28:37Theo - t3․gg · Talking Head
FABLE IS BACK! (And Sonnet 5 is here too)
A 28-minute benchmark teardown of Claude Sonnet 5, plus the government letter that brought Fable back from the dead.
July 1stThe argument in one line.
GPT-5.6 Sol is now a legitimate default coding model — more determined and far cheaper per task than Claude Fable — but it over-writes code and needs steering, so the real decision isn't whether to use GPT-5.6, it's how to split work across its Sol, Terra, and Luna tiers.
Read if. Skip if.
READ IF YOU ARE…
- You already pay for Codex or ChatGPT Pro and are trying to decide which of Sol, Terra, or Luna — and which reasoning level — to default to.
- You're comparing GPT-5.6 against Claude Fable for daily coding work and want cost-per-task numbers instead of vibes.
- You run multi-agent or orchestration workflows and need to know which models can actually manage sub-agents competently.
- You want a fast, opinionated summary of a long product-launch blog post without reading the whole thing yourself.
SKIP IF…
- You don't use AI coding agents at all — this is inside-baseball for people already running Codex, Claude Code, or similar day to day.
- You want a definitive 'GPT-5.6 vs Claude Fable, which wins' verdict — Theo explicitly defers that comparison to a future video.
The full version, fast.
OpenAI shipped GPT-5.6 as three models — Sol (flagship), Terra (balanced), and Luna (cheap, agent-only) — plus a new Ultra mode that runs parallel agents. Sol posts state-of-the-art scores on coding and long-horizon agent benchmarks while costing roughly a quarter of what Claude Fable costs for comparable results, and it's dramatically better at compaction, context handling, and computer use than the prior GPT-5.5. Early testers who lost access during a preview window described the withdrawal as demoralizing, which the reviewer treats as the strongest signal of all. The catch: Sol over-writes code by default, burns tokens when it can't find a clean stopping point, and isn't meaningfully better at design. The practical recommendation is Sol on high reasoning for most serious work, Terra on medium as a budget workhorse, and Luna reserved for tasks your agent delegates itself rather than ones you pick from a dropdown.
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 – 01:22
01 · Cold open — the promise of a real review
Theo states his thesis: hard numbers and aggregated opinion instead of just his own take, plus a preview of the Sol/Terra/Luna/Ultra decision matrix he'll help untangle.
01:22 – 04:02
02 · PostHog sponsor read
Live demo of PostHog's AI chat-based analytics dashboard, showing real usage-split queries for Theo's own t3.gg codebase across Windows/Mac/Linux users.
04:02 – 16:30
03 · Reading the OpenAI launch blog, chart by chart
Walkthrough of the GPT-5.6 announcement: Sol/Terra/Luna naming, Agents' Last Exam, Artificial Analysis Intelligence Index, coding agent index, Deep SWE, programmatic tool calling, design capability, cyber safeguards, internal OpenAI usage stats, and API pricing.
16:30 – 21:49
04 · What other early testers are saying
Reviews and reactions from Dax, Jay, Max, Mitchell (Terraform/Ghostie creator), Tim (Next.js team), Corey, and the Every team, mostly on losing and regaining preview access.
21:49 – 32:10
05 · Theo's own strengths and weaknesses, written live
Theo hand-writes his verdict on a whiteboard app: determination, frontend improvement, computer use, efficiency, orchestration, and compaction as strengths; over-writing code, excessive determination, weak design sense, and poor self-awareness of errors as weaknesses.
32:10 – 35:03
06 · Which model to actually pick
Concrete guidance: Sol for anything uncertain or long-running, Terra as a budget coding workhorse, Luna as an agent-only utility tier, plus the reasoning-effort cost/score tradeoffs on Deep SWE.
35:03 – 36:00
07 · Outro and subscribe CTA
Theo teases a dedicated Sol-vs-Claude-Fable comparison video and asks viewers to subscribe.
Lines worth screenshotting.
- GPT-5.6 Sol scored the highest result ever recorded on Deep SWE — 73% on max reasoning — while costing under half of Claude Fable's equivalent run, $22 per task versus $8.39.
- On Agents' Last Exam, a 55-field long-running-workflow benchmark, Sol set a new high of 53.6, beating Claude Fable by 13.1 points, and still beats it by 11 points at medium reasoning for roughly a quarter of the cost.
- GPT-5.6 Terra and Luna outperform Claude Fable at around one-sixteenth the cost, extending the efficiency gains down to the smaller model tiers.
- On the Artificial Analysis Intelligence Index, Sol on max reasoning comes within one point of Claude Fable while finishing tasks 61% faster at roughly half the cost.
- Cyber safeguards on Sol block roughly ten times more potentially harmful activity than prior models, which OpenAI acknowledges creates friction for legitimate use.
- The programmatic tool calling API lets GPT-5.6 write and run lightweight programs that filter large amounts of intermediate tool output before it ever reaches the model, cutting context bloat from tool-heavy tasks.
- GPT-5.6's default context window is 350k, up from prior generations, and it manages compaction well enough that long agent runs no longer reliably lose track of their own goals.
- Sol will turn a five-line requested change into a 300-line file rewrite plus 2,000 lines of unnecessary tests if you don't explicitly steer it away from that behavior.
- Running Sol in Max mode disables its efficiency-focused post-training and can burn a five-hour usage window in under an hour if the model gets stuck in a difficult task loop.
- Internal OpenAI research usage of coding inference compute grew 100-fold over six months, and internal agentic token usage grew roughly 22-fold, with GPT-5.6's per-researcher daily output more than double GPT-5.5's peak.
- GLM 5.2 is cheaper per token than GPT-5.6 Sol but is so token-inefficient that it ends up more expensive and slower in practice on the same benchmark runs.
- The jump from high to max reasoning on Deep SWE moves the score from 69 to 73 but the cost from $3.40 to $8.39 per task — a large cost increase for a small capability gain.
- Multiple early-access teams described losing access to GPT-5.6 during a preview window as genuinely demoralizing, with one team reporting they stayed 'literally depressed' until access returned.
- OpenAI now bills for prompt cache writes, which it didn't do before — a net cost increase even though caching itself is more predictable.
- The reviewer frames the three tiers as direct competitive responses: Luna targets Gemini Flash's price tier, Terra targets Claude Sonnet's tier, and Sol targets replacing GPT-5.5 itself as the default.
How to actually pick between GPT-5.6's three tiers.
MODEL SELECTION
The newest frontier model isn't automatically the right default — the real skill is matching task size and budget to the right tier and reasoning level instead of reaching for the biggest option every time.
- A model that's cheaper per token isn't necessarily cheaper per task — token-inefficient models can cost more overall despite a lower sticker price, so compare total task cost, not rate cards.
- The jump from a mid reasoning-effort setting to the highest one often buys a small accuracy gain for a large cost increase — check the specific benchmark delta before defaulting to maximum effort.
- Cheap, fast models are best used as a delegated tier that a smarter orchestrating agent calls for small subtasks, not as a model you manually select for primary work.
- A model that writes too much code by default needs explicit system-prompt and skill-level constraints, or it will quietly turn small requested changes into large unreviewable diffs.
- Uncapped or 'max' reasoning modes can burn through a usage allowance in a fraction of the expected time if the task doesn't have a clear stopping condition — use them deliberately, not by default.
- Losing access to a tool you've built workflows around is one of the clearest signals of how much value it was actually providing — pay attention to what breaks, not just what improves, when a tool changes.
- When evaluating any new model release, weigh independent third-party benchmarks alongside the vendor's own blog post numbers, since vendor-reported comparisons are optimized to look favorable.
Terms worth knowing.
- Sol
- GPT-5.6's flagship model — the most capable and most expensive of the three GPT-5.6 tiers, positioned as the default for any task you aren't sure a cheaper model can handle.
- Terra
- GPT-5.6's balanced, mid-tier model — cheaper and faster than Sol, pitched as a budget-friendly workhorse for implementation and code review.
- Luna
- GPT-5.6's cheapest and fastest tier, designed to be called by other agents for small tasks like titling or data processing rather than selected directly by a developer.
- Ultra
- A new OpenAI capability setting that coordinates multiple agents in parallel to finish complex tasks faster, distinct from a simple reasoning-level bump.
- Max mode
- A setting that disables GPT-5.6's efficiency-focused post-training so the model can reason for far longer, at the cost of burning through usage limits much faster.
- Agents' Last Exam
- A benchmark evaluating long-running professional workflows across 55 fields, used to measure how well a model sustains complex agentic work over time.
- Deep SWE
- A coding-agent benchmark that scores models on realistic software engineering tasks and reports both accuracy and per-task dollar cost.
- Artificial Analysis Intelligence Index
- A third-party composite benchmark spanning agentic work, coding, scientific reasoning, and general capability, used to compare frontier models independent of any single vendor's claims.
- Programmatic tool calling
- An API pattern where the model writes and runs a small program to call and filter tools directly, instead of passing every tool response back through the model — dramatically cutting wasted context tokens.
- Context pollution
- When irrelevant or incorrect information lodged earlier in a model's context causes it to lose track of its actual goal, common in earlier GPT-5.5 long-running sessions.
- Cerebras hosting
- A promised but not-yet-live inference option for Sol that would run at up to 750 tokens per second, versus the standard 40-60 tokens per second on NVIDIA-based inference.
- Claude Fable
- Anthropic's competing flagship model family, used throughout the video as the primary point of benchmark and cost comparison against GPT-5.6.
Things they pointed at.
04:02linkOpenAI GPT-5.6 launch blog post
09:05toolDeep SWE benchmark
05:48toolAgents' Last Exam benchmark
07:23toolArtificial Analysis Intelligence Index
29:45toolBrowseComp benchmark
08:00toolTerminal-Bench 2.1
09:50productGLM 5.2
27:30productCerebras hosting for Sol
17:50productTerraform and Ghostie (Mitchell)
Lines you could clip.
25:31
“This model will turn a five line change into a 300 line file rewrite in 2,000 lines of tests.”
28:59
“But how do you pick between Terra on X high and Sol on medium? I don't fucking know.”
34:52
“Luna is their attempts to kill Flash... Terra is their attempts to kill Sonnet... and Soul is their attempts to kill GBT 5.5.”
16:35
“I've never hyped a model release... but Five Six has had a massive impact on our team. We're using five times the tokens that we used to.”
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.
metaphoranalogy
00:00GBT five six is here and it's time for a real review. As I've mentioned in other videos, I've been using it a ton, but I don't wanna just share my opinions anymore. I've talked about that in plenty of stuff and I will in the near future.
00:11I'm here for hard numbers and to do my best aggregating what other people think, as well as offering a little bit of advice on how to use the model, how to pick between the Soul, Luna, and Terra versions, the different reasoning levels, Ultra, Pro, and all of the chaos OpenAI has given us. Because when you multiply out all the different options in the Pro version and all of that, there's like 30 plus choices you have here, and it's not easy to get right.
00:34And I wanna do my best to help you while also explaining what the strengths and weaknesses of this model are. And the strengths are showing themselves pretty clearly when you look at charts like Deepest WE, where Five Six Soul on max got the highest score ever at 73%, while costing under half as much as Fable did on its max equivalent, $22 per task versus $8.39.
00:57And again, Soul got a higher score. Benchmarks are far from the only thing that matters here though. The sentiment on this model is wild, from people saying it's the best thing they've ever used, to it's changing how they write software, to others saying Fable is so much better that they basically stopped using 5.6.
01:13There's a wide range of opinions and things to go over here, and I'll do my best to cover all of it right after a quick break for today's sponsor. I've built a lot of different solutions to a lot of different problems. I have so many repos that they're hard to keep track of, but all the ones that matter have one thing in common.
01:28And no, it's not that they use TypeScript, some of them even use other things. It's this little hedgehog that just snuck into my inbox. Posthog is truly something else.
01:37They're an all in one suite of product tools that make it way easier to understand your users and give them good experiences. From their analytics to feature flags, their data warehouses, and so much other stuff, they were always the obvious solution for knowing your users better when you were building real products. And they always had a nice attitude too.
01:53But recently, their attitudes have been changing because of AI. And unlike most of the companies in the space that are just using AI to give you automatic insights that aren't particularly useful, Post Hog is going full hog with this one. They're leaning in.
02:06These guys replaced the default dashboard with a chat box that I honestly thought I would hate. If you don't know this about me, I've spent a lot of time in analytics dashboards throughout my career. I would often use them to settle arguments I was in at big companies, And learning things like ClickHouse SQL and Mode and all those tools was super useful for me when I was working at a real company.
02:27As such, at my companies, everyone just kind of expects me to be the guy to go do the analytics. That changed because of how much better this tool is. You can just ask it to create a chart and get useful insights, and it does.
02:39Like, t three code is used by Linux and Mac and Windows people. Let's just ask about this. What's the split like between Windows, Mac and Linux for t three code users?
02:49Is there any unique insights you have about how those different groups use t three code differently? Maybe one of them uses the app more heavily than the other, and now it can respond like an agent, but also write SQL to get information, generate charts, and more. Okay, we now have insights, and I'm already learning things about my users.
03:08Apparently, Linux had a slump two weeks ago, but it is on the line of overtaking Mac on Apple Silicon, and also, Intel Mac is basically none of our usage. I should probably just drop it, because like, let's be real, Intel Mac is no.
03:21Plea please don't use my app with an Intel Mac. Just go upgrade. It's time.
03:24But we can also see how many threads the average user makes across platforms. And you can see that Linux users are making more and more threads. Mac OS users are slightly so, but the Linux users are becoming more and more valuable as you might have guessed from my recent Linux arc.
03:39You can see the average project count distribution per OS as well. This is really cool stuff for me to learn from. And in a world where our products are gonna become more and more self improving, they need to know what to improve.
03:49So you need the data on what your users are doing and that's what Postlog is building for. A future where you know enough about your users to autonomously improve things for them. So if you're ready to better understand your users, check them out now at swaydiv.link/posthoc.
04:01Let's start with this unnecessarily beautiful looking blog post. We got all three of the celestial bodies they named things after, the sun, earth, and moon.
04:11Curious if they end up sticking with Soul, Terra, and Luna going forward. Who knows? So let's dive into what OpenAI has to say.
04:17I've actually not read this yet. We're launching the 5.6 family of models for general availability following their limited preview, our new flagship, Soul, alongside Terra, which is a balanced model for everyday work, and Luna, which is their most cost efficient model. The concept of dropping three models like this at once is a bit much, and I think it's gonna confuse a lot of people, and I'll do my best to explain how to think of and use all of them later.
04:40Most of this article is probably gonna be focused on Soul, though, which I would expect it's the big one. It's the one that matters. Soul sets a new standard for both intelligence and efficiency, achieving state of the art results across coding, knowledge work, cybersecurity, and science, while outperforming previous and competing frontier models with fewer tokens and at lower estimated costs.
04:57If you're curious how they make their models so token efficient, I have one of my personal favorite videos out about that that I put out a few weeks ago. Check it out if you haven't. The result of all this work is stronger performance per dollar, more successful work at the same spend, or comparable results at a lower total cost.
05:12We also introduced a new way to accelerate the most demanding work, which is Ultra, their highest capability setting, which coordinates multiple agents in parallel to finish complex tasks faster. Gonna do a dedicated Ultra video as well, because people really struggle to understand what it is and what it brings. Stronger computer use and design judgment make Five Six sole our most polished collaborator yet, helping it inspect, refine, and deliver ready to use results.
05:34We've trained Five Six to get more useful work from every token. On the agent's last exam, which is a newer bench that they've been working on, which is an evaluation of long running professional workflows across 55 fields, Sol set a new high of 53.6, eclipsing Fable five with adaptive reasoning by 13 points.
05:51Even at medium reasoning, it beats Fable five by 11 points at roughly one quarter of the estimated cost. That efficiency extends to smaller models, which are essential to making intelligence more abundant and affordable. Five six Terra and Luna outperform Fable five at around one sixteenth the cost.
06:07On the artificial analysis intelligence index, a broad measure of intelligence spanning agentic work, coding, scientific reasoning, and general capabilities, Five six Sol with max reasoning comes within one point of Fable five while completing tasks 61% less time at roughly half of the estimated cost.
06:24I'm guessing for agent last exam that there's a lot of computer use stuff here. Yeah. Most major fields of virtual work performed on a computer because OpenAI's models are way better at computer use in general, but especially now with five six.
06:39God, this agent last exam score is terrifying. Soul on x high outperformed Soul on Max. That checks out for reasons we'll discuss.
06:48The X High version got that 53.6% at around $760. Meanwhile, the best run Fable got was at $3,985, and it got a 45 oh, no, that was Opus.
07:02Babel's in the middle here with adaptive reasoning at $2,300. So Opus outperformed, but also out cost too. It has crazy safeguards, yada yada, you know all that.
07:12Let's talk about the efficiency and performance on demand. On the artificial analysis coding agent index, five six SOL with max reasoning sets a new state of the art at 20 points. That is a pretty big leap, actually.
07:25Yeah, the next highest was Fable at a 77, and then Grok Build at a 76. Grok build was tied with five five, if I recall, but five six is now meaningfully leading, specifically on Max, though.
07:39Terra also tied with Fable there, which is kinda crazy. I wanna see costs. Five six Soul was so expensive, it even beat out really expensive options like GLM five two.
07:48If that confuses you because you heard GLM five two is cheap, you need to watch more of my videos, man. I covered that a lot at this point. GLM five two is so token inefficient that even at its cheaper cost per token, it ends up being more expensive and way slower.
08:03Sadly, it looks like they don't have scores in for medium, high, and other options for SOL just yet, but Terra, which was the third highest score if I recall, was $2.76, making it cheaper than almost any other Frontier thing they've tested.
08:18It's neck and neck with five five, while getting a slightly higher score. It seems like Terra is an underrated GOAT for a lot of this. As I was reading though, they got state of the art on Terminal Bench 2.1 as well as Deep SWE, and they show a lot of these numbers here.
08:33And again, Five Six Soul, the benchmark leader, the state of the art, got the highest score we've ever seen on this on high for $1,400, whereas Fable costs $3,700 and got a slightly lower score.
08:49Oh, no. They're basically tied. 77.2 versus seventy seven point one.
08:53Yeah. Not great. But then, x high scored 78, and of course, the max, which got an 80.
09:01I think the cost difference between these isn't worth it, and we'll talk about that later on. But five six high, a very, good option, and it seems like Tera on x high and max might also be as well.
09:13Five Six can write and run lightweight programs that coordinate tools. Big deal there. It can process intermediate results, monitor progress, and choose the next action as work unfolds.
09:23This lets tool heavy tasks advance with fewer tokens, fewer model round trips, and less guidance. Instead of requiring devs to script every step or pass every tool response back through the model, the programmatic tool calling option that's built into the responses API can now filter large amounts of intermediate data, retaining only what matters, and adapt its workflow along the way.
09:44Check out my other videos on code mode. If you don't know what programmatic tool calling is, the super quick TLDR is that when you have the agent call each tool individually, you end up filling context with nonsense.
09:55If you do it programmatically, you can do things like query the database, filter, find the three rows that matter, and then only pass those back to the model instead. They talk a bit about the max option here, which basically turns off the super efficient post training.
10:09I'm sure that it still does the grug mode in its reasoning traces, but it lets the model go way, way longer. So much longer that it in like fifty minutes, not even, probably closer to thirty. Be very careful with both Max and Ultra.
10:24We'll talk about them in the future. The next section we have is design. And I know this is one of the questions y'all have been asking about the most.
10:31Is it better at design? From my experience, it is definitely better than GPT five five, although that is not saying much. Can it make good novel designs?
10:41That's that's a little tougher. If you steer it carefully, you can get good designs out of it, but you can also get some absolute slop.
10:50And believe me, I've seen some slop. Even just this UI that it made for summarizing all the work I did with the model, which, yeah, did a lot of work with this model and the window I had to use it for. I had to do some refinement passes, and this is as unugly as I could make it quickly.
11:06Our boy Dara did some designs using Five Six. Thank you again, Dara, the help here. And you can see the things it made.
11:13They're definitely less bad than I would have expected from an OpenAI model. This is also all with the design skill on. So I will turn that off and see what we got instead.
11:22Oh god, that first one here, I hate what it did with the font there. That's awful. This is fine.
11:28This is This has some ideas that are cool in it, but I don't love it. And this is garbage.
11:34Yeah. Again, without steering, this model is not going to make things that look good. To compare with Five Five, though, yeah, all of these came out nearly identical before.
11:45So it's it's it's not shit the same way, but it's still not good. It still needs a lot of hand holding. Thankfully, much like other OpenAI models, it listens when you tell it shit, so you can steer it towards better designs if you have an idea in your head.
11:59But it's not going to bring you a good design. You have to point it towards one. I made this museum website.
12:05It really likes doing this marquee style thing. I've seen that on a ton of things it designed. It's pretty good at three d.
12:11I talk about this a bit in my All the Things I Built video. I'm surprised at how well it can, like, handle three d spaces. But it's far from, like, really good at design.
12:22It does not surprise me at all that the end to end knowledge work is way better too, because again, computer use is a massive improvement, which tends to be a huge part of these types of work. I got two more Macs because I wanted to let Codex control them entirely, and I do not regret it at all. It has been awesome.
12:39I'm almost at the point where I'm willing to give it access to my Gmail, and I never thought I would get there. Here's the Browse Comp Benchmark, and they finally fixed this, so it actually has the different options. We have Sole Ultra here, industry leading at $12.17 per task, but a 92% pass.
12:55Google and Anthropic didn't give costs for theirs, so we don't actually know what it cost them to run. But you could see the massive improvement here, both how much cheaper the runs are with Sol to complete the work compared to things like Five Five, but also the improvement in score. If you look at this with latency, it's even cooler.
13:11You can see something like Five Six Medium can complete the tasks in two minutes instead of ten, which feels so much better. But also when you use Ultra, you're gonna be burning a shitload of tokens, so be careful about that. Yeah.
13:23Wait, how is Five Six Soul so much cheaper when it did so many more output tokens? Is it just super input token heavy?
13:33That's a little weird. I want more info on those numbers later. It's also good at spreadsheets.
13:39I can't believe they put this in as like a sincere example with this opening slide. Pushing the frontier on cyber and science. It did well in Exploit Bench and Exploit Jim, yada yada.
13:49It's smart, it's good at these things. It's not as good as Mythos five was, allegedly. Again, we don't have access, we can't test that.
13:56But it is still way better than it was before. Seems like cybersecurity stuff, Mythos is still the best, but we can't use it, so yeah.
14:04GeneBench Pro and LifeSciBench, it's slaughtered, MedChemBench, also seems to have done very well on. And now they talk about how they used it internally, which I think is one of the most interesting pieces of these blog posts. In particular, this section here.
14:16Over the past six months, the share of research compute devoted to internal coding inference grew 100 fold, while internal agentic token usage increased approximately 22 fold. These adoption metrics do not measure research progress on their own, but they show how rapidly AI assistance is increasing for research and across other teams like sales, marketing, user ops, finance, and more.
14:36So it looks like the research team is using the models way more than they ever have. Five point six's daily output tokens per active researcher was more than twice the highest levels observed with Five Five. And that lines up with my experience too.
14:48I've never burned quite as many tokens as I did with Five Six. Not because it's super inefficient, just because I wanna throw it at everything. There are some problems with the safety stuff.
14:56I've even experienced this. I had it block a request I did trying to clean up my lake bed code base. They called out that compared with previous models, Five Six Souls Cyber Safeguards block roughly 10 times more potential harmful activity.
15:11That is bad. Because these measures can create friction for benign use, we provide an option in ChatGPT and Codex to easily retry prompts on lower capability models. This feature was also really broken when they first shipped it, and I gave them a lot of feedback, but it seems to be better since.
15:26They plan to keep shifting this around over time and making it less likely to block, but for now, it's pretty aggressive. They're rolling out all of the models, basically all the paid accounts, free and go users, you want that like, I think it's 5 or $8 tier, they only get Terra.
15:42They don't get Sol, but they get Terra and obviously Luna, and then everything else available to pretty much everyone else, especially in Codex. And the API rates are what we discussed before, $5 per mil in, 30 per mil out for Sol, $2.50 per mil in and 15 per mil out for Terra, and dollar per mil in, $6 per mil out for Luna.
16:00They also have better and more predictable prompt caching, but that comes at the cost of them billing for cache rights, which they didn't do before, which is going to be a net increase in cost, sadly. So by the judge of this, the model slaughters and it's obviously so much better, right? Well, let's go talk about what others had to say.
16:18One of the most interesting details is how long people have had it for. Many of the early testers have had it since May 27. But also, when the original announcement happened, a lot of the early testers lost access.
16:28So we had to go through the strange experience of getting used to this model and then losing access and having to get over the fact that we didn't have it anymore, and falling back has never felt worse. Dax said the following, I've never hyped a model release.
16:44We're generally conservative with how we use these things, but Five Six has had a massive impact on our team. We're using five times the tokens that we used to. It's not even smarter than Fable or anything, but it's just so reliable and fun to use.
16:56He also calls out how miserable people were when they lost access. Jay, his boss, said even more on this. I don't wanna talk too much about 5.6 versus Fable because that's gonna be a deeper dive in the near future, so for now we're just gonna cover 5.6 itself, but I wanna talk about this.
17:10They tested earlier versions of 5.6 for a couple weeks, had a great time, it felt like a step change improvement enabling new workflows. They tried Fable and don't think it's not as good. Personally, I would've taken the experience with a grain of salt, tends to be biased in trying something new.
17:21Fable and 5.6 are taken away because of regulatory issues. Team is literally depressed that five six is gone. We're looking for anything that could even partly replace it.
17:30I actually had the same thing where I tried to make skills to distill the things I liked about five six into five five, not needed at all anymore. Fable comes back, and here's where it gets interesting. You would think Fable would be enough, but no, the team is still depressed that five six isn't available.
17:44And then five six is back, and it's immediately clear to them that it's just better than Fable. I have a lot of controversial thoughts there. I wanted to talk about how much they sucked losing the models though, because it made it really clear how much better they were.
17:57I like what Max had to say here. He thinks the most impressive part about this model is that it never gives up. If you throw it in Max's reasoning, it will just keep working until it's done.
18:07And as such, it is his favorite model by far. And this I absolutely agree with. It's the thing that makes Five Six feel special.
18:15It'll just keep going. It will try so goddamn hard to solve the problem more than any model I've ever used.
18:23And that's so different from how Five Five behaved. It's almost like jarring in a way. Five Five was really quick to stop and be like, okay, I finished one and a half of your seven step plan.
18:33Can I keep going? Over and over. Or it would have something bad in context and get lost and die.
18:39I knew five six would fix a lot of the problems I have with five five. I did not think a post training run and an RL pass would be able to make the model go from, yeah, that's smart, but a little annoying, to holy shit, this is unbelievably capable. Because remember, that's all five six is.
18:56It's not a new base model. It's not a new pre training. It's not more parameters than it was before.
19:01It is a refinement on five five, which is why its capability is so insanely impressive, and gets me even more excited for the next pre training run they do. And I'm almost positive they're working on GPT six right now. I did hear rumors that GPT six would come out end of month.
19:15I'm gonna call bullshit on that now. There is no way they're going to be able to get that done in time, just realistically speaking. More reviews.
19:22Mitchell, the creator of Terraform and Ghostie, absolute legend, has had early access as well. Souls is default. It's faster.
19:29Plans and judges just as good as Fable, and he thinks it produces better overall work. He calls out a few things Fable is good at, which I'll save for the future, because again, dedicated Fable versus video coming soon. Tim from the Next.
19:40Js team said that he's been testing five six Soul for over two months. It's incredibly good in his day to day work on Next. It understands architecture trade offs.
19:47It can investigate complicated Next issues. It considers other areas of the code base when fixing bugs. It needs very little guidance and short prompts are enough.
19:55There's some big refactors of the Next server that it implemented end to end with him pointing at high level possible improvements. The PRs are ready to merge after Next sixteen three has been released. Pretty nuts.
20:05And then Corey had the following to say, I'm really confused by OpenAI's strategy for 05/2006. I can't find the date the price jumps by 50%, the date that it leaves the inclusion and subscriptions where it's probably missing, and what their employees say is somehow aligning with what the docs say too.
20:21Clearly something's up. For those who suck at sarcasm, this is a hilarious burn on Anthropic for doing all of those things incorrectly constantly.
20:28An OpenAI employee replied, sorry to disappoint, but you can use 100% of your quota on the plan that you're paying for on Five Six forever. Additionally, the price is going to stay the same. To which Corey replied that they need to hire a VP of rug pulling.
20:40One more review that I'll summarize pretty quick from the guys over at Every. They lost access and felt like they were going insane, like they were trying to . Soul is their favorite model to work with.
20:50It's really fast, which changes how you use it. It finds the context it needs. This is a huge difference with Five Five, which was so much worse at that in particular.
20:58One thread can carry a project in production. Yes, again, it doesn't lose track of what it's doing in a thread. I probably have way fewer threads than I used to with 5.6, but I'm doing way more work because the compaction fucking works again.
21:09It's great. It plans well, but it may build too much. We'll talk about this momentarily.
21:14And it works best when you plan to steer. Soul gets better when the surrounding system supplies sources, examples, style guides, and clear outcomes. Review its choices and redirect it as the work changes.
21:25I'll talk a bit about this though. Obviously focusing on Soul for now. I'll talk about the others in a bit, don't worry.
21:31Hopefully, establish at this point. GBT 5.6 Soul is a phenomenal model, and if you use AI for writing code, it absolutely has a place in your toolbox. It doesn't mean we've answered the question though.
21:42The question of course being, should this be the model that consumes all of your usage? Should this be the model that runs all of your other things? Should this model command Fable or should Fable command Sol?
21:52That is again, a dedicated video coming up, but I wanna set the groundwork for it by talking about its strengths and weaknesses. First strength is that it's determined as hell. If you give this model a task and it possibly can complete it, it'll find a way to.
22:07Whether or not you want it to, it will find a way to. It is better at front end. It's still not better than other labs, but it's better than Five Five.
22:15It's industry leading at computer use, which honestly is one of my favorite things that I never thought I would love. I just let it control my computer, and it it blows me away what it can do.
22:26I forgot how much worse Fi five was at it until it was taken away, and I just kinda stopped using my computer because I was so much less happy with my experience. It was kind of jarring to have it taken away and then get it back, because that window without it was rough. It's also really efficient, even compared to other models in a similar capability tier.
22:46It uses way fewer tokens and comes out way cheaper at a better base price as well, which also means it's really fast, even without the Cerebras stuff. If you're not familiar, OpenAI promised us that there would be a version of Soul on the Cerebras hosting that would be up to 750 tokens per second instead of the usual 40 to 60, and I'm very excited to try that because it already feels so fast.
23:07The fast mode isn't that though. The fast mode is just more provisioning on the normal NVIDIA based inference. So if you're using fast mode and you're like, that's not as fast as I expected, it's because that's not the Cerebras hosting.
23:18That's apparently coming soon. Also on the topic of coding, it's way better at mobile dev stuff. It's also really good at like navigating around a problem space.
23:27So things like environment setup, using other machines, controlling things over SSH, provisioning, orchestrating, that type of stuff it's great at.
23:37And on that note, its capability of orchestration things with sub agents in particular is next gen. The only models that have this taste right now, this capability of understanding how to break up their work for sub agents are Five Six Soul, Fable Five slash Mythos, and Sonnet Five.
23:54Potentially, Terra can do this, but I haven't had a chance to try yet. But this new generation of orchestration tasks are a huge strength that only a couple models have, and now OpenAI ones do too.
24:07And that's enough for this to be your default model for a lot of reasons. Other things really quick that I could think of, way better at compaction.
24:18This is an important one because in codecs, you still don't get to use the full million token context window. It still limits to around 200 k. So for it to do these long running things, it needs to be able to manage the context.
24:29Now that it can compact and not lose track of what it's doing, it is way better at those types of long runs. Where before, I would basically never let Fi five run for a while because it would just get lost and do something stupid. On that note, it's much better about context pollution than it used to be.
24:45This is a huge problem I had at Fyfe five, where if it read the wrong file and put something in its history, it would just lose track of its goals and be miserable to work with. That's resolved now. And it's also better at understanding intent now.
24:58When you tell it to do something, it doesn't make a bad assumption anywhere near as aggressively as it used to. Apparently, the context it has by default is 350 k, which is better. Cool.
25:07It did not used to be that high. So now that we have the strengths, I think it's time to write the weaknesses.
25:13First and foremost, by default, it writes way too much code. You gotta adjust your system prompts and your skills and a bunch of those things to convince it to not do that.
25:26To tone down its over eagerness to just write everything. This model will turn a five line change into a 300 line file rewrite in 2,000 lines of tests. It loves writing unnecessary tests, like far too many of them.
25:42I can't tell you how many times I've had to have other models come in and clean up because five six did too much. It was too convinced those things were necessary. I also hinted at this before, but it's a little too determined.
25:53The point where it will break things if they're in the way. This model loves to work around the problems that it runs into, sometimes in ways that are a little too clever. Like when it can't launch something that doesn't have pseudo permissions, so it finds something else that does and then convinces it to launch it.
26:09Weird, sketchy shit I've seen this model do. It means that it doesn't get blocked, but it also means I kinda wanna run it in a VM more often than not.
26:17As I mentioned before, it's far from frontier at design. It's better, but it doesn't bring exciting designs to me very often.
26:26And on that note, it's also pretty bad at understanding its own limitations. It is eager to use its tools to confirm information it's not sure about, but if it thinks something is the case that isn't, it will fight you tooth and nail on that.
26:39It does not like being wrong, and it does not know when it is wrong. It has only come up a few times for me, but it was really annoying when it did, and it often sent it down crazy rabbit holes where it would try to to come up with some novel solution to a problem that didn't actually exist. Back to the determined part though, because there's other important pieces here.
26:59It can burn tokens aggressively if it doesn't have a clear stopping point. If it thinks it can eventually get somewhere, it will go and go and go even without having like a goal that it's running against. There's a lot of prompts where Five Five would have given up and stopped, and Five Six will just keep trying.
27:16That could result in massive token burn, especially if you combine that with Fast Motor Ultra. This model's the first time I have gotten into my limits on Codex ever. It's also confusing how many options we have with it.
27:28Not just Soul, but when you combine the different reasoning efforts on Soul and Soul Pro, alongside Terra and Luna, there's a lot to decide between, which can be rough. There's one more thing I wanna say about the weaknesses that's hard to put into words.
27:41The simplest I can put it is that the model's capable, but not necessarily thoughtful. It does whatever it has to, but it doesn't always think about what it's doing, and the results are that it just writes more and more and more code, or tries again and again and again to work around something instead of taking the step back and rethinking it.
28:00So on one hand, I trust this model more than ever to go off and do shit, but on the other hand, I also check-in a lot to make sure it's not getting lost in the sauce. I think that's most of what I have to say here.
28:11Is it the best model in the world? Have You to wait for my video comparing Soul and Fable for that, but I do wanna talk a little more about the options you have. Ultra's gonna have to wait for later because there's a lot more to say there, but when we're talking about Soul, Terra, and Luna, as well as all of their different reasoning options, I understand why you might be getting confused.
28:33For code tasks, I'm gonna try my best to resolve this with DeepSWE. This benchmark is great, and there's some really useful info to get here, even if it's a little clogged up, despite the fact that I just put a lot of time into hiding all the things that aren't really relevant anymore. I know a lot of people really like five five medium and are confused about what they should go to now.
28:53Well, have good news, because almost every version of Terra ends up cheaper than five five medium was, and even five six on medium is cheaper than five five was. So five six medium is probably a good enough replacement for what you're used to with five five medium.
29:10But how do you pick between Terra on X high and Sol on medium? I don't fucking know. I only had access to Sol when I was testing.
29:16That said, it seems like the spacing across options with Terra is really good. Specifically, the jump from X high to max is actually a meaningful score bump.
29:28Whereas with Soul, the bump from x high to max wasn't that big. I'll also throw in Luna here, and you can see how much messier things get when it's there. A lot of its scores are low, but it also just kinda gets in the way a lot.
29:41Somehow, Luna on max scored really, really well, but that doesn't mean I think you should use it.
29:48So here's how I would think about these options. I will start with Luna. Luna is not for you as a dev.
29:56It's really cheap, really fast, really capable, but the point of Luna is to be one of the things that is orchestrated by a smarter agent or for it to do things like bulk data processing, title generation, stuff like that. We'll probably start using Luna for doing like branch naming and title gen inside of t three code by default, so you won't have to go click any buttons for it, it'll just work.
30:17And I'm very excited that it is at the point it is at and that it is so damn good, but it is not there for you to click on it in a drop down. The point of Luna is to be surprisingly capable at that really really cheap price tier. You should let the models be the ones to call Luna, not you.
30:33And maybe if you're doing like your own programmatic stuff, like you're analyzing data or chat reads or stuff like that, you could use Luna in code, but I wouldn't select it in a drop down other than experimentation. So now we need to talk about Terra and Sol because these are the pieces that are much more interesting.
30:50Sol is the smartest model ever, depending on who you ask. It is surprisingly efficient for the price, and you should definitely throw it at work when you're not sure if a model is capable of solving it.
31:02If you're planning on setting off a task and expect it to take more than ten minutes, Soul's almost certainly the right choice. But what about Terra?
31:11Again, I've only tested Terra for about 20 or so prompts over the last two days because I did not have access during the previous early access window. So what do I recommend Terra for? First and foremost, coding on a budget.
31:23It is incredibly capable for its price. So if you're on the $100 tier or the $20 tier for codecs, Terra is probably a much better default. And you find yourself at the end of a usage window with a bunch of remaining usage, switching up to solo makes sense even on the cheaper tiers.
31:37I wrote a bit more about when I think Terra makes sense. It's great for reviewing work because it'll just read through whatever and give you good feedback on it. It doesn't have quite the same level of determination that Soul does.
31:49It will stop more aggressively. But that's also kind of a good thing if you wanna be sitting there going back and forth with the model.
31:57It's also a pretty solid workhorse for implementation. If you have a model like Sol, do all of the deep diving, figuring out what needs to be done, having Tara come in and write the code is not a bad idea, especially because from my limited experience, it seems a little less aggressive about overwriting. Doesn't do the too much code thing that I've experienced with Soul.
32:16And as I was sending out there, having a human in the loop where you're sending a prompt, waiting for it to make a couple changes, and then looking at it, it's pretty damn good for that. All that said, if you've been dealing with Fable at all at its absurdly high price and usage limits, Soul's probably going to be a good default.
32:33And my honest recommendation is that you should start with Soul, you should push it to its limits, push your usage to its limits too, and once you start getting close to hitting those limits and running out of usage, bump some of your work down to Terra and see how it goes. I do wanna talk a little bit about reasoning efforts.
32:51Again, we're not talking about Ultra yet. My honest advice there is don't use it unless you know what you're doing. Ultra is not just a new reasoning level, so be very careful.
33:00It will burn your usage. Same with fast mode, because this model will go so much longer if you get sole in one of those impossible task loops where it keeps on going.
33:09On fast mode, you will burn through your weekly usage in a few hours. It's rough. And again, when we go to benches like Deep SWE, you can see that from high to max is not a big difference in capability, but is a massive difference in cost.
33:24Where high scored a 69 and max scored a 73, with x high only at a 71, like smack dab between the two. But you went from $3.40 per task to $4.70 to $8.39.
33:37It also ends up much slower when you do that because of the number of output tokens it is generating and the additional steps that the agent is taking. So for all of those reasons, I think medium and high on SOL are really, really good defaults.
33:51My current default is SOL on high, and I do expect to stay on that for a bit. I've dropped down to medium and noticed it missed a few things that it wouldn't have on high, and I went up x high and noticed that tasks took a lot longer without getting meaningfully better results. So to really briefly summarize my recommendations, Sol on high for most things, especially once you get into the habit of orchestrating your work and doing lots of sub agents.
34:14Tera on Medium's a budget king. It's a very, very good option for real world work while also being super fast and cheap. And Luna is a thing your agent should use a hell of a lot more than you do for when they're looking for specific things, they're breaking down lots of data, or you're just processing things and generating simple text outputs.
34:32It's pretty cool having a model that is that capable, that intelligent, that cheap, and also good at calling tools.
34:40Honestly, Luna kills my use cases for something like Flash from the Gemini series, while also being cheaper, more efficient, faster, and way more reliable. So, yeah. One way to think of this list is that Luna is their attempts to because Google fumbled Flash.
34:58Terra is their attempts to kill Sonnet because Anthropic fumbled Sonnet. And Soul is their attempts to kill GBT 5.5 because they wanna have a better, smarter, more capable model that can do much longer tasks. This has been a lot to go over.
35:13I know this is a bit much. I was hoping that breaking this up in multiple videos would keep me from going off as long as I did, but I wanted to do a real review here. And my real review is that Five Six is a phenomenal model and it's the default that I go to for most things.
35:26But does that mean it's my favorite model? Does that mean that it's better than Fable? This video is already too long, so you'll have to wait for the next one to learn my answer to that.
35:34You might be surprised though. In fact, you'll probably be surprised. I have a feeling that no one quite knows my take on this overall.
35:40So if you wanna hear my thoughts there, make sure you hit the subscribe button and that little bell next to it so you know when my next videos are coming out. We're gonna have a lot to talk about with both Fable and Five Six now finally available for everybody, especially with Fable being taken away from the subscription plans in a very, very short amount of time.
35:56I'm gonna go back to prompting, so until next time. Peace, nerds.
The Hook
The bait, then the rug-pull.
Theo opens with a promise: no vibes, hard numbers. What follows is a chart-by-chart read of OpenAI's GPT-5.6 launch post, cross-referenced against what early testers actually said once the model was pulled from them and then given back.
Named ideas worth stealing.
30:22list
Sol / Terra / Luna tier selection
- Sol — flagship, default for any task you're unsure a smaller model can handle, especially runs expected to take more than ten minutes.
- Terra — budget coding workhorse; good for implementation with a human in the loop and for reviewing work, less prone to overwriting than Sol.
- Luna — not meant to be selected directly; extremely cheap and fast, intended to be called by other agents for things like branch naming, title generation, and bulk data processing.
Theo's practical rule for which GPT-5.6 tier to reach for, based on task size, budget, and who is actually initiating the call (a human or another agent).
Steal forAny product routing logic that needs to pick a model tier automatically based on task type rather than defaulting everything to the flagship.
31:48model
Reasoning-effort cost ladder (Deep SWE)
- High: 69 score / $3.40
- X-high: 71 score / $4.70
- Max: 73 score / $8.39
Moving up reasoning-effort levels buys diminishing accuracy gains at steeply rising per-task cost, so Theo defaults to Sol on high rather than max for most work.
Steal forDeciding default reasoning-effort settings for any agent product where cost per run matters.
How they asked for the click.
VERBAL ASK
35:30subscribe
“make sure you hit the subscribe button and that little bell next to it”
Standard end-of-video ask, paired with a tease for a dedicated Sol-vs-Claude-Fable comparison video, which gives viewers a concrete reason to come back.
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.
