OpenAI has rethought what it means to ship a flagship. Instead of a single model with adjustable "thinking" settings, GPT-5.6 arrives as three distinct large language models, Sol, Terra and Luna, each trained separately, priced separately, and capped at a different level of ability. The contest worth watching is Sol versus Claude Fable 5, the strongest model Anthropic currently puts in front of the public.
Three models where there used to be one
Sol is priced at $5 per million input tokens and $30 per million on output. Fable 5 sits at $10 and $50, double the cost, and it is now slipping behind on several of the benchmarks that developers actually send real work through. Luna, the budget option at $1 input and $6 output, already ranks above Anthropic's Opus 4.8 on coding. That single fact turns into a genuine headache the moment July 19 arrives.
A punishing month for Fable 5
The past few weeks have been unkind to Fable 5. On June 12 the U.S. government blocked it after Amazon researchers uncovered a jailbreak that quietly repurposed the model into an unintended vulnerability scanner. Anthropic pulled it worldwide for 19 days, engineered a fresh safety classifier, and reinstated it on July 1, this time with a narrower window of access.
A deadline that will not sit still
Ever since, the model has lived on borrowed time. The plan was to shift it behind a usage-credits paywall on July 7. That slipped to July 12, and then to July 19. Every extension landed just hours before the deadline expired, and none of them came through an official post. In its July 12 message on X, Anthropic said it was keeping Fable 5 available on all paid plans and holding Claude Code's weekly rate limits 50% higher until July 19.
The logic is not subtle. Should Fable drop out of subscriptions after July 19, the best model left for paying customers becomes Opus 4.8, and Luna already outperforms that on coding for a sliver of the price. Keeping Fable on the menu, even at half the weekly allowance, is the one thing stopping Anthropic's subscription tier from looking weaker on paper than OpenAI's mid-range.
The benchmark scoreboard
When the two go head to head on published tests, the margins are thin. On the Artificial Analysis Coding Agent Index, Sol posted 80 to Fable's 77.2, and did it with roughly half the tokens, in under half the time, at close to a third of the cost. On Agents' Last Exam, a test that runs professional tasks across 55 fields, Sol reached 53.6% while Fable managed 40.5%. On Terminal-Bench 2.1, Sol running in ultra mode, which fires four subagents at once, hit 91.9% against Fable's 83.1%. Yet on the wider Intelligence Index, which rolls up 9 separate benchmarks, Fable 5 edges GPT 5.6 by a single point, so the overall gap in raw ability is almost impossible to feel.
Most benchmarks lean so heavily on coding that they say little about how a model performs elsewhere. So beyond one casually built game, the trials leaned on prompts that stray from the usual coding drill. Here is how each one played out.
Test one: a time traveller who never learns
Both models received an identical brief: take Jose Lanz from the year 2150 back to the year 1000, trap him in a time-travel paradox, and keep him in the dark about what he set in motion until he is safely home. Each returned something nearer a novelette than a short story, and each fumbled the single instruction that counted, that Jose should catch the paradox only once he is back in the future.
Sol's Jose works it out mid-story instead, recognising that "the unknown traveler was not someone he had come to stop. It was him." Fable is blunter still, letting Jose grasp it while still in the past: "He was the seed event."
Sol's piece, titled "The First Fire," plays it as clean genre sci-fi, with Jose unwittingly introducing the furnace that triggers the very climate collapse he came to head off. The opening line is genuinely strong: "Only thunder. Only insects. Only the wet breath of the world before machines." The trouble is that Sol never lets the image speak for itself. It spells out the loop, then spells it out again, and finally hands an older Jose a recording to explain it a third time, insisting his attempt to fix the problem was what created it. Lucid, certainly, but wearying by the third pass.
Fable's entry, "Lo Que Arde, Vuelve," constructs the same loop from Lake Maracaibo, Catatumbo lightning and an Añu village, with Jose accidentally birthing the prophecy he came back to wipe out simply by comforting a frightened child. The whole mechanism folds into one line: "The grief that sent him backward was the cargo he delivered." Fable's weakness is the reverse of Sol's, leaning too hard on its own prose, piling metaphor on metaphor until a line such as "You cannot pull the thread, you are the thread" feels like the model preening rather than the story asking for it.
On the writing alone, Fable's "Lo Que Arde, Vuelve" comes out ahead of Sol's "The First Fire," winning on cultural detail, a tidier causal loop, and a finish that resolves through action rather than a speech. Sol takes the round on sheer clarity, the version to hand someone who wants the machinery laid bare rather than hinted at. Both are good without being great, and the leap in quality over the previous generation is hard to spot.
Test two: from a twig to a lettuce
The second exercise probed associative thinking rather than politics. The instruction: describe a twig, use that description to explain the exploitation of workers and the blind worship of the wealthy, then let the whole thing melt into a description of a lettuce. The point was to see whether the metaphor could carry the argument without the model stepping outside it to narrate what it was up to.
Sol came out of the gate well, describing how twigs form the trunk and keep the tree alive before mapping that onto workers who "build homes they may never afford" and "manufacture goods they can barely buy." A standout sentence notes that "the worker does not merely surrender labor, but imagination as well." Even so, Sol keeps puncturing its own illusion, announcing the comparison outright when it says much of the modern working class is treated the same way, instead of letting the image do the lifting. The lettuce finish never really fused with the rest, so the association fell flat.
Fable took the opposite route, hiding the argument entirely within the object. Its twig "moved water it never drank" and "held leaves it never owned," letting the exploitation surface through plain physical description with no label attached. The cleverer stroke was recasting the fallen twigs as believers, each convinced it is an "early-stage branch" merely enduring "a temporary setback," sure it will still reach the canopy "with hustle and hydration," a neat stand-in for chasing riches that were never on the way. It does overreach here and there, as with "ninety-five percent water and one hundred percent unimpressed," and its ending keeps the metaphor on display rather than dissolving it, calling the vegetable something with "no trunk, no canopy, no upward dream" instead of simply letting it be a lettuce.
The round ends in a tie, and the winner depends on taste. For a reader who wants everything spelled out, Sol is the pick. For one who prefers to uncover the message alone, Fable takes it.
Test three: the rewritten bridge puzzle
The logic round used a fresh puzzle, because the models had started answering the old one the same way every time, a hint that it sits inside their training data rather than being reasoned through on the spot. Taken at face value, four people share one torch to cross a bridge, each walking at a different speed, with "A" quickest at 1 minute and "D" slowest at 10 minutes. The question is simply how long the group needs to get across.
Sol replied 17 minutes without any working, running the classic five-move shuffle from the original riddle: A and B cross, A comes back, C and D cross, B comes back, then A and B cross again. Nothing in its response notices that the prompt never limited how many people can stand on the bridge at once. It reads less like a puzzle solved and more like one pulled from memory. Fable arrived at the same wrong figure of 17 minutes but defended it at length, arguing that sending the two slowest together is the more efficient move and framing the cost of the naive method as an "escort tax." Its reasoning is easier to follow than Sol's and just as off target, since neither model bothered to check whether the constraint it was solving even appeared in the question. For the record, the right answer is 10 minutes, reached by having all four cross together at the pace of the slowest.
Test four: a browser game in one shot
The final trial was a single-shot build: one prompt each for a typing-based shooter, where the player fires by typing words, with whatever came back accepted as is, no follow-ups, no revisions, no do-overs.
Sol appears to have rethought its visual taste, now favouring flat, square interface elements that sit closer to Windows 8.1 than the glossy purple-to-blue diagonal gradient nearly every AI image tool defaults to. It was also the only model to draw the weapon as a bullet-firing typewriter rather than a literal gun, a genuinely fresh choice. Still, the backgrounds stay flat and lifeless across every setup, the crosshair sits still instead of following enemies, and the geometry, from the foes to the gore on kills, looks nearer a late-90s engine than anything modern. It is a clear improvement on GPT-5.5 and more inventive than Opus, just not enough to top Fable in one shot.
Fable won this one comfortably. It delivered music, atmosphere and sound effects that Sol's build left out entirely, and its enemies use a similar geometric-retro look but assembled with more care, closer to Minecraft than to late-90s bargain-bin fare. Its interface is more imaginative and gorier, with real animation in place of static states, and it tracks words per minute, a touch that actually honours the prompt's aim of using the game to practise typing speed. It also includes power-ups, which Sol's version lacks. Benchmarks and professional coders may score it differently, but on this identical prompt the edge clearly belongs to Fable.
So which one should you pick?
Outside of coding, neither model is likely to astonish anyone. Even so, Fable 5 feels like the sturdiest all-rounder, though which model counts as "better" hinges entirely on which of those four things you are paying for. For someone who does not live inside a terminal, drafting emails, asking questions, using a chatbot the way most people do, the trials favour Fable on quality alone, but that verdict gets muddied by something that has nothing to do with intelligence. The pricing gap can be the deal breaker. Sol, Terra and Luna are all bundled into ChatGPT's paid plans with no expiry attached, while Fable 5 is on its third deadline extension in three weeks and flips back to $10/$50 usage credits on July 19 unless Anthropic moves the goalposts once more. If it does revert, paying by the token may lose its appeal.
