model here (GPT-5.6-Sol) actually uses the fewest tokens, and Opus 5 uses more tokens than Fable, not fewer.
08Vulnerabilities and usage accounting
Opus 5 is still good at finding vulnerabilities in code but meaningfully worse at exploiting them than Mythos/Fable were — the safety filtering held even though bug-finding skill didn't get lobotomized.
Real account screenshots show a full day of coding work cost about 12% of a weekly Opus quota, versus roughly 1.5x the weekly Fable quota burned in a comparable day.
09Reason #1: cost and limits
Claude subscription plans only allocate 50% of the weekly usage limit to Fable 5, but Opus 5 can use 100% of the same limit — switching alone roughly doubles how much work a subscription covers.
Because Opus 5 is cheaper to host and has less demand right now, Anthropic can subsidize it more generously — a pricing situation the video expects won't last forever.
10Sponsor: Convex, plus Reason #2 (ZDR)
Fable 5 and Mythos are audited and logged by Anthropic on every request even on enterprise plans, which rules them out for companies that legally cannot allow that; Opus 5 doesn't carry that same restriction.
11Reason #3: diligence and instruction-following
Opus 5 is the first Anthropic model that reliably does what it's told instead of guessing at unstated intent — significant enough that Anthropic's own team published new 'context engineering' guidance saying users no longer need to over-specify instructions in CLAUDE.md files and skills.
It lands in a genuine middle ground: more thorough and code-quality-conscious than GPT-5.6-Sol (which will write excessive code just to close the ticket), less prone than Fable to quietly working around a problem instead of solving it.
The tradeoff for that diligence is speed — Opus 5 takes noticeably longer per task because it double- and triple-checks its own work before finishing.
12Hallucination scores and the full comparison
Opus 5 scores 31 on the AA Omniscience bench (measuring whether a model avoids confidently making up an answer it doesn't know), a real improvement over Opus 4.8, though Fable 5 is still the least likely of the three to hallucinate.
Final three-way cheat sheet from the video: GPT-5.6-Sol is a 'tool' that doesn't care about code quality but is efficient and thorough; Fable 5 writes code you actually want to look at and knows more, especially on frontend and niche topics; Opus 5 is the pick when you don't want to reason about any of this.
13Final verdict and sign-off
Even on 'auto' permissions mode, Opus 5 opened a non-default browser three separate times uninvited to run a visual check the user never asked for — the smaller-model failure mode is still annoying interruptions, versus GPT-5.6-Sol's rarer but more destructive failures.
Closing recommendation: default to Opus 5 for most coding work, keep GPT-5.6-Sol for tasks you don't intend to merge or as a second-opinion reviewer, and reach for Fable 5 specifically when frontend polish or niche/obscure knowledge is the bottleneck.
The video's actual ask isn't 'trust me' — it's to run the same prompt across models yourself and have them review each other's output, since that cross-review is what surfaced the real differences here.
Glossary
Terms worth knowing.
Frontier Code
A coding benchmark that grades whether a model's patch is mergeable and follows a codebase's own conventions, rather than just checking if it passes unit tests. Referred to as 'Frontier Bench' partway through the video before the mix-up is corrected on screen.
GDPval / GDPval-AA
A benchmark that scores models on realistic, specific economic and business tasks rather than abstract reasoning puzzles.
ARC-AGI-3
The third version of the ARC-AGI pattern-recognition benchmark, redesigned to penalize a model that solves a problem correctly but takes far more reasoning steps than a human would need.
BrowseComp
A benchmark measuring how well a model can autonomously search the web and browse pages to complete an agentic research task.
Humanity's Last Exam (HLE)
A multidisciplinary reasoning benchmark meant to be near the edge of what current models can answer correctly.
DeepSWE
A realistic software-engineering benchmark considered more representative of real coding work than benchmarks that just measure whether a patch merges cleanly.
Zero Data Retention (ZDR)
A policy where a provider does not log or retain the contents of a request after processing it, which some companies are legally required to have before they can use a given AI model at all.
AA Omniscience Bench
An Artificial Analysis benchmark measuring not just whether a model answers correctly, but whether it avoids confidently stating a wrong answer when it doesn't actually know.
Distillation
A training technique where a smaller model is trained to reproduce the useful behaviors of a larger, more capable model, while leaving out capabilities the trainers don't want to transfer.
Mythos
Anthropic's larger, more capable but higher-risk internal model, described in the video as the source model that Opus 5 was distilled from to keep its useful capabilities while filtering out riskier ones.
31:40link"The New Rules of Context Engineering for Claude 5 Models" (Thoric's article)
Quotables
Lines you could clip.
00:45
“This is probably the only model you need.”
the video's whole thesis in one line, delivered as a cold-open spoiler→ TikTok hook↗ Tweet quote
23:40
“Get your shit together, Anthropic. Seriously, this model is too good to be gimped by your shit.”
raw frustration at a real workflow-breaking bug, high energy→ IG reel cold open↗ Tweet quote
19:20
“How do I convince the world that the smarter models don't use more tokens? The smartest models use less.”
a contrarian, quotable correction of a common misconception→ newsletter pull-quote↗ Tweet quote
42:20
“Use Opus because you don't wanna think about all this bullshit.”
the entire recommendation distilled into one blunt line→ TikTok hook↗ Tweet quote
36:50
“Stop opening my fucking browser, and stop it from running.”
the funniest concrete failure moment in the video, very shareable→ IG reel cold open↗ Tweet quote
The Script
Word for word.
Read-along
Don't just watch it. Burn it in.
See every word as it's spoken — crank it to 2× and still catch all of it. The same dual-channel trick behind Amazon's Kindle + Audible.
17px
metaphoranalogystory
Anthropic just dropped Opus five, and I hope the counter guy's around because it's looking pretty good. In every benchmark I've been able to find, Opus five is coming out on top, even beating out Fable and Five Six Soul, which is particularly crazy when you know how much cheaper Opus five is from either of those models. It's under half the price of Fable five, and it's a little cheaper than five six Soul as well.
But wait, Theo, I thought Fable was Mythos. Why is Opus doing better? That makes no sense.
It's cheaper and better? Well, you're not the only one who feels that way. Hank Green just replied to my tweet saying that he's feeling overwhelmed when I asked how people felt about Opus five, and I understand.
This model is weird. It's confusing because there's a bigger, better model by the same company that is somehow benching worse. It's even crazier because they're not bench maxing.
We've all seen models that score well on the benches that aren't great to use, and we've seen models that are great to use that don't score well on benches. What's really strange here is it's a model that does score well on benches, is cheaper, and somehow is actually one of my favorite models.
I've been coding with it all day, that's why this video is out so late. Huge shout out to my editor for getting it done still. And I have some interesting conclusions.
I'll spoil the end for you now. This is probably the only model you need. I promise I'll do my best to justify that after a quick break for today's sponsor.
I'm gonna be so real with you guys. Users don't want yet another app or website to go to. They just wanna use the things they're already familiar with.
Things like SMS and WhatsApp. But how do you get those to use your services? Great question.
And thankfully, have a great answer too. Today's sponsor, Centium. These guys build a platform that makes it as easy as possible to access your users where they are on their phone numbers.
When you have a phone number to contact, you put it in the service and it will figure out what they prefer. If they like SMS, it'll route there. If they have RCS, it'll route there.
And if they have WhatsApp, you bet your butt they have that covered too. They have great packages for every ecosystem and language you'd want to use, but more importantly, they now have an MCP server for agents.
So if you're building a service that needs to be able to text users or you just wanna set up your own services to text you when things change, if they have bindings for agents and MCP, you can now do it through Sent. This is great for a lot of things that would normally be annoying, not just for sending messages, but for looking up numbers, going through your analytics, all the types of things that you would normally be doing in the background while you wait for your agent to go write the fun code.
Now you can have your agents do this too. You can copy paste this one command, sign in to Sent, and now you have access to everything you need to do good interactions with your users in the apps they already use. I'm not gonna pretend every app could be a text message, but a lot of them could and most could benefit from it.
So take advantage of where your users are at soidiv.link/sent. Let's start with the official article from Anthropic. Introducing Opus five.
It's kinda strange this model came out on a Friday afternoon because it's been rumored for weeks. I put out my conspiracy theory about this recently, which is that I didn't think Opus five was beating Kimmy k three in benches, so they delayed it accordingly. Now that I've seen the numbers, that does not seem to be the case.
Opus five's available today. It's a thoughtful and proactive model that comes close to the frontier intelligence of Fable five at half the price. I also have to talk a bit about that half the price thing, because it's it's not as true as it might seem.
On coding and knowledge work evals like Frontier Bench and GDP Val AA, Opus five is the new state of the art, though it remains behind Mythos five on cybersecurity tasks. This I will also explain.
It's designed to be used every day, it works more efficiently than other models. It's the new default model on Claude Max, and it's the strongest model on Claude Pro. They worded this this way here because they removed Fable from the Pro plans, the $20 plans.
You can only get it on the $102,100 dollar plans now, and it's only for half your usage limits. Also, things we'll have to talk about more. But I wanna lean into this part first.
It's designed to be used every day, because this is this is the interesting piece, and I actually really agree with them here. We'll start by looking at the numbers. It beat out Fable handedly in agentic terminal coding with Frontier Bench.
You're not familiar with Frontier Bench, it's not my favorite bench. They actually removed the part of it I hated the most, which was their diamond subsection, because it had the weirdest noisy chart I've ever seen. I actually just discussed this bench with a recent hire at Cognition, and they're doing a deep dive trying to make it less strange.
But the goal of Frontier Bench isn't to be yet another code bench seeing if it can solve unit tests. It's an attempt at measuring the maintainability of code, and more importantly, the mergeability of code.
Does it follow the heuristics of the code base? Does it touch things that it shouldn't? Those types of things that they can use to measure the quality of the code it's actually writing.
And this is one of the few code benches that OpenAI was behind on, especially 5.5. It was way far behind other models like Fable and even Opus 4.8. Opus five is crushing, but Soul also crushed when they fixed the bench a bit.
So 34.4% was the highest before. Opus five's now in at a 43.3.
Believe me, we'll talk plenty more about code. We also have GDP val, which is a bench I don't love, but it's interesting to see models in very specific economic circumstances, and it does pretty well there.
It slaughtered Arc AGI three, which is one of the most controversial benches to come out in recent time, This is no longer just measuring can the model complete these weird geometrical pattern tasks, which is what ARC AGI was. It's weird pattern recognition things that LMs are bad at.
The LMs got good at it, they kept making it harder. And v three is no longer just measuring if it could solve the problem or not, it is penalizing based on how many steps it takes and how much reasoning it does. So if the model gets it right, but it takes multiple steps, and a human would have solved it in one step, the model gets scored really poorly.
Our KGI three's only a few months old, and when it came out, there was no model scoring above a 1%. So to already be at 30 is crazy. We might actually saturate this bench, even though I didn't think that was possible.
We also have a gentic search through browse comp where it's doing really, really well. It's comparable again to five six SOL there, 90.8 versus 90.4. Multidisciplinary reasoning with Humanity's last exam, with tools, it did very, very well.
Without tools, it did slightly worse than Fable. This is also gonna be a recurring thing, so remember this. When Fable is just being quizzed on knowledge, it wins.
When it's being quizzed on its ability to get work done with the things it has around it, it does slightly worse. I don't trust OS World two as a bench, because I've just never seen it reflect reality particularly well. It seems to think Opus is a best in class computer use model, and that Fable was before it.
I know for a fact Five Six Soul is significantly better at computer use, so I just don't trust that one. DeepSUE is another one of my favorites. It's a much more realistic CodeBench, although it's not measuring mergeability in the same ways as something like Frontier is.
It came behind Fable here, but it also came out behind Soul, too, by quite a bit. Soul is the industry lead there. Did I mix up Frontier Bench and Frontier Code?
Yeah, I did. Apparently Frontier Bench is a new thing built by the only Terminal Bench and Harbor, and Five Six Soul was the lead until today.
So yeah, everything I said earlier about Frontier Bench, I meant about Frontier Code, not Frontier Bench. So apply that all here. Fable's still the winner, but it's actually very, very narrow, which is interesting.
Business Workflows with Automation Bench, it did very well in as well. It seems like the business use case is more and more a focus for the labs. In the Legal Agent Bench, it did okay, don't think it matters too much.
In Health Bench, it's still behind Mythos in a meaningful amount. And in Bio Bench, it actually did quite well. These are very, very interesting numbers.
For code stuff, it generally comes out about as well as Fable, if not slightly ahead. For everything else, it varies wildly which side it's on. Next, have performance and cost effectiveness.
Opus five provides greatly improved performance for the same cost as its predecessor, Opus 4.8. Cost as in the token cost, which is mentioned at the bottom here.
$5 per mill in and $25 per mill out. Whereas Fable five is $10 per million in and $50 per million out. So it is, literally speaking, half the price of Fable.
Thankfully, they don't brag about that here because it's not a realistic way measuring what's up. The costs on this are measured in the actual amount it would cost to run the thing instead. Because running a task takes different amount of tokens for different amount of reasoning in different models.
And while Opus five is much more token efficient than Opus 4.8 was, it's still not token efficient enough to really compete with Fable's numbers. Fable is still the most efficient model Anthropix ever put out. It's still nowhere near as efficient as a model from OpenAI would be, but it's better enough to be meaningful.
This chart looks pretty good because the cheaper is left and the more expensive is right, and if you're measuring Opus versus Fable purely on dollar amounts, it looks pretty good. You're getting a higher score for a cheaper price. And if go over to CursorBench or the Artificial Analysis Index, you'll see the same.
You'll also see a weird little dip there, which is what happens when you have a max reasoning level on that you probably shouldn't have. But the point I'm trying to make here is that the cost is better than Fable.
But all stories have multiple sides, and now that we're looking at the actual cursor bench numbers, we can turn off cost and switch to tokens. And what you'll see here is that Opus five is using a lot of tokens. It's using as much, if not more, than Fable for a lot of things.
It's not quite as bad as Fable was on max reasoning levels, but it's still doing similar token utilization. And if we hop over to artificial analysis, you'll see that on Fable five, it did around 33 k tokens per task, and on Opus five, it did around 37 k tokens. What that means is this model is meaningfully less efficient.
So the tokens are cheaper, it's going to use more of them, which means it will take longer to respond, it'll take longer to complete tasks, it will use your context window faster, which means it will go off track slightly sooner, and it means the cost difference isn't quite as big as it might seem. For real world use cases, it does not seem to come out to that 50% discount.
It comes out a lot closer to like 20 to 25% off. For artificial analysis, cost per task for Fable was around $2.75, and for Opus five, was about $2.03.
Okay, Theo, so it's slower, it's smaller, it's similarly expensive. Why do you like it so much? The main reason I like it is, to be frank, I used it, and it's good.
And I'll do my best to explain what makes it good and why I think it's such a good default model in a bit. But I'm gonna start with this weird example. When I started using Opus five today, I wanted to use it where I normally do.
I wanted to use it in the overhauled t three code, which is so far ahead of other agentic experiences now, it's kinda stupid. We'll We'll talk more about this in the near future. I'm genuinely so proud of what the team has been cooking.
The new sidebar system with settling is great. But there was a problem. For various reasons, we have chosen to hard code the models inside of t three code for Claude.
We don't do this for Codex or OpenCode or Cursor, we only do this for Claude because the manifest that we'd get out of Claude directly wasn't great. I wanted to fix that though, because I didn't wanna have to go manually add Opus five or cut a release just to have it inside of t three code. So I spun up Cloud Code, the real thing, and asked it to use Opus five with high effort to go explore and figure out how we could do this in a better way.
But I wanted to compare with Fable, so I gave the exact same prompt to Fable as well. They both dug in, and they both made plans. And the first thing that happened, and this was very, very annoying, it was unable to actually upload my plan using my plan skill because someone at Anthropic tuned their auto mode a bit differently.
And now auto mode classifies a action I told it to take from a skill that I wrote as potentially harmful and refused to do it. And then we go copy paste the command to do the upload myself.
Actually obnoxious. It didn't even give me a like, you can hit yes to let it happen thing. It just blocked it and told me I had to go do it manually.
Obnoxious. And no, that's not just an Opus thing. The same exact thing happened to my Fable thread next to it.
Get your shit together, Anthropic. Seriously, this model is too good to be gimped by your shit. Anyways, after I manually uploaded the file, after, first off, I told it, I want the plan uploaded, please run it, and it got blocked again, same classifier denial, not a transient failure.
Thanks, Opus. Same deal here. The classifier doesn't see our conversation, so it keeps denying the post plan upload regardless.
Thanks. You know whose auto mode does see the context of the conversation?
You can guess. Anyways, I did the upload for both, and then I asked each model to independently review the plan from the other.
I won't spoil the results. I want you to just think about this first. How do you think Opus felt about Fable's plan, and how do you think Fable felt about Opus's plan?
Now I will show you the scores from one of these two. This model thought its plan was a 7.8 out of 10, and it thought the other was an 8.2. And there were some very big gaps in here, like in the external research and alternative analysis section.
It felt like its plan was way better, but in the honesty about residual hard coding and the failure modes, it felt like the other plan was much better instead. Okay. That's the first one.
I want you to guess which model you think wrote this and said the other plan was better. Not gonna give the answer just yet because we have the other model. And here's the fun twist.
Model two also thought the other model had the better plan. It didn't see as big of a gap, but both model thought the other had a better plan overall.
The one that perceived the bigger gap was Fable. The second one, the one I'm showing you now, this is Opus.
So Fable said Opus' plan was meaningfully better in ways that mattered. Opus said Fable's plan was better in smaller ways, and that was pretty close. So I did what any person with sufficient AI psychosis would do.
I asked both models to update their plan based on what they learned from the other, and then review again. And both still said the other plan was better, but they gave more useful information this time.
What's funny with this one is that Fable gave Opus a nearly perfect score. Ten, ten, ten, ten, ten, ten, nine, nine, nine, and then a seven, because it didn't like the external research not being included in the plan. Opus gave Fable much lower scores overall, nine nine nine five nine nine seven seven five nine averaging to an eight, but it still said their plan was better.
So Opus said to ship Fable's plan as the base. It's better in round one and the gap widened, but there are things we can improve in it.
That's not what I did. I just had Opus use its plan and go build it. But there was one last opinion I wanted, which is Souls.
I had Soul review both plans. I labeled the plans o five and f five for Opus five and Fable five. I didn't know that, And it gave much different scores.
It did an 8.3 for one of the plans and a six point o for the other. Ready for the mind blowing part? The 8.3 was Opus.
What the fuck? I thought Fable was the god model. How is Opus doing so much better a job where everyone other than Opus thinks the Opus plan is better?
This is what makes this model special. Remember this post from Peter that I quoted in my Fable versus five six video? I think it's very useful here.
My overall view is that Fable is a wise owl who is very thoughtful and very well spoken. Five Six Soul is like a rottweiler who will grab the problem by the throat and not let go until it is done. This is why I think Opus is so cool.
It's somehow perfectly between the two. I want you to remember this as we continue, because this is the theme I want to beat into your heads, and we will we will talk about it a lot throughout. If we're gonna be comparing Opus to Mythos, then we need to talk about safety and alignment, because those are the concerns that Anthropic had around Mythos specifically, is that when they put all of this knowledge into this giant model, alongside that knowledge came the ability to hack shit, they were very concerned.
And now that we've seen what happened with GPT six, pony hugging face, those concerns are probably valid. According to Anthropic, Opus five is the most aligned thing they've ever made. It adheres to their constitution better than four eight, Sonnet five, or Fable five.
It exhibits the lowest rate of deceptive behavior, and it's the least susceptible to being tricked into misuse. It's also their safest model yet in terms of avoiding reckless actions that could have hard to reverse side effects. All good other than the constitution.
I think this document is insane. I have a video where I read the whole thing. It it hurt me.
But in their misalignment benches, it did quite well. It also does not advance the frontier in risky dual use capabilities. I saw a lot of people who were upset that this model came out without all the restrictions, especially when compared to Mythos and Fable, saying that, oh, I guess Amprophic was just lying with Mythos then, because suddenly they'll put it out and not have all those restrictions.
I still can't believe people think this way, but they do. This is an individual saying that they are confused how Opus five is better than Fable in every single bench, but it didn't need to be, like, government approved with all that Ultron marketing. To which I replied, hey, it's me, Anthropic's number one hater reporting in.
Opus five is a new model trained on what they learned from Mythos. Hell, it was probably largely, I said disrupted by Mythos. I was slide typing in the shower when I typed this.
What I meant was distilled by Mythos. Let's say, hypothetically, that Mythos is a big model, and as a result of its training, it has a 100 theoretical capabilities.
Most of them are good, but five of them can be bad. This happened because they weren't working from a fixed set of capability. They were trying to squeeze in as much as possible.
This is like, imagine that you have a bowl that you want to have all of the best possible food in.
So you put all of the food in. There's a lot of good food in there, but there's a lot of shit in there that you might not want. Some people might want it for some things, but you probably don't want it for these other things.
So when they distilled, when they shook out a bunch of things from that bowl trying to filter it to just the parts they wanted, they ended up with less food, but the food that's left is more of the food we want. Opus five started in this way.
It started from mythos. It's almost like they saw all of these things. They're like, okay, this is the subset that we actually want.
Let's get everything out, and maybe we can improve a few things in here while we're tinkering with it. A model like Mythos isn't trained by another model, at least in a traditional sense, because it's the best model. It has the most info.
So what they've done here is they took the model with the most information and used it to get the right pieces of info into someone else. You can almost think of this as like a teacher student relationship, where the best teachers aren't the ones that never made a mistake.
It's the ones who made all the mistakes and can now guide the students into avoiding them and not wasting time on things the teacher did. Like if the teacher spent three years studying bullshit and then six years studying good stuff, they can steer the students to only study the stuff that matters. It's kind of how it works.
It's the simplest I can put it. And OPUS five is a much smaller model, so obviously they're trying to put in the parts they want. And also because the hacking stuff is scary, they're trying to get as much of that out of it as possible.
And the alignment benches that have been published confirm they were definitely successful with this. It does seem contradictory, but the more you look at it, the more reasonable it seems. Pardon me for a moment because I have to do another crash out real quick.
I used to love Ars Technica. It was a really good journal. When I was looking for sources for today's video, I found their article about Opus five.
Opus five is about token efficiency, not a capability leap. There's one sentence in here that makes me question if anyone knows what they're talking about anymore.
Companies like CursorMeta have been building model routers, systems that automatically select models of varying size and capability from an array of options based on the nature of the prompt. The idea is that you save a lot of tokens by not using something like Fable for every task.
How do I beat this into your guys' fucking skulls? How do I convince the world that the smarter models don't use more tokens? The smartest models use less.
Again, some of the smartest models, the second or third, depending on how you measure it, best model ever, which is five six SOL, is basically as far to the left as you can go on the token efficiency chart. And Sonnet five, which is a garbage model that has almost no real uses, is the least token efficient of modern models, getting literally four to five times more tokens used than five six SOL, which is a much smarter and more capable model.
Fable is smack dab in the middle here, and Opus five is less efficient. Opus five uses more tokens than Fable. So to whatever author at Ars Technica decided to write this sentence, the idea is that you save a lot of tokens by not using something like Fable, I hope you have someone who knows what they're doing review your work in the future because you are spreading this information that I'm stuck cleaning up now as the YouTuber, so thanks.
It is still good at finding vulnerabilities, which is awesome because that means you can use it to solve bugs in your software. But its ability to actually exploit those things is much, much worse.
While it can find the bugs, its ability to exploit them is much lower, which is awesome. It means they successfully lobotomize the model in the right ways. Previous models have been hurt much more by changes.
That's part of why Mythos and Fable don't necessarily have the training to keep it from behaving this way, they just have the classifier in front. Because Fable is Mythos, the only difference is what requests are allowed in and what responses are allowed out, it's just a guard in front. So Mythos and Fable, same exact thing, it's just a difference of what's allowed in and out.
They didn't restrict those models because they just poured everything into them. Opus had these restrictions baked into it.
Historically, that made the model much, much worse at real world use cases for patching bugs and things. Opus five does not seem to have taken that same hit while also still maintaining the safety difference, which is a nice change.
It means that it's going in the right direction. Hopefully, you already watched my Solo versus Fable video. If not, I still recommend it.
I think I did a great job of breaking down the, like, philosophical differences of these models. But I think this is a really good place to talk about Opus from. As I mentioned before, I've been using it for different code tasks all day.
We can hop through these, Opus five, Opus five, Opus or Fable five, so I was doing a test having Fable review a plan Opus wrote. Opus five, Fable five again reviewing a plan from something else. Actually, this one was funny.
It wasn't reviewing the plan. I had both trying to fix some remote dev connection stuff. And Opus took way longer than I wanted it to.
So I took his plan, I threw it at Fable and Solo. Both said it was a good plan, had some feedback, I had to adjust the feedback, and then I told it to implement. It took, like, four hours to do the implementation, and I will be real, I had very little faith in it until I told Fable to do the same, and after six hours, it was still going.
I ended up burning a shitload of usage from these tests. But this PR is actually coming out decent, and I'll hopefully, fingers crossed, get to merge it tonight. I also have a bunch of others that I merged today that Opus five did.
Opus five did this. Five six Soul did this, because I was sticking around. You get the idea.
I got a lot of work done with Opus, and I'll be frank, I was really impressed with it. So what impressed me so much? Well, if you scroll back here, you'll see a lot of my other work has been Fable.
Fable five, Fable five, Fable Fable Fable five. I've been using enough Fable five that I'm maintaining three subscriptions for Fable right now.
I mostly have these because I'm using Fable so much. And what you'll see at the end of a given window is that I have most of my seven day limit left. I have about 50% left, and the Fable will be at zero because you only get half your weekly limit for Fable.
So let's say you got theoretically $200 of credit for a week, you only get a $100 of Fable if you use the other 100 for other things. It's effectively how this works out for the subscription plans, to be clear. What this means is that Opus is a much better deal not just because it's cheaper and will burn through your limits less quickly, but also you get 100% of your limit, not the 50% that you have with Fable.
What you'll see here is on this account, have used almost all of my Fable. I only have 2% left. And for my seven day limit, I have 50% left, which means I've only really used Fable from that account.
Same deal with my other account at the bottom here. But this one on the left, this is the one that I've had my router pointing things to for Opus because I already burned all my Fable on it, I might as well use it for Opus, Sonnet, and Haiku for the things I need that for.
And all the work I did today, and it was a a meaningful amount of work, I went from 50% of my seven day to 38%. So a full day of work only used 12% of my weekly. Meanwhile, I had almost all of these accounts reset in the last two days.
I burned through one and a half of the limits of Fable across all the accounts in a day. I can do one and a half weekly limits for a day of work with Fable. I did about 12% of a weekly limit with Elpis.
That in and of itself should kinda tell you the whole story, like this model is so much better at using your limits. The cost difference isn't quite as severe here. This is mostly the arbitrary nature of how Anthropic subsidizes these subscriptions.
Because Opus is lot cheaper for them to host, and there isn't as much demand for it right now, they could be more generous with it. There are three key reasons I think you should use Opus most of the time. First, as I said before, is cost slash limits.
If you're trying to get the most bang for your buck at frontier levels of intelligence, I still think you should probably just be using Five Six Soul on medium or high settings. It is way more efficient, it is way cheaper because it uses so many fewer tokens. But there is a a taste that comes with anthropic models that is only really present in the highest end ones, models like Opus and Fable, primarily Fable, but even Opus four eight, four seven to an extent, was better at, like, getting syntax in a way I didn't hate or trimming down PRs to be simpler.
Again, it's the Rottweiler thing that the OpenAI models do. They want to win so bad that they will make a mess in the process. Inphallic models, I found to make fewer messes.
So what are these three reasons? I'll give you number two super quick, because it's a relatively easy one. ZDR, zero data retention.
One of the annoying parts of Fable is that in Anthropic's pursuit of preventing any bad usage, they have been actively auditing and logging every single request that goes to it, even if you're a big company on an enterprise plan. There is no way you can call Fable or Mythos without Anthropic getting to save that data, and a lot of companies don't want that.
A lot of companies legally can't do that. As such, none of them could use Fable really at all. Opus doesn't have those same restrictions.
So immediately, a shitload of potential use cases that you could not use Fable for just because of this policy are now opened. So that huge win, suddenly a frontier model from Anthropic is useful for real world enterprise work again.
So what is number three? This is the important part, the thing I've been dancing around because it's gonna be hard for me to put this into words. It's not that clever.
This one's gonna be hard to justify, so I need a sec to figure that out, and while I'm doing that, we'll do another ad real quick. Today's sponsor's Convex, and if I'm being real with you guys, it's strange they sponsor me because I show them so much for free. I could sit here and tell you all the reasons I love using Convex for my back ends, but I'd rather just show you.
I also haven't used the plugin yet, so let's give that a go. I'm gonna open Cloud Code, I'm gonna paste the install command. We now have it ready.
There's some examples here, so I'm just gonna go with the one it showed there. Build a Kanban board with Convex. Let's see what it does.
While this is running, I should tell you guys a bit about why agents like Convex so much. The biggest reason is that they don't really need anything to use it. They just have a folder in your code base that describes everything your infrastructure can do.
All of the endpoints that can be hit, all of the data that's accessible, and all of the sync that makes the app actually good and nice to use. They actively benchmark how well every model's able to use Convex, and the results speak for themselves. The scores are insane, and they keep going up because Convex is way easier for agents.
You ask an agent to build an app with Postgres, it just won't come out as good, and it will take significantly more work. That's why I ported all of t three chat to Convex. It's just so much better, especially with the mobile app coming.
And even companies like OpenAI have started leaning on Convex for their powerful primitives. And there you have it. It built the full app from scratch.
But it wouldn't be a Convex demo if I didn't show you guys the magic. It syncs across all users, all browsers, all things. No more weird states because one person's page is behind.
I promise you these guys solve a ton of problems. See what I mean at soidiv.link/convex. Why would I ever say that not clever is one of the reasons to use a model?
One of the reasons to default this model. Let's go back to this comparison of Sol versus Fable. The first benefit was cost, because as I said, it is way more token efficient.
Second benefit was time to complete, because the token efficiency plus the WebSocket hosting layer just makes OpenAI models complete work faster in the right harness. Computer use, I still think OpenAI is in the lead, especially the software side. Diligence and eagerness.
This is the one I want to focus on. The diligence and eagerness side.
The other ones I wanna focus on here are the follows instructions really well, and the writes way too much fucking code. These three things are important because the top two are what I loved Soul for, because it would solve the problem at all costs.
But it also would write way too much code, and it would write a lot of things I didn't wanna merge. I almost only use Soul now for code I don't intend to merge, or as an implementation agent doing work that was specked out by Fable, or now specked out by Opus. We go over to the Fable side here.
I argued that it understood intent better, that it hallucinated less, it had better taste, it would write less code, and it was clever in a big model.
The negatives were the cost and the speed, but also it would give up too early.
This one, I should have worded a little differently in retrospect. It's not that it gave up too early so much as it tried too hard to be clever and find workarounds to the problem instead of just working through the problem. Bable was always a little too quick to go, but have you considered?
And it almost never got it right. Okay, it did actually get it right more often than I expected, but I often had to get Fable to think bigger and go read more code than it planned to. Fable really wants to do that thing that the clever, overpaid, ten year in the company engineer does of working around the problem so they can ignore the problem.
Opus is like these two models had a kid. It has so much of this eagerness and instruction following behavior of Five Six Soul.
It has so much of it that poor Thoric just had to write an article that he rushed out today, The New Rules of Context Engineering for Claude five Models. I would argue that these rules apply way more for Opus than they do for Fable, because the main thing he is emphasizing here is that you don't have to repeat yourself so much anymore.
Overall, we found that we were overconstraining Cloud Code, both through our system prompt and in our CloudMD files and skills. For example, when we read transcripts of our own internal usage of Cloud Code, we saw several conflicting messages in a single request, like leave documentation as appropriate or do not add comments as our system prompt skills and user requests would all clash with each other.
System prompt would say leave documentation as appropriate, the skill would say don't add comments, and your request would say just make it work like the old one. This, yeah.
You had to do this though, because anthropic models didn't just do the thing you told them to. They would often go kind of mad and try to guess what you really wanted.
So you'd have to remind them over and over again, hey, don't do that. Please stay away from that thing. Don't run this unless I tell you, that type of thing.
Opus five is the first model from Anthropic that I do not feel does that. It does what you ask. And if it's not sure, it just asks questions and clarifies, and asks some really good questions, some that have really impressed me throughout the work I've been doing with it.
And these are the things that I always loved Soul for, to be clear. I just love OpenAI models and the fact that they do exactly what you say and nothing more and nothing less. I'm just sad I couldn't get OpenAI models to write good code by just telling them to go write good code.
They won't do that. You can give enough examples and it helps. But OpenAI models tend to write TypeScript like Python devs, and I just I hate it.
I really hate it. They write way too much code, they smoke test the shit out of everything, but you can tell it to not do things than it usually does. Fable has more taste.
It still does. I would still argue Fable is the model that I I like the results from the most. The issue is in its pursuit of being the most knowledgeable and the most likable, it also cuts corners.
And I have often had to have Fable build a thing, do it as minimally as possible, then have Five, Six, Soul come in and say, hey, what did Fable miss? Soul would come in and find all the edges that Fable pretended didn't exist, and then it would ask either or or both to simplify further to try and trim down the fat and make the simplest, most concise solution to the problem.
With Opus, I just don't feel the need right now. Opus five has found a pretty solid in between. It writes slightly more code than Fable, and it takes meaningfully longer than Fable because it checks way more shit than Fable does most of the time.
Opus is diligent now. It wants to know its change will be good, and it will sit there and spin and double, triple, quadruple check to be sure. It's almost a little insecure about what it does.
I remember when Opus four eight came out and I was using the new OpenAI model early. I would have Opus review the code from five six, and it would say, oh my god, this code is surprisingly well architected. Then I would show it to Fable and it would be like, you think that's good code?
Wait till I show you something. Opus still has that aspect to it. It still has the uncertainty that it's trying to plug.
It's weird talking about these models like they're people with personalities, but but it's much better than these benchmarks that don't mean shit, if I'm being honest. The thing that makes Opus great is that it found this in between where I don't feel the need as much to reach for one or the other like I was before.
So much of my life was hopping between Fable and Soul that I started working on features in t three code to make it easier to spin up multiple tabs on a given task and rotate between the two. Since I started using Opus, I haven't felt the desire anywhere near as much. I'll still occasionally go to Five Six or go to Fable and say, hey, go look at my history, make sure everything Opus did here is okay.
And every single time I've done that, the both of the models have been impressed. Opus is just doing the right thing. All of this said, it is still a slightly dumber model, and it comes out in weird places.
Like when I was working on updating the t three Connect logic for t three Code so that you can have it auto start on a Mac, whatever you turn on the computer, It spent like twenty five minutes making its plan, and it doesn't show the right timing here. It ended up spending a lot more time verifying the results after.
So this took like forty five minutes to an hour to do. But it wrote a pretty good plan. I read through it, and it's like, this makes sense.
There are some annoying edges, but it's clever. It also has to ask good questions too, which was very useful. I asked it to make the plan into an HTML plan and host it using my HTML plan skill.
You want a person to publish it. I want you to send me the link so I can see it easily here. This is me adding too many additional things because I wanted to work through auto because I forgot that I just didn't have it on auto here.
Had it on full access. I only have Claude code on auto. The auto is weird in Claude code.
But it proceeded to do things it shouldn't. Within six minutes of starting, it starts to try to use the in app browser.
I noticed this because suddenly I had a different browser than I used opening taking over my full screen when I was working on other shit. And I was like, what the fuck? I'm using my computer.
Why are you doing this? I said, get out of the browser, Jesus fucking Christ. Understood, skipping the visual check.
It didn't understand. It opened the browser two more times, and I had to curse out the model. Stop opening my fucking browser, and stop it from running, and then let it continue from there.
But here is why I'm so mad. Sorry, Emdash.
That was me. Host plan opened it. Host plan is a minimal shim to Railway's s three bucket.
Don't blame my slop CLI when you're the one opening the browser fucker. Post plan did not open it, you liar. It opened in a browser that isn't even my default.
That was you. Yeah. It was opening in a browser I didn't have set as my default.
That's not happening when you call the generic open command on macOS. That is a thing that happens when it is being programmatically done, which is what Opus was doing.
You're right, and I shouldn't have said that. I don't actually know that post plan opened anything, em dash. I assumed it and stated it as fact.
That was wrong, and blaming a tool for something I likely caused is worse than just owning the mistake. So that was me. Three separate attempts to open a browser you never asked for.
You asked me to build an HTML plan and host it. Visual verification was my own addition to the task, and it wasn't worth interrupting you over. Yeah.
So it still does the smaller, dumber model thing in this way. I have found that Five Six doesn't do this in the immediate, annoying way. It does it in the bigger, destructive way, a smaller amount of the time, like the people who have had their whole home directories get wiped by five six soul.
That's why they've added that huge egregious warning in Codex, by the way. It's because so many people have lost their systems. They're trying to make you feel bad for using full access.
Yeah. All that said, it is generally less bad about hallucinating than Opus four eight was. It's got a 31 on the AA omniscience bench, which is a way of measuring how likely the model is to not just get an answer right, but to avoid lying when it doesn't know how to get it right.
And for what it's worth, $5.06 Soul actually scored quite a bit worse than any of the modern Opus models did, or Fable Five. Meaningfully better than Sonnet in this, but yeah, if hallucination is a problem for you, Fable Five is still the least likely to make shit up. Opus still has that tendency to just lie if it makes the answer easier.
Not that humans don't do the same thing, but it's worth knowing when you're paying for a technology. According to artificial analysis themselves, the factual knowledge is still just behind Fable five because it's a smaller model. There's just less there in its brain.
What's there has been fine tuned and refined to be as effective as possible for the real world use cases they can measure, but a lot of the more general stuff isn't there. But with that, it is strange just how closely it performs with Fable.
It really does feel like Fable's little brother in a way. For example, on Skatebench, I ran Opus five x high and max against Fable five x high and max.
Fable five x high got an 84%. Fable five max got an 82%. Opus five x high and max both got an 83 smack dab in the middle between those two scores.
Nothing comes close to Gemini three one Pro still at a 95%, but you get the idea. It's weird to see numbers this close with models that are different tiers, and it really does show that they're trying to get as much of Fable's capability baked into Opus as possible, but the factual knowledge just isn't there.
So to summarize and give you guys the guidance on what to use, Five Six Soul, if you want it to feel like a tool, like you just wanna tell it what to do, have it go do the thing and come back with a thumbs up or down. Use it if you don't care about code quality. So for side projects, one off things that you're building, scripts that are automating your life, weird things you do in like a Hermes agent or whatever, it's great for that.
On that note, if you're using it for an AI assistant, Five Six Soul is just so much better at like resolving requests, because it will do what it needs to, and it will often do it so fast that it's much nicer to use with like an Open Claw or a Hermes agent. I still prefer it for those types of things.
But most importantly with Sol is the efficiency. If you want the best bang for your buck with a Frontier model with that level of capability at a reasonable price, Five Six Sol is gonna come out way cheaper, especially if you're using a sub. I've been using Five Six incredibly heavily.
I actually accidentally spun up 32 agents. Well, I didn't do it.
Opus Five did it when I was testing. And despite all of that and my heavy use for the last few days, I've only done 30% of my whole weekly limit on just one account. So while I can barely get my work done with three accounts with Claude code, one account with Codex is more than enough still.
As long as you're not using Ultra, remember, don't use Ultra. Those are the things I liked Sol for the most. For Fable, the biggest benefit is what I would refer to as code you look at.
When I look at Fable's code, I don't hate looking at code as much. When I look at Five Six's code, I question if I ever wanna read code again. Fable Five writes good code that is actually pleasant and fits systems well and has taste.
Obviously, with that, it's way better at front end. It also just knows more. And it's hard to describe why that matters, but if you're stuck on some niche bug, you're working on a weird obscure platform or something that's just less common of knowledge, especially in our space as developers, Fable is much more likely to find a good answer to the problem.
It's also really good at orchestration, which was great for like having Fable make a plan and then spin out a bunch of work for other things to complete. It was just a really good model. But it also wasn't thorough enough to rely on.
So again, in 05/06, I would have put, like, you want every stone to be turned. You want every single possibility to be known before you get your implementation done. Fable is more, it's probably good enough, but it's likelihood to be good enough is higher, and the quality feels better.
So what about Opus? Why would you use Opus? To put it simply, use Opus because you don't wanna think about all this bullshit.
If you're tired of listening to hour long videos of me rambling about the personalities of the models, you're probably gonna really like Opus five. Especially if you've been an anthropic user historically, and you've never had a model with the, like, chat GPT level of autism, it's great. OPUS five really feels a lot more like an OpenAI model than any amphibic model has before.
It does what you tell it, it stays on task better, it's way too thorough, and it's got a little bit of self doubt that leads it into good directions. So going forward, my honest plan is for things I'm actually hoping to get merged into the code base, I'm gonna use Opus.
For the next few days or weeks or however long it takes for me to build the confidence, I'll be using Opus and then asking Fable and or Soul to take a look at the code and give a thumbs up or down as to if it thinks it's good enough, and then use that to tune my understanding of the capabilities of Opus. But I like it enough that it will be my default for the foreseeable future, and I would recommend you give it a similar shot, especially if you're paying API prices for these things and you're not interested in using Five Six Sol.
So if you're at a company that's on top of AWS and you're using Bedrock for all your inference, and you're still stuck sitting waiting for Five Six Sol to get good enough on Bedrock to be usable, you now have something very similar, slightly more expensive, meaningfully slower, but with more taste, better code output, and generally just feels a little nicer to use.
Soul is still the ultimate robot. It just does what it's told. It completes tasks well.
It will not stop until it has completed its goal. Babel is the wise ass that is surprisingly good at what it does. We have to build a relationship with it.
Opus is the reasonable in between. It feels a hell of a lot more like Soul than it does Opus four eight. It honestly feels as much like Soul as it does Fable.
It really does feel like that in between. I didn't know how much I wanted it till I have it. Fable's still my favorite model.
I'll be honest. Having something that capable just feels surreal still, and every time I click it, almost feel like a little, like, jitter when I do it. It's just, it's crazy what it can do, and the things I've seen it figure out are insane.
But Opus gives you a taste of that, a meaningful taste of that for way cheaper, and if you're using the subscriptions, way less usage.
Seriously, it's like more than four to five times as much on Opus as Fable for similar work on the subscriptions. The actual cost difference is a lot less, but again, ThoughtWorks put themselves in this position where Opus is going to use your usage way, way slower.
And as someone who has burned over $45,000 in the last twenty days of inference across my four accounts here, I have a good intuition for which ones are using things faster. And I have been surprised at how slowly I've seen the numbers tick down as I've used Opus throughout the day.
So if you already like Anthropic Models, you should default to Opus. If you don't like Anthropic models and you usually use OpenAI models, you should give Opus five a shot. You might be surprised.
I know I certainly was. And if you feel Fable burning through your limits far too fast, try those same tasks on Opus. The best thing you could possibly do though is to not just blindly trust the YouTubers and the journalists who you're hearing things from.
Now that I've seen more of how other people cover these things and how obviously factually wrong they are, I'm not gonna sit here and pretend I know everything. What What I'm gonna ask is what I have been a lot more recently, that you try it yourself.
If you're already using Fable, maybe take a prompt you were gonna do and spin it up with both, and then have each one review each other's work. Maybe have an independent model or an independent thread come in and take a look at both. You can tell Sol in Codex to look at your Claude code history and give feedback.
One of the best ways to learn how models differ is to use them for the same things and then compare the results. And if you're using Fable and you feel like your work can only be done in Fable, I implore you, go try running those same tasks on Opus and Soul, and see how the results differ. I know I have been surprised in that Opus isn't just as good as Fable, it sometimes is better.
It often catches things Fable missed, and has code that is more likely to actually work for the problems that I wanna solve than Fable does. And I don't feel this compromise I had to make before, where Soul would solve the problem at the cost of my sanity, Fable would make me feel great at the cost of the problem not being properly solved.
Opus is the in between, and I'm really liking it. So stop listening to me, go try it out yourself, and let me know how you feel in the comments.
I hope I'm not the only one out here saying that Opus five is really good. I've tried to ignore all the commentary so I could give you guys my honest thoughts. I have a feeling I'm gonna check Twitter and be really mad in a bit though.
So I'm gonna go do that. Until next time. Peace, nerds.
The Hook
The bait, then the rug-pull.
Anthropic dropped Opus 5 on a Friday afternoon, and within a day Theo had a strange problem: a smaller, cheaper Claude model was benching ahead of Anthropic's own bigger, more expensive one — without bench-maxing. He spoils the ending immediately: this is probably the only model most people need. The rest of the video is him proving it to himself in real time.
Frameworks
Named ideas worth stealing.
14:44concept
The Wise Owl vs. the Rottweiler
Fable 5 — the wise owl: thoughtful, well-spoken, sometimes too clever for its own good
GPT-5.6-Sol — the rottweiler: grabs the problem by the throat and doesn't let go until it's done
Opus 5 — lands in between the two
A framework the video borrows from a Peter Gostev post and uses as its throughline for the rest of the review: each model has a distinct working 'personality' that predicts how it behaves on real tasks better than any single benchmark score does.
Steal forFraming any multi-model comparison around behavior/personality instead of just benchmark tables
20:00concept
The Distillation Bowl
Anthropic's larger, riskier internal model Mythos is described as a bowl with 100 theoretical capabilities in it — mostly good, a handful bad (like offensive hacking ability). Opus 5 was built by 'shaking out' that bowl down to a smaller model that keeps most of the good capabilities while leaving more of the risky ones behind, similar to a teacher who made every mistake first passing only the useful lessons to a student.
Steal forExplaining model distillation to a non-technical audience
CTA Breakdown
How they asked for the click.
VERBAL ASK
30:50product
“check them out at soydev.link/convex”
Live demo, not just a read — installs Convex's Claude Code plugin on camera and builds a synced Kanban app from a single prompt to show the multi-user sync working in real time.
Theo spends a day stress-testing Moonshot's 2.8-trillion-parameter open-weight release — and comes away convinced it's frontier-class, cheap enough to matter, and genuinely dangerous once the weights go public on July 27.
Theo reacts line-by-line to Boris Cherny's post arguing that automation — CLAUDE.md rules, lint checks, CI — matters more than ever in the agent era, not less.
Theo runs OpenAI's GPT-5.6-Sol through Claude Code instead of Codex and gets visibly better designs and cheaper orchestration — then reads Codex's system prompt on camera to find out why.
Theo spends 36 minutes putting real numbers behind the GPT-5.6 hype — Sol, Terra, and Luna, benchmarked against Claude Fable, one blog chart at a time.
00:50
01:15
01:50
02:25
03:03
04:10
04:43
05:16
05:50
06:33
07:30
08:03
08:42
09:06
09:54
10:17
11:23
11:57
12:30
13:04
13:37
14:10
15:15
15:50
16:18
16:57
17:30
18:04
19:12
19:52
20:07
21:06
21:24
21:57
23:04
23:37
24:11
24:44
25:17
25:51
26:48
27:26
28:04
28:38
29:11
29:30
30:45
31:24
31:58
32:31
33:05
33:38
34:45
35:18
36:02
36:38
36:54
37:31
38:38
39:12
39:45
40:18
40:52
41:58
42:32
43:05
43:38
41:35
24:43
19:12
30:48
30:09
36:00