The argument in one line.
Claude Opus 5 launched priced the same as its predecessor but performing close to or above Anthropic's larger, pricier Fable 5 model, which the video argues proves cost-per-task — not price-per-token or raw benchmark score — is now the metric that actually matters when picking a model.
Read if. Skip if.
READ IF YOU ARE…
- You build with or choose between frontier LLMs (Claude, GPT, Gemini) and want to know if Opus 5 changes your default model.
- You track AI benchmark releases and want the actual numbers from Anthropic's Opus 5 launch without digging through the blog yourself.
- You're deciding between models based on price and want to understand why price-per-token comparisons can be misleading.
- You're curious how AI labs are now measuring and marketing models on cost-efficiency rather than just capability.
SKIP IF…
- You want a hands-on coding test of Opus 5 rather than a benchmark-chart walkthrough — this video only reacts to Anthropic's own published numbers.
- You're not currently using Claude models and have no stake in the Anthropic model lineup.
The full version, fast.
Anthropic released Claude Opus 5 at the same price as Opus 4.8 ($5/$25 per million input/output tokens) but with performance that beats Fable 5 — Anthropic's larger, more expensive model — on nearly every benchmark: agentic terminal coding, GDPval, ARC-AGI-3 (30% vs single digits before), BrowseComp, and OSWorld computer use. The one exception is cybersecurity exploit development, where Anthropic appears to have deliberately capped Opus 5's capability while still improving vulnerability-finding accuracy. Anthropic frames the release around cost-per-task rather than price-per-token, arguing that's the metric that actually predicts what a task costs to run, since a cheaper-per-token model can still burn more tokens completing the same task. Sponsor Box AI's own internal benchmark shows a similar jump on real document-grounded enterprise work. The reaction is unusually bullish: Opus 5 is described as outperforming Anthropic's own bigger model at half the price, comparable to GPT-5.6 Sol.
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.
Create a free account →Where the time goes.
00:00 – 01:02
01 · Cold open: "beating Fable at half the price"
Berman opens on the claim that Opus 5 beats Fable 5 on price and performance, recaps Anthropic's model tier history (Haiku/Sonnet/Opus, then Fable), and pulls up the first benchmark table.
01:02 – 02:28
02 · First benchmark table: ARC-AGI-3, BrowseComp, OSWorld, DeepSWE
Walks the Opus 5 vs Fable 5 vs Opus 4.8 vs GPT-5.6 Sol table: ARC-AGI-3 jumps to 30%, BrowseComp improves slightly, Humanity's Last Exam is flat, OSWorld computer-use gains 4 points, DeepSWE dips slightly.
02:28 – 02:41
03 · The declines: legal and health benchmarks drop
Frontier Code is flat, AutomationBench improves 9 points, but legal benchmark drops from 13.3% to 11.7% and health benchmark professional also declines — called out directly rather than skipped.
02:41 – 05:06
04 · The real story: cost-per-task efficiency charts
Pivots to Anthropic's efficiency framing — cost-per-task charts for OSWorld computer use and AutomationBench business workflows show Opus 5 sitting above and to the left of Fable 5, Opus 4.8, and GPT-5.6 Sol, meaning better score for lower cost. Explains why price-per-token comparisons (using Kimi K3 as the counterexample) are misleading.
05:06 – 06:30
05 · Sponsor: Box AI's Complex Work Eval
Box AI segment — their own 12-industry document-grounded benchmark shows Opus 5 scoring 68% overall vs Opus 4.8's 63%, with the biggest gains on due diligence and data analysis.
06:30 – 08:20
06 · More benchmarks: Humanity's Last Exam, Frontier-Bench coding, ARC-AGI-3 explained
Returns to benchmark charts — HLE cost/performance curve, Frontier-Bench agentic coding (Opus 5 nearly matches its own high-effort score for less), then explains how the ARC-AGI-3 game-based benchmark actually works using the interactive game board UI.
08:20 – 09:14
07 · Cybersecurity: capability without exploits
Novel-problem-solving cost chart shows the 30% ARC-AGI-3 jump again, then the OSS-Fuzz cybersecurity results: Opus 5 beats Opus 4.8 at finding vulnerabilities but stays far behind an uncensored model ("Mythos 5") at developing exploits — read as a deliberate Anthropic guardrail.
09:14 – 10:48
08 · Official announcement: pricing, Fast mode, beta features, wind tunnel demo
Switches to Anthropic's own blog post — pricing confirmed at $5/$25 per million tokens (same as Opus 4.8, half of Fable 5), Fast mode at 2.5x speed, mid-conversation tool switching, automatic safety-classifier fallbacks, and a physics demo of Opus 5 building a working wind-tunnel simulation.
10:48 – 12:35
09 · Community reaction and sign-off
Shows a tweet confirming Opus 5 briefly fell back to Opus 4.8 due to safety filters, revisits the ARC-AGI-3 and OSS-Fuzz charts, reads a tweet recommending Opus 5 as daily driver paired with Fable 5 for hard problems, notes three models are now on the cost/quality Pareto frontier, and signs off.
Lines worth screenshotting.
- Claude Opus 5 launched priced identically to Opus 4.8 ($5/$25 per million input/output tokens) — half the price of Claude Fable 5 — while beating Fable 5 on most benchmarks.
- Opus 5 scored 30% on ARC-AGI-3, up from roughly 8% for the previous best model — the single largest jump discussed in the video.
- Opus 5 improved on GDPval, a real-world-task benchmark built by OpenAI, by a full 100% over Fable 5 (43% vs 33% on a related coding metric).
- Cost-per-task, not price-per-token, is the metric Anthropic is now marketing around — a model can be cheaper per token and still cost the same per task if it uses more tokens to finish the work.
- Kimi K3 illustrates the price-per-token trap: it's less than half the price of Fable 5 per token but used roughly twice the tokens to complete the same task, netting out to the same real cost.
- Opus 5 dropped on two benchmarks versus Fable 5: legal reasoning (13.3% to 11.7%) and health benchmark professional — both declines Berman calls out explicitly rather than glossing over.
- On cybersecurity, Opus 5 improved vulnerability identification over Opus 4.8 but stayed well behind an uncensored model called Mythos 5 at actually developing exploits — Berman's read is that Anthropic deliberately capped exploit capability while preserving detection accuracy, which is unusual since safety guardrails normally hurt general performance too.
- Sponsor Box AI's own internal benchmark (12 industries, document-grounded enterprise tasks) showed Opus 5 scoring 68% versus Opus 4.8's 63% — a 5-point jump, with the largest gains on the most exhaustive, multi-step analytical work.
- Opus 5 briefly fell back to Opus 4.8 due to safety-classifier flags shortly after launch, confirmed by a third-party account inspecting Claude Code's system prompt strings.
- Anthropic's own recommendation pairs Opus 5 as the daily-driver model with Fable 5 reserved for planning, brainstorming, and the hardest bugs — not a full replacement.
- Three models now sit on what Anthropic-adjacent commentary calls the quality/cost Pareto frontier: Opus 5, GPT-5.6 Sol, and the open-source Kimi K3.
Cost per task, not price per token, is now the real model-selection metric.
WHAT TO LEARN
Anthropic's Opus 5 launch shows that a model priced the same as its predecessor can out-benchmark a pricier flagship, and the only way to see that clearly is by comparing total cost to finish a task rather than the sticker price per token.
01Cold open: "beating Fable at half the price"
- A new model priced the same as its predecessor beating a pricier flagship model is itself the headline — price and performance don't have to trade off on every release.
02First benchmark table: ARC-AGI-3, BrowseComp, OSWorld, DeepSWE
- A benchmark jumping from single digits to 30% (ARC-AGI-3) is a meaningfully different kind of result than a few-point gain — treat large jumps on novel-problem-solving benchmarks as the most newsworthy signal in a release.
- Flat or slightly-declining scores on some benchmarks (DeepSWE) alongside big gains elsewhere is normal — no model release improves uniformly across every measure.
03The declines: legal and health benchmarks drop
- Check domain-specific benchmarks (legal, health) separately if your use case lives there — a model can regress on your exact use case even while its aggregate score improves.
04The real story: cost-per-task efficiency charts
- Cost-per-task charts (score on one axis, total dollar cost on the other) reveal efficiency gains that a simple percentage-improvement table hides.
- Comparing models on price-per-token alone is misleading — a cheaper-per-token model can still cost the same or more per completed task if it needs more tokens to finish the work.
05Sponsor: Box AI's Complex Work Eval
- Third-party evals from companies that integrate a model (not just the model vendor's own benchmarks) are a useful corroborating signal, since they're run on real document-grounded enterprise tasks rather than academic test sets.
06More benchmarks: Humanity's Last Exam, Frontier-Bench coding, ARC-AGI-3 explained
- A benchmark that gives a model zero instructions and just a game to figure out (ARC-AGI-3) tests something closer to general problem-solving than a benchmark built from known task types.
07Cybersecurity: capability without exploits
- A model can improve at detecting a risky capability while being deliberately held back from the harmful version of that same capability — safety tuning doesn't have to uniformly degrade performance.
- When a model is compared against an explicitly uncensored competitor on a dangerous-capability benchmark, the gap tells you about guardrails, not necessarily about raw underlying ability.
08Official announcement: pricing, Fast mode, beta features, wind tunnel demo
- A lab's own recommended usage pattern (daily driver vs. reserve-for-hard-problems model) is often better guidance than benchmark tables alone for deciding which model to reach for on a given task.
- New API features like mid-conversation tool switching without invalidating the prompt cache, and automatic fallback routing when a request is safety-flagged, change the practical cost and reliability of using a model in production, not just its benchmark score.
09Community reaction and sign-off
- Read vendors' own benchmark marketing skeptically but not dismissively — when a lab's own charts show declines (legal, health) alongside the gains, that's a sign the comparison is being presented honestly rather than cherry-picked.
Terms worth knowing.
- ARC-AGI-3
- A benchmark from the ARC Prize foundation that drops an AI model into an unexplained interactive game with no instructions, testing novel problem-solving the way a human would encounter an unfamiliar puzzle.
- GDPval
- A benchmark created by OpenAI that measures a model's ability to complete real-world, economically valuable professional tasks rather than academic test questions.
- OSWorld
- A benchmark that scores a model's ability to control a computer directly — finding on-screen elements, clicking buttons, and completing multi-step tasks inside a real operating system.
- Cost per task
- The total dollar cost to complete an entire task end-to-end, as opposed to price per token — the metric Anthropic argues actually predicts what running a model costs in practice.
- Pareto frontier (in this context)
- The set of models that aren't beaten on both quality and cost simultaneously by any other model — each one is the best choice for some combination of budget and performance.
- OSS-Fuzz
- An open-source fuzzing benchmark used here to test a model's ability to find and exploit software vulnerabilities in real open-source code.
- Humanity's Last Exam (HLE)
- A very difficult multidisciplinary benchmark designed to test expert-level reasoning across many academic fields, intended to remain unsaturated by current models.
- AutomationBench
- A benchmark measuring a model's pass rate on agentic business-workflow automation tasks.
- Frontier-Bench
- A coding benchmark used in the video to compare agentic coding performance across models at different cost/effort levels.
Things they pointed at.
04:05toolKimi K3
Lines you could clip.
00:01
“Better than Fable, half the price. This is Opus five.”
06:45
“It's no longer sufficient to just look at the price per token and think, okay, well, that's the price I pay... The main metric you need to be looking at is cost per task. That is it.”
14:04
“Opus 5 is stronger than Opus 4.8 on cybersecurity tasks, but it remains substantially behind Mythos 5 at developing exploits.”
17:19
“It's crazy to me that Opus is outperforming Fable. Does that mean the government is going to take it offline soon?”
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.
analogystory
00:00Better than Fable, half the price. This is Opus five.
00:04We have a new model, Claude Opus five. Now if you remember, Opus was their kinda top tier in their family model.
00:13There was Haiku, Sonnet, and Opus. Then, of course, Fable came out. Fable is their biggest best model they make.
00:20And now we have a version five of Opus. I tested it again. I didn't have a lot of time.
00:27They gave it to me very late, but let's look at the benchmarks. Woah. Wait a second.
00:33Wait a second. Are are you all seeing this? Is Opus beating Fable five on almost every benchmark?
00:40Agentic terminal coding, Frontier Bench, an incredibly important benchmark for coding? 43 as compared to 33 on GDP Val, which is an OpenAI created benchmark, which tests real world practical tasks, a 100 improvement from Fable five.
00:59How could this be? Woah. Arc AGI three.
01:03Shout out to Greg at ArcPrize. Woah. 30%.
01:08That's crazy. This is BrowseComp, 90% versus 87%.
01:14Humanities last exam, basically the same. With tools, basically the same. A slight improvement.
01:21We have OS World, which is computer use, the model's ability to control your computer, literally figure out where things are on the screen, click buttons, do tasks, a four point improvement for DeepSUI, the best benchmark in my opinion.
01:36DeepSwee, I'm so glad they're including this now. It had a slight drop. So I think if I were to point to any benchmark and say this is probably the vibes you all and myself are going to feel, this is it.
01:49So a very slight drop, basically the same. Now I can't wait to find out what the price is of this model.
01:58Okay. So Frontier Code, basically the same. Automation bench, a very nice improvement, nine point improvement there.
02:07On the legal benchmark, we actually had a decently significant decrease from 13.3 to 11.7. Sorry, lawyers.
02:15Health bench professional, we did have a decrease as well. BioMysteryBench looks to be just about the same.
02:23This is crazy. I did not expect this. Opus five.
02:27Wow. Opus five is also highly efficient. It outperforms other models for a similar or lower cost per task.
02:34Now if you've been watching my videos over the past week, you know I've been talking so much about cost per task. Wow. Wow.
02:43Look at this. So what we're seeing here on the y axis is the total score of OS World. And remember, that is the computer use benchmark.
02:53On the x axis, we're seeing cost per task. And what we're seeing is this orange is Opus five, and it is higher than this kind of gold yellow and blue, which is Fable five and Opus 4.8, but it's also to the left more, which means it's better and cheaper.
03:13This is a really big deal. Now here's GPT 5.6 Soul, which is interesting. It has such a wide spectrum of price.
03:21But even at the very top to match performance of Opus five at its lowest setting, it is costing over twice as much. Wow.
03:31Crazy. Crazy. Unreal.
03:33This is like this is very impressive. This is automation bench. This say oh my.
03:38Look at this. This is so crazy. So okay.
03:41Once again, a substantial improvement. And by the way, I want to say I'm very happy that they're putting so much emphasis on the cost per task.
03:51That is what matters. For example, with Kimi k three, it was half the price of GPT 5.6 sold, less than half the price of Fable five, yet it took twice as many tokens to accomplish the same task. So thus, it's a wash.
04:07You're paying the same price. It's no longer sufficient to just look at the price per token and and and think, okay.
04:14Well, that's the price I pay. Well, it's half the price. It's gonna be less expensive.
04:17No. That's not it. It is the main metric you need to be looking at is cost per task.
04:23That is it. Now what we're seeing here, this is incredible.
04:27So it's it's not only cost per task, but it's also pass rate per task, which I think is fantastic. So we have OPUS five coming in way above everything else, way above, and significantly less expensive.
04:41I'll also I'd like to say, I said this a bunch of times in my videos in the last few weeks. I said, you know, people looked at Fable, and they said, wow. It's so expensive.
04:50It uses so many tokens. And I said, well, that's their first version. They're going to optimize the hell out of it.
04:58And that's what we're seeing here with Opus. They basically used Fable, probably, this is my assumption, to to really get the most out of Opus. Thanks to the sponsor of this video box.
05:08They tested Cloud Opus five quite a bit. So our benchmark of realistic document grounded tasks across 12 industries benchmarked here against the prior generation Claude Opus 4.8. The task mirror the analytical work that knowledge workers actually do, reading source documents, reconciling numbers, running due diligence, and reviewing expert output for errors, here's what we found.
05:30So we do have a pretty sizable bump. Now this does kinda reflect the bump we saw from Fable.
05:38I I can't believe I'm saying this. From Fable to Opus five. So from '63 to '78, great due diligence, 65 to 76, report drafting from data, 67 to 69, expert review just about the same, and then data analysis, a nice big six point jump.
05:54On complex enterprise knowledge work, Claude, OPUS five is a clear step up from the prior generation. It's advantage concentrated on the exhaustive multistep analysis that drives real decisions.
06:05OPUS five is coming soon to Box AI. Thank you to Box for sponsoring. Shout out to Box.
06:10Go check out Box. We use it at forward future. All of my agents use Box.
06:15We set up the Box CLI. They just know to put stuff there. It's really nice.
06:20And go build on top of Box if you're a developer. They're awesome. They've been a great partner, so thank you to them.
06:24So here's humanity's last exam. Once again, what we're seeing, the pricing actually looks to be about the same.
06:32Um, you know, it depends on your thinking level. That's what we're seeing. Each of these dots most likely represents a thinking level.
06:38We have Fable five and Opus 4.8, but Opus is the best. It's crazy to me that Opus is outperforming Fable. Does that mean the government is gonna take it offline soon, or we're gonna have a lot of those kind of weird, hey.
06:52We can't let you use Opus five on this. You're gonna have to fall back to Opus 4.8, which I do get pretty often. But this is a really good score.
07:00Okay. Frontier bench, we have here's GPT 5.6 Soul.
07:06So right here is the max score coming in at about 35, 36% and coming in at about $12.
07:15Now we have essentially a slightly better score for a little bit more of a cost at this thinking level for Opus five, but for just, like, another $2 per task, you actually get another 5%. And and then interestingly, the score actually went down at the highest thinking level, and it was more expensive.
07:34So that's interesting. Keep that in mind as you're thinking about which model to use. Okay.
07:39Arc AGI three. The benchmark is it gives the model a a game to play or a set of games and nothing else.
07:47It doesn't even give it the name of the game. It doesn't tell it how to play. It doesn't tell it how to complete it.
07:53All it does is give it a game. Right? So, like, we'll click start, and you have to basically just figure out what to do.
08:01And you have a certain amount of moves, certain amount of time based on moves to complete it. And, again, you have no idea what it is. And so an AI model's dropped in here and just says, go.
08:12And that's it. The games are solvable by humans, but apparently very difficult for AI.
08:18Now look at this. The best model in the world prior to this was coming in at, let's see, maybe 8%. Now we had this massive jump up to 30%.
08:31This is such a massive jump. I don't think r three prize expected it. Opus five is stronger than Opus 4.8 on cybersecurity tasks, but it remains substantially behind Mythos five at developing exploits.
08:43Very interesting. Very interesting. So I think, like, the Mythos model has no guardrails on cyber, and I think that's the key.
08:50And so Opus five probably has not only guardrails, but they probably built it in a way in which they tried to remove some of that cyber capability, which is interesting.
09:01And, again, this is all speculation. Which is interesting, typically, when you remove capabilities from a model, you hurt the model generally. Right?
09:10It it performs worse generally, but we're actually seeing the opposite effect here. Let's check out the blog. Let's see if it has any new information here.
09:17Yeah. I mean, look. This is stunning.
09:19Opus five outperforming Fable five at almost everything except for cybersecurity, and that that was probably their entire intention.
09:27Claude Opus five builds a working wind tunnel. Nice. Very cool.
09:32I mean, it's not the first time that we've seen fluid simulation, but still impressive nonetheless. Claude Opus five is available today on all platforms priced at $5 per million input tokens and $25 per million output tokens, the same price as OPUS 4.8, half the price of Fable, half the price of Fable, and comparable to GPT 5.6 Sol.
09:59So this is Anthropic's answer to 5.6 Sol. Now also just keep in mind, 5.6 Sol is likely the last iteration of the GPT five series of models before OpenAI actually releases their next GPT six model.
10:15Okay. Interesting. So automatic fallbacks on the API.
10:18Users can now choose to have requests that are flagged by our safety classifiers on Opus five, so they're doing the same thing. If they're safety classifiers, and hopefully, they're not overly aggressive, it will fall back automatically route to a new model.
10:32You do pay the price of the fallback model, to be clear, but it's annoying. I don't like when they do that. Yeah.
10:38So so we're seeing already people saying Opus five has fall back to Opus 4.8 due to safety filters. They're probably more lax than Fable, but they still exist.
10:47Look at this. Novel problem solving by cost. Look at this score.
10:5230% on the Arc AGI prize, and it was able to do so less expensive than GPT 5.6 SOL. Less expensive and more than tripled its score.
11:04That is so crazy. Do we think that the government is gonna leave it? Do you think they've already shown the model to the US government?
11:10Do you think they've already gotten approval for it? I would be doubtful that they did not share this model with the government, that they didn't share the scores. And I think the interesting thing to note, and I'm gonna just go back one more time to this, this is probably why we're not gonna see a rollback.
11:25This is probably why OPUS five is here to stay and was already reviewed by the government.
11:32But what we're seeing here, so exploitation success, a 13 with Mythos, zero with 4.8, and a four with Opus five.
11:40So they really reduce the performance of the model's ability for cyber capabilities, but also improve the quality.
11:47And, again, it's kinda crazy to be able to do that. Usually, when you put guardrails on a model, you also just reduce the overall performance of the model, but that's not what we're seeing.
11:57Here's to Rick. Opus five rounds out our Claude five family beautifully. Interesting.
12:03Does that mean Haiku five is not coming? I think it's an incredible daily driver. Pair it with Fable for planning, brainstorming, or fixing the hardest bugs.
12:11Pair it with Fable. Why would he say pair it with Fable? Interesting.
12:15There are three models now at the Pareto frontier, which means a combination of quality and cost. Those models are now OPUS five, GPT 5.6 Sol, and Kimi k three, an open source model.
12:30And I made an entire video about that model. Check it out right here.
The Hook
The bait, then the rug-pull.
Matthew Berman opens mid-sentence, already stunned: Anthropic's new Claude Opus 5 is beating the company's own flagship Fable 5 on almost every benchmark, at half the price — and he spends the next twelve minutes screen-sharing the receipts.
Named ideas worth stealing.
04:11concept
Cost-per-task over price-per-token
Anthropic's efficiency charts plot benchmark score against total dollar cost per completed task rather than sticker price per million tokens, because a cheaper-per-token model can still cost the same or more if it needs more tokens to finish the same job — illustrated with Kimi K3, which is under half the price of GPT-5.6 Sol per token but uses roughly twice the tokens, netting to a wash.
Steal forAny cost comparison between AI models or vendors — measure total cost to complete the actual task, not the unit price.
How they asked for the click.
VERBAL ASK
05:06product
“Thanks to the sponsor of this video, Box... Opus 5 is coming soon to Box AI.”
Mid-video sponsor read woven into the benchmark narrative — Box AI's own internal eval is presented as corroborating evidence for Opus 5's gains rather than a hard break in the video, then closes with Berman's personal endorsement ("we use it at forward future").
MENTIONED ON CAMERA
FROM THE DESCRIPTION
PRIMARY CTAWhere the creator wants you to go next.
OTHER LINKSAlso linked in the description.
Visual structure at a glance.
