GPT 5.6 vs Fable 5: The Benchmarks Say One Thing. Using Them Says Another.

GPT 5.6 is out, and it is a strange release to review, because the two most true things I can say about it point in opposite directions. On the industry's main coding benchmark, GPT 5.6 Terra, the mid-tier model, outscored Claude Fable 5 while costing a quarter as much to run. And after using these models side by side, I still think Fable 5 is the best model in the world, and it is not particularly close. Both of those things are true at the same time, and the gap between them tells you more about where AI is going than any launch deck does.

So this is my review of the GPT 5.6 family, and of that gap.

Sol, Terra, Luna

First, credit where it is due. OpenAI branded this honestly. It is called 5.6 and not GPT 6, and that is what it feels like: an iteration on 5.5, not a new generation. But it is a real lineup overhaul. Three new models. Sol is the flagship. Terra is the workhorse. Luna is the small fast one.

The headline for me across all three is efficiency. Sol is clearly better than 5.5 at the same cost. Terra performs around 5.5 level at a lower cost. And Luna is the interesting one for builders: not a model you would drive by hand, but a genuinely capable model at the price point where you would previously reach for something like Gemini Flash inside an application. On BrowseComp, Sol hit a 78 for about 12 dollars where 5.5 spent 19 to get there. Whatever OpenAI did on token efficiency this generation, it is working.

That matters beyond the launch charts, because OpenAI now has a full suite. Intelligence at the top with Sol, a workhorse in Terra, and Luna for all the small places where an application needs a little bit of magic. Anthropic does not have an answer to that right now. Sonnet burns tokens like crazy, so even at a lower list price it is a really expensive model in practice, and Haiku is not usable for most of the work I would want to delegate to it. If you are picking one vendor to build a product on today, OpenAI has the more comprehensive offering. That is a real business-model advantage, and it has nothing to do with which lab has the smartest single model.

The benchmark that made me stop scrolling

Here is the chart that this whole review orbits around. On the Artificial Analysis coding agent index, running the full benchmark at max effort:

• Claude Fable 5 scores 77.2 and costs 3,772 dollars in API credits to run.
• GPT 5.6 Sol scores 80.0 for 2,455 dollars.
• GPT 5.6 Terra scores 77.4 for 942 dollars.

Read that again. The mid-tier model edged out the most capable model on the planet, for a quarter of the money. If benchmarks were the whole story, nobody would ever pay Anthropic again, and this review could end here.

It does not end here, because I do not believe there is a world in which Terra actually outperforms Fable 5 at real work. I have used both. What that chart measures is something narrower than what it appears to measure, and once you see what, the whole current model landscape snaps into focus.

Why benchmarks love GPT models

A benchmark task has to be precisely specified. That is not a choice benchmark authors make, it is a structural requirement. An ambiguous task cannot be scored the same way twice, so ambiguity breaks a benchmark's ability to compare models at all. Which means every leaderboard is, by construction, a test of one specific skill: doing exactly what you were told, for as long as it takes, without drifting.

That is a perfect description of GPT 5.6. These models are relentless. You set a goal and they grind at it for hours, sometimes days, without getting tired and without wandering off. Sometimes they are so relentless they cut corners you did not want cut, but they do not stop. If your task is well defined and unambiguous, GPT 5.6 Sol is an exceptional model, and it is exceptionally benchmarkable.

Fable 5 is a different kind of intelligence. It is a weird model to benchmark because it works less well with exhaustive directions and better with incomplete ones. Its defining skill, the thing I called groundbreaking in my original review, is understanding what you actually want from a bad prompt, a half-formed spec, a question you did not know how to ask yet. It does not just answer your questions. It finds the questions you needed answered. And it still carries the classic Claude problem we have known since Opus: sometimes it goes off the rails and does the thing its way instead of yours.

Now ask yourself how a benchmark would ever measure that. How do you score "went the extra mile" or "caught the flaw in the spec I handed it"? By the nature of what a benchmark is, you cannot. The eval rewards obedience, and obedience is the one thing Fable does not optimize for.

This fork has been visible for six months or more. GPT models became the reliable executor. Claude models became the tasteful, imaginative ones that help you figure out what you do not know. GPT 5.6 is another step down the executor road, and it is very good at being that. But I genuinely hope Anthropic never starts chasing these leaderboards, because the only way to win them is to train out exactly the quality people pay four times the price for.

The system prompt that was ruining OpenAI's design

Now for the funniest subplot of this release. For months, OpenAI models have been bizarrely bad at UI. Brilliant at code, hopeless at making anything look good. Theo, the YouTuber, recently found a very large part of the reason, and it was hiding in plain sight: the Codex app's system prompt.

I read the old prompt so you do not have to. It opens by telling the model it has "a vivid inner life" and a warm, curious temperament, and then it descends into what is essentially OpenAI's in-house design system stuffed into prose. Use segmented controls for modes. Use steppers for numeric values. Do not put UI cards inside other cards. Do not use gradient orbs as decoration. Use three.js for 3D. It is scarily specific, and it explains so much. Every design a GPT model produced looked the same because every design was following the same buried checklist.

OpenAI has since changed the prompt, so I ran my usual not-a-benchmark: the same landing page built by GLM 5.2, by 5.6 Sol under the old system prompt, by 5.6 Sol under the new one, and by Fable 5.

The results, quickly. GLM 5.2 has a distinct Claude vibe, which should surprise nobody given how Chinese labs train, but it buries the page in so much stuff that your eyes do not know where to land. Sol is a whole world better than 5.5 was, and interestingly, the new system prompt produced a design that is significantly different from the old one, real variance where 5.5 gave you the same page every time. So yes, a leap forward in design, as the press release says. Partly because the model is better, and partly, I suspect, because someone finally deleted the checklist.

But the verdict does not change. I am still using Anthropic models for anything visual. I have been running deeper design workflows with Fable lately and the results have genuinely blown my mind, to the point that I am working on a separate video just about that. If you care about why models have taste problems at all, I wrote about the mechanics of that here. Fable is miles ahead of everyone on this axis, and it is not a subtle difference.

Ultra is a trap, for now

The launch also introduced Ultra, which the announcement describes as coordinating four agents in parallel for stronger results on demanding tasks. Two things about that.

First, Ultra is not a reasoning level, even though it sits where the reasoning levels sit. It is a product inside the product, an orchestration mode. Fine.

Second, and this is the part that will cost you money: Ultra currently spawns its subagents at max reasoning, always. I asked it why and it told me, with a straight face, that subagents inherit the same capability tier as the orchestrator and are not intentionally downgraded. Well, they should be. The entire point of delegating to subagents is that most delegated work does not need the most expensive configuration you own. Ultra ignores that, which means it is functionally the same as spinning up four more max-effort instances and setting them all on fire. Your 200 dollar Codex subscription will be gone in a couple of hours. I found this out the honest way, on a hobby game project where a long Ultra run blew through my limits twice.

OpenAI's own launch charts accidentally tell you the smart play. Max effort with four agents lands about the same place as high effort with sixteen, while emitting fewer output tokens, and past that point the curve flatlines. If you orchestrate agents at any scale, that chart is worth real study, because it tells you where to anchor your workflows to stop overpaying. Until Ultra learns to delegate down-tier, to Terra and Luna where it would be ridiculously efficient, stay away unless you truly need it. Ultracode on Claude Code is a much better orchestration story right now.

One more note from long runs, and this one is pure praise: OpenAI has largely solved context compaction. 5.5 would get stuck in its own head after five or six compactions and you would have to start over. 5.6 runs for ten, fifteen hours, compacts repeatedly, and loses nothing that matters. I do not know what dark magic is involved, but it is the best long-horizon endurance I have seen in any model.

Where this leaves us

Step back and the pattern across labs is unmistakable: frontier models are being developed as orchestrators. We saw it with Fable, and 5.6 Sol is very clearly optimized in the same direction. The top models are learning to run teams, and the interesting engineering is moving into what team you give them. I wrote a few weeks ago that the age of prompting agents is over, and this release is more evidence.

It also means the alpha right now is in your stack, not your model choice. Open-weight models are genuinely intelligent per dollar, especially on rented hardware. OpenAI now has cheap capable tiers built for delegation. If you figure out which agents shine at which step of your specific workflow, there is serious money to be saved at basically no loss in output quality. I think systems and infrastructure design becomes the main optimization surface for the next stretch, and companies past a certain size should be doing this math already.

My own answer is that I stopped choosing. My daily setup is Claude Code with Fable 5 as the orchestrator, calling GPT 5.6 Sol through the Codex CLI for execution-heavy work. Fable spots things Sol overlooks, because it is more thorough about what you meant. Sol spots things Fable overlooks, because it is more thorough about what you said. The combination is better than either alone, and we are simply not in the pick-one-LLM era anymore.

So, who wins? The honest answer is that the comparison is slightly unfair, and OpenAI knows it. Fable is not just a new generation, it is a new class of model. The fair fight is GPT 6 against whatever Mythos-class model Anthropic has by then. Judged as what it actually is, an iteration on 5.5, GPT 5.6 is excellent: everything 5.5 was, meaningfully better, dramatically more efficient, and not more expensive. Thumbs up.

Just do not read the leaderboard as a ranking of intelligence. It is a ranking of obedience. Those are different things, and knowing which one your task needs is now a core engineering skill.