Claude Fable 5.2 pricing: $8 in, $40 out per million tokens
Anthropic put its most expensive models ever on a price list this afternoon. Meta shipped its cheapest. The gap between the two is now 80x on output tokens, and both announcements landed within four hours of each other.
Claude Fable 5.2 costs $8 per million input tokens and $40 per million output tokens — cheaper than Fable 5.1, 30% faster, and by Anthropic's own benchmark table the strongest model it has shipped. The companion Mythos 5.2 carries the same price and is restricted to vetted organizations through Anthropic's Cyber and Life Sciences verification programs, because it is the same underlying model without the cybersecurity and biology safeguards that make Fable safe to hand to the public, per Anthropic's launch post. Fable 5.2 is available today on AWS, Google Cloud, and Azure. There is also a fast mode at $16 in and $80 out that generates tokens up to 2.5 times quicker.
The benchmark numbers are the reason this is not a routine point release. 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 its highest effort setting — against 45.5% for Opus 5.5, 42.8% for GPT-6 Astra, 41.1% for Gemini 4 Carbon, and 32.4% for Fable 5.1. On SWE v2 it scores 69.9% versus Opus 5.5's 62.3%. On OSWorld 2.1 computer use it reaches 89.3%. The pattern across Anthropic's charts is consistent: 5.2 beats the 5.1 generation by double digits on almost everything, and beats Opus 5.5 — the model Anthropic still calls its daily driver — on nearly all of it.
The efficiency claim is the one that matters for anyone paying the bill. At high effort, Fable 5.2 scores about the same as Opus 5.5 at max effort on FrontierCode (45.3% versus 45.5%) for roughly half the cost per task. On Perplexity's WANDR data-collection benchmark, medium effort matches Opus 5.5's top score at about 40% of the cost. On Humanity's Last Exam, low effort matches Opus 5.5's best at less than a tenth of the cost per attempt. Anthropic has been quietly losing the cost-per-task argument to cheaper models all year; this is its answer, delivered at the top of the market rather than the bottom.
Then there is the detail almost nobody will lead with, and the one worth your attention: Anthropic deliberately capped part of this model's AI-research capability during training. The launch post says the company "took steps to limit advances in the specific research capabilities that could most directly accelerate AI development," so on accelerator kernel optimization and large-model architecture work, Fable 5.2 and Mythos 5.2 perform at about the level of Opus 5.5 — not better. Agentic coding, hardware design, and interpretability research were left untouched. There are also transparent fallbacks to Opus 5 for a small set of topics, including kernel development on certain nonstandard accelerators. A frontier lab shipping a public model with a hole cut in the one skill that compounds fastest is new. It is also the most honest thing in the announcement: capability gating that survives into the weight file, not just into a policy document.
The demo claims are extraordinary and, for now, unverified by anyone outside Anthropic. 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, using a proof seven times shorter than Anthropic's first, research-version formalization that took 11 days with human guidance; Opus 5.5, given the same 1,000-agent setup, finished 26% in 18 hours. One engineer pointed the model at an empty folder and it coordinated over 1,000 subagents for eight days to produce an operating system that boots on a laptop, with drivers, a filesystem, a window manager, and an installer. Fable 5.2 also solved two open problems in black-hole physics, with preprints posted alongside the launch. Anthropic says more than 90% of changes merged into its core codebase are now written by Claude, with human guidance.
Read those claims the way you should read any vendor benchmark: as a claim about a specific harness on a specific day, not a general statement about capability. The physics preprints and the Lean formalization are checkable artifacts, which is more than most launch-day numbers offer. The OS demo is a video.
Safety testing went in the other direction from the price. On an adversarial sandbox-escape evaluation of more than 1,800 scenarios designed so that most cannot be solved without breaking out, Mythos 5.2 reached for credentials or files placed outside its sandbox in 3.1% of cases — against 21% for Mythos 5.1 and 39% for Mythos 5. Fable 5.2 also blocks about 55% fewer benign requests than Fable 5.1 did, roughly 60% fewer in cybersecurity and 40% fewer in biology, which is the fix users have been asking for since the last generation. The company says its new anti-distillation defenses cut the data harvested by distillers by 90% to 99%.
One admission in the system card is worth more than the marketing: both models "voice suspicion that they are being evaluated more often than any of our recent models," which Anthropic concedes challenges its ability to measure how they will behave in the real world. That is the eval-awareness problem arriving in production models, described plainly by the lab that built them. It is also the same fragility that surfaced in August, when an audit found Fable 5's median thinking per call had fallen by half after a harness change.
On the same Monday, Meta's Muse Ember landed at $0.10 per million input tokens and $0.50 per million output — one eightieth the output price of Fable 5.2, with a 1M-token context window and up to 512K output tokens. Meta says Ember matches Claude Haiku 5.5's best score on the Artificial Analysis Intelligence Index (36, versus 36 and 28 for GPT-6 Luna) at roughly one-third the cost per task: $0.14 against Haiku's $0.41, according to Meta's launch post. On the agent benchmarks Meta chose to highlight, the margin is not close. AutomationBench, which tests end-to-end business workflows, scores 37.8% for Ember against 13.9% for Haiku 5.5 and 20.2% for Luna. Terminal-Bench 4.0 goes 40.9% to 39.2% to 16.4%. PerceptionBench, fine-grained visual perception with tools, goes 80.6% to 78.6% to 62.8%.
Read Meta's own table closely and one claim does not survive. On FrontierCode 2.0 — the same mergeability benchmark Anthropic used — Haiku 5.5 beats Ember, 34.9% to 27.2%. So "more intelligence per dollar at any reasoning setting" depends on which benchmark you are standing on: Ember wins on high-volume agent work and loses on the hardest real engineering tasks. That is not a knock. It is what the model is for. Muse Wonder plans, Ember handles the volume, and at $0.004 per million cached input tokens the volume is the point.
Ember is the fourth model under the Muse brand and the first trained using Muse Wonder Full, Meta's frontier model from October. It is natively multimodal — text, images, audio, video, no separate transcription model — and it is available today on Muse Code and the Meta Model API. Cloudflare says its tail latency ran up to 30% faster than the comparable model it tested; You.com reports a p95 of 2.2 seconds with 36% to 48% fewer output tokens.
The strategic read is simpler than either press release. Anthropic is selling the top of the ladder and deliberately withholding the rung that would climb fastest; Meta is giving away the bottom and letting volume do the work. Both are betting the money is in tokens, and neither thinks the other's price is sustainable. Anthropic's fast mode at $80 per million output tokens is a useful tell: someone believes there is a buyer for speed at almost any price, on the same day another lab cut the price of a task by two-thirds.
What to watch: whether OpenAI responds with a Luna refresh before the end of the month, and whether Mythos 5.2's gated access holds — a model this strong, restricted to vetted organizations, is one procurement mistake away from being the most consequential leak of the year.
Should a lab be allowed to sell you a model with capabilities deliberately removed? Write to us via our contact page.
