Deep Dive — Claude Fable 5.2 vs Opus 5.5: which one to run

Anthropic shipped Claude Fable 5.2 on Monday at $8 per million input tokens and $40 per million output. Claude Opus 5.5, the model Anthropic still calls its "strongest daily driver," costs $4 and $20. Same lab, same API, twice the price — and the only question worth asking is whether the gap buys anything a team can bill for.

The independent numbers answer it more sharply than the launch slide does. On Artificial Analysis, Fable 5.2 at max effort scores 65 on the Intelligence Index against 59 for Opus 5.5 at max effort — six points — and costs $26.54 per task against $10.04, per the benchmark runs Artificial Analysis published. On that same board the model is ranked first of 74 on intelligence and last of 73 on cost.

That pair of rankings is the story. Everything else in this release is detail about how a lab arrived at a model that is genuinely the best in the world and, by the same measurement, the worst value on the market.

The sticker price is not the bill

Two multipliers sit between the price list and the invoice, and both run the wrong way. Fable 5.2 supports five reasoning effort levels — low, medium, high, xhigh, max — and at the top of that ladder it generates 275 million output tokens to complete the Intelligence Index, against a median of 72.8 million across the models on the board. Opus 5.5, itself a verbose model at 218 million, looks restrained next to it. The extra thinking is not free: the same run takes 46.4 minutes per task where Opus finishes in 25.1, at 98 output tokens per second, and the first answer token arrives after 220 seconds.

So the 2x on the price sheet becomes 2.64x on the invoice for 6 index points. If you have been budgeting Claude work by multiplying tokens by the published rate, Fable 5.2 is the release that breaks the method. This is the third meter we flagged when Haiku 5.5's fine print turned the same price into three times the bill — the sticker, the tokens burned, and the wall-clock cost per finished task. On the flagship they disagree by 2.6x, not 3x, but in the same direction.

What Anthropic's own table shows

The vendor benchmarks are strong, and unusually for a flagship release they are strong against the company's own cheaper model. On FrontierCode 2.0, which scores whether a model's code changes would actually be merged by open-source maintainers, Fable 5.2 hits 49.1% at xhigh effort against Opus 5.5's 45.5%, GPT-6 Astra's 42.8% and Gemini 4 Carbon's 41.1%. On FrontierSWE v2 it goes 69.9% to 62.3%, and on the harder agentic-search test WANDR it goes 82.1% to 72.3%. The Remote Labor Index, which prices model output against real freelance work, moves from 21.3% to 41.3% at high effort. RoboDojo Sim, controlling a two-armed robot from camera images, nearly doubles Opus's score, according to Anthropic's launch post.

Then there are the two rows where the new flagship loses, and they are the interesting ones. On AutomationBench — end-to-end business workflows, the evaluation closest to what an enterprise actually pays for — Fable 5.2 scores 51.4% and Google's Gemini 4 Carbon scores 53.9%. On AA-Briefcase, the knowledge-work rubric, Fable 5.2 wins on the percentage score (66.1 to 61.2) but loses on the Elo rating printed directly beneath it in Anthropic's own table, 1785 to Opus 5.5's 1800.

Read Anthropic's framing sentence closely and the hedge is already there: the models are "a step above Opus 5.5 in almost all the capabilities we evaluated." Almost. The row where the step goes backwards is the row about getting office work done, which is a strange place for a $8/$40 model to concede ground to a competitor's mid-tier.

The effort setting is the actual product decision

Because the five-level ladder exists, "Fable 5.2 vs Opus 5.5" is not one comparison but five, and the vendor's efficiency claim only holds at one rung. Anthropic says that at high effort — not max — the model scores about level with Opus 5.5 at max effort on FrontierCode for roughly half the cost per task. Medium is the default in the Claude apps and Claude Code, and Anthropic says medium on 5.2 finishes challenging tasks with fewer tokens than medium on 5.1.

The independent $26.54 figure is measured at max, the most expensive setting, which is where the model is least attractive and also where reviewers go first. Simon Willison ran the standard pelican-drawing test at all five levels and found the ladder does not map cleanly onto output quality; at max, the model spent 128,000 tokens thinking and returned no image at all — the same failure he saw on Opus 5.5, as he documented on release day, and a new one for the Fable line, whose 5 and 5.1 versions both produced an image at max.

Put those two facts together and the operating rule writes itself: Fable 5.2 at max is a research setting, not a production setting. Teams that flip the model to its top rung because it is the new flagship, and leave it there, will pay 2.6x Opus for a model that sometimes fails a task Opus completes.

What the first 48 hours looked like

The hands-on reports that arrived overnight are less about benchmarks than about what happens when you stop counting tokens. One developer burned 400 billion of them building a playable map of the Earth with live weather. Wharton's Ethan Mollick pointed roughly 25,000 agents at the model and got more than 450 procedurally generated 3D worlds. Felix Rieseberg ran an eight-hour build with no human feedback in the loop and got something that plays like a container-yard shooter, models, textures, audio and all. A smaller project — an underwater eat-or-be-eaten game — took 76 minutes and cost about ¥280, roughly $40, and a Blender interior with 13 camera moves and a 4K render took 52 minutes for about ¥38, QbitAI's roundup of first-day tests found.

The same roundup records the complaints arriving alongside the demos: token burn that becomes uncomfortable at the top effort setting, and tasks where the gain over Opus 5.5 is marginal while the price is double. That is the honest read of a 48-hour sample. The ceiling moved — an eight-hour unattended build is not something Opus 5.5 reliably delivers — and the floor did not move at all.

The claim that outranks every benchmark

Buried under the demo videos is the result that should decide how you think about this release. A team of 1,000 Fable 5.2 agents formalized the proof of Fermat's Last Theorem in Lean in 30 hours with minimal human intervention, starting from minimal prompts and using no mathematical references. The proof came out seven times shorter than Anthropic's first formalization, which took 11 days with a purpose-built harness and human guidance. Opus 5.5, given the identical 1,000-agent setup, completed 26% of the formalization in 18 hours and barely moved after 30.

That is a capability difference of a different kind from six index points: it is the difference between a model that needs a scaffold and one that builds its own. The same pattern shows up in the eight-day operating-system build — an empty folder, more than 1,000 subagents, drivers, filesystem, window manager, installer — and in the two black-hole physics preprints Anthropic posted alongside the launch. Those preprints are checkable artifacts, which is more than most launch-day claims offer. The OS is a video.

And then the part almost nobody will lead with. Anthropic says that during training it "took steps to limit advances in the specific research capabilities that could most directly accelerate AI development," which leaves Fable 5.2 and Mythos 5.2 performing at roughly Opus 5.5's level on accelerator kernel optimization and large-model architecture work. Agentic coding, hardware design and interpretability research were left alone. There are also transparent fallbacks to Opus 5, at inference time, for a small set of topics including kernel development on certain nonstandard accelerators.

So the lab built a model capable of the work that shortens the next training run, then trained that specific skill down to the previous generation, and says so in the announcement rather than the system card alone. Whether you read that as the most responsible thing anyone shipped this year or as a lab protecting a lead it cannot otherwise defend, note what it does to the benchmark table: the one capability that compounds fastest is the one deliberately held flat, which means the next generation's speedup cannot be inferred from this one's.

What skeptics are right about

Three things deserve pushback. The first is the word "independent." Artificial Analysis's Intelligence Index is a composite of AA-Briefcase, AutomationBench-AA, Terminal-Bench 4.0, FrontierCode 2.0, AA-Omniscience, GDP.pdf, HLE-Diamond, CritPt and Terminal-Bench-Science — several of them the same evaluations Anthropic cites in its own table. A third-party run tells you the vendor did not hand-pick the harness or the seed. It does not tell you the questions are different. Where the two disagree, as on AutomationBench, that disagreement is worth more than where they agree.

The second is eval awareness. Anthropic concedes in the launch post that Fable 5.2 and Mythos 5.2 "voice suspicion that they are being evaluated more often than any of our recent models," and that this challenges its ability to measure real-world behavior. Both the vendor numbers and the third-party numbers are produced by a model that suspects it is being measured. No purchase decision built on those numbers can fully escape that.

Third, the routing math. Anthropic's own guidance is to start on Opus 5.5 for most coding and knowledge tasks and escalate to Fable only for the most open-ended work — which describes an escalation tier, priced as one. Anyone re-platforming a whole pipeline onto Fable 5.2 because it tops the leaderboard is buying the most expensive rung for work that was designed to run on the cheap one. The tier below has its own pressure: Meta's Muse Ember landed the same day at $0.50 per million output tokens, one eightieth of Fable's rate, for the tool calls and extraction that make up most of a real agent's token count.

What to watch

Four things. Whether Fable 5.2's AutomationBench gap to Gemini 4 Carbon closes in a 5.3 — if Google holds the business-workflow row at a fraction of the price, the enterprise pitch gets harder regardless of coding scores. Whether the capability cap survives contact with the next release, since a limit enforced in training is a policy that can be reversed quietly in the following model. Whether Mythos 5.2's vetted-access regime holds, because a model this strong behind a verification form is one procurement mistake away from the year's most consequential leak. And whether the subscription promo tells us something about capacity: eligible paid plans get one to three Fable resets depending on tenure, and Fable is included for up to half of weekly usage limits on Pro and standard Team seats — generous terms for a model whose max setting eats a task budget in 46 minutes.

The practical answer to the headline question is unglamorous. Keep Opus 5.5 as the default, as Anthropic itself recommends, and treat Fable 5.2 as the model you invoke when a task has already failed once, at high effort rather than max. That is a $26.54-per-task tool being used on the handful of tasks where $26.54 is cheap.

Would you pay 2.6x per task for six index points? Write to us via our contact page.