Fable 5.2 & Mythos 5.2

Anthropic cut the price of its top model for the first time, and at high effort Claude Fable 5.2 now matches Opus 5.5's best coding score for about half the cost per task. Anthropic shipped Fable 5.2 and its unfiltered twin, Claude Mythos 5.2, on November 9, 2026.

Fable 5.2 and Mythos 5.2 are one model sold two ways, the same split Anthropic has used since June. Fable 5.2 is for everyone: it runs on the Claude API and on AWS, Google Cloud and Azure from day one, with safety classifiers that block most cybersecurity and research-biology tasks and hand them to Opus 4.8 (cyber) or Opus 5 (biology) instead. Mythos 5.2 is the same model with those classifiers off, available only to organizations in Anthropic's Cyber and Life Sciences Verification Programs, plus Claude Enterprise customers through Claude Security. Both cost $8 per million input tokens and $40 per million output, down from Fable 5.1's $10 / $50 but still exactly double Opus 5.5's $4 / $20. Both have a 1M-token context window and 128K of output, and both get an optional fast mode that generates up to 2.5 times faster for $16 / $80. Subscribers get a launch promotion: Pro and standard Team seats get Fable in their plan for a week (up to half their weekly limit, through November 16), and Max, Ultra and premium seats get one to three Fable resets depending on how long they've paid. Free gets nothing.

The evidence is a 15-row table against Fable 5.1, Opus 5.5, GPT-6 Astra and Gemini 4 Carbon, where Fable 5.2 comes first on everything except Gemini 4 Carbon's AutomationBench score and Opus 5.5's head-to-head work-product rating (see ## The benchmarks), plus a pile of demos. A team of 1,000 Fable 5.2 agents formalized the proof of Fermat's Last Theorem in 30 hours (Anthropic's first attempt took an unreleased research model 11 days), one engineer had it coordinate more than 1,000 subagents for eight days to build an operating system that boots on a laptop, and it designed three working gadgets with their own circuit boards (a voice-assistant hub, a USB speakerphone and a handheld virtual pet) that worked without any electrical fixes. Anthropic also says it solved two open problems in black-hole physics, with preprints. The catches are in the same post. Anthropic's own advice is to start on Opus 5.5 and only move up to Fable when that falls short, it deliberately held back the model's skill at AI research work (kernel tuning, model architecture) to about Opus 5.5's level, Fable 5.2 still requires 30-day data retention, and every API account now gets a 400 error if code edits earlier turns of a conversation. The launch also landed the day after a Hacker News thread full of people convinced Opus 5.5 had been made worse ahead of the launch. Anthropic hasn't said anything about that.

What's new

Most of the changes are about cost and the filters. The model also got a lot better at the long jobs where it runs hundreds of agents at once, which is what Anthropic spent most of the launch post showing off.

• The first Fable price cut. $8 / $40 is 20% under Fable 5.1, and cache reads drop to $0.20 per million. Anthropic says outputs come up to 30% faster than 5.1, and at Medium effort (the default in the Claude apps and Claude Code) it finishes hard tasks with fewer tokens than Fable 5.1.

• Fewer pointless blocks. Anthropic measured about 55% fewer classifier blocks than Fable 5.1 across claude.ai, Claude Code and Cowork, about 60% fewer on cybersecurity and 40% fewer on biology. When it does block, you still get an Opus answer instead.

• A boss for other models. Anthropic pitches Fable 5.2 as an Advisor (Haiku, Sonnet or Opus consult it on hard problems) and an Orchestrator (it hands simple work to smaller models). The Fermat run and the operating system build were both swarms of a thousand-plus agents.

• A big jump in robotics. On RoboDojo, a test of controlling a simulated two-armed robot from camera images, Fable 5.2 more than doubles GPT-6 Astra and Fable 5.1, and Anthropic's video shows it solving a puzzle cube and unscrewing a bottle cap with real robots.

• Mythos 5.2 finds more bugs, and fewer fake ones. Anthropic says it found 10.2% more known vulnerabilities than Opus 5.5 and 14.4% more than Mythos 5.1, Mozilla says it roughly doubled Mythos 5.1's bug finds on Firefox, and one tester found earlier models overstated about one in twenty findings while Mythos 5.2 overstated none.

The benchmarks

Anthropic ran every model at its highest available effort unless a row says otherwise, with safeguards on for Fable 5.2, Fable 5.1 and Opus 5.5. The cyber and protein rows show Mythos 5.2 against Mythos 5.1 and an Opus 5.5 with no safeguards. On CyScenarioBench, Anthropic notes each lab's score comes from a different set of tasks. RoboDojo is Anthropic's own run of every model. BenchCAD is scored 0 to 100 with partial credit.

Anthropic's numbers

• FrontierCode 2.0 (code maintainers would merge): 49.1% (xhigh). +3.6 over Claude Opus 5.5 (45.5%), +6.3 over GPT-6 Astra (42.8%), +8.0 over Gemini 4 Carbon at medium (41.1%), +16.7 over Fable 5.1 at xhigh (32.4%).

• FrontierSWE v2 (frontier software engineering): 69.9%. +4.4 over GPT-6 Astra (65.5%), +7.6 over Claude Opus 5.5 (62.3%), +8.7 over Gemini 4 Carbon (61.2%), +13.6 over Fable 5.1 (56.3%).

• AutomationBench (business workflows, Zapier): 51.4%. −2.5 under Gemini 4 Carbon (53.9%), +8.9 over Claude Opus 5.5 (42.5%), +10.0 over GPT-6 Astra (41.4%), +20.0 over Fable 5.1 (31.4%).

• Terminal-Bench-Science 0.1 (scientific research in a terminal): 74.6%. −0.7 under Mythos 5.2 (75.3%), +6.5 over GPT-6 Astra (68.1%), +9.8 over Gemini 4 Carbon (64.8%), +15.9 over Claude Opus 5.5 (58.7%), +22.0 over Fable 5.1 (52.6%).

• OSWorld 2.1 (using a computer, offline set, partial credit): 89.3%. +2.1 over Claude Opus 5.5 (87.2%), +5.8 over Fable 5.1 (83.5%), +7.8 over Gemini 4 Carbon (81.5%), +9.8 over GPT-6 Astra (79.5%).

• WANDR (big data-collection jobs, relaxed score): 82.1%. +9.8 over Claude Opus 5.5 (72.3%), +13.4 over Fable 5.1 (68.7%).

• Remote Labor Index (paid freelance projects): 41.3% (high). +20.0 over Claude Opus 5.5 (21.3%), +20.5 over GPT-6 Astra (20.8%), +23.4 over Fable 5.1 (17.9%).

• Humanity's Last Exam (with tools): 74.2%. +6.5 over Claude Opus 5.5 (67.7%), +8.6 over Fable 5.1 (65.6%), +17.0 over GPT-6 Astra (57.2%).

• Chartography (reading charts, no tools): 73.4%. +0.9 over Gemini 4 Carbon (72.5%), +2.4 over GPT-6 Astra (71.0%), +9.0 over Claude Opus 5.5 (64.4%), +28.6 over Fable 5.1 (44.8%).

• RoboDojo Sim (controlling a robot): 64.9%. +25.5 over Claude Opus 5.5 (39.4%), +30.7 over Gemini 4 Argon at high (34.2%), +35.1 over GPT-6 Astra (29.8%), +37.3 over Fable 5.1 (27.6%).

• BenchCAD 2.1 (hardware design, with tools): 83.5. +18.2 over Claude Opus 5.5 (65.3), +39.1 over Fable 5.1 (44.4).

• AA-Briefcase v1.1 (work products, Elo): 1785. −15 under Claude Opus 5.5 (1800), +116 over Fable 5.1 (1669), +174 over Gemini 4 Carbon (1611), +216 over GPT-6 Astra (1569).

• CVE Rediscovery (Mythos 5.2): 89.6%. +8.3 over Claude Opus 5.5 (no safeguards) (81.3%), +11.3 over Mythos 5.1 (78.3%).

• Irregular CyScenarioBench (Mythos 5.2): 76.3%. +8.7 over Claude Opus 5.5 (no safeguards) (67.6%), +14.6 over Mythos 5.1 (61.7%), +15.3 over GPT-6 Cyber (61%).

• De novo binder design (Mythos 5.2): 85.1%. +2.5 over Claude Opus 5.5 (no safeguards) (82.6%), +6.0 over Mythos 5.1 (79.1%).

Independent numbers

ARC Prize verified Fable 5.2 on its puzzle tests within 15 minutes of launch (leaderboard). On ARC-AGI-3 it scored 81.2% on ARC's standard harness, a new high there, and 97.7% with the provider's own harness, just under GPT-6 Astra's 99.9%. The full ARC-AGI-3 runs cost about $21,900 and $18,600. Artificial Analysis hasn't published its own numbers yet. The AA-Briefcase rivals above match Artificial Analysis's board exactly, so Anthropic is quoting that run; Fable 5.2's 1785 is Anthropic's figure until Artificial Analysis posts its own.

• ARC-AGI-3 (standard harness, ARC Prize): 81.2%. +18.5 over GPT-6 Astra (62.7%), +51.0 over Claude Opus 5 (30.2%).

• ARC-AGI-2 (visual puzzles, ARC Prize): 93.8%. −1.2 under GPT-6 Astra (95.0%), +1.3 over GPT-5.6 Sol (92.5%), +3.4 over Claude Opus 5 (90.4%).

• ARC-AGI-1 (visual puzzles, ARC Prize): 98.5%. ties GPT-6 Astra (98.5%), +1.0 over Claude Opus 5 (97.5%).

See how it ranks. Every score above is on the Handy AI benchmarks board, where you can see where Claude Fable 5.2 lands against every model we track, open and closed.

First impressions

The models are a couple of hours old, so these are mostly reactions to the numbers and the price. The Hacker News launch thread passed 130 points inside 20 minutes.

The positives

Ethan Mollick (@emollick), the Wharton professor who writes One Useful Thing, had early access and went further than he usually does.

Mollick is careful with the AGI word, so "a sort of Jagged AGI" from him carries more weight than the same line from a hype account. His examples (starting a business, academic research, hacking hardware) line up with what Anthropic demoed.

ARC Prize (@arcprize), which runs the ARC-AGI puzzle tests, posted verified scores before most people had read the announcement.

This is the one outside check so far, and it holds up: a new high on ARC-AGI-3's standard harness, 18.5 points over GPT-6 Astra. It's also a reminder of cost, since each full ARC-AGI-3 run came to roughly $20,000.

awongh on Hacker News went straight to the row I care about most.

Doubling Remote Labor Index (RLI) - at 41.3% compared to around 20% for the other models - I wonder what kinds of new agentic work this is going to unlock. (comment)

Remote Labor Index scores models on real freelance projects, and every other model on Anthropic's table sits around 20%. Twice the field on paid work is the number that would make a business switch, if it holds up outside Anthropic's own run.

The negatives

skeledrew on Hacker News had the most repeated complaint in the thread, and it's about Opus 5.5.

Damn people were right about 5.5 being nerfed in preparation for a version release. This is not at all a good look. (comment)

There's no proof Opus 5.5 was changed, and Anthropic says it doesn't alter capabilities at serving time. But a launch whose pitch is "a step above Opus 5.5" landing a day after a front-page thread about Opus 5.5 getting worse is bad timing, and Anthropic's comparison numbers for Opus 5.5 come from its own runs.

guybedo on Hacker News read the fine print on those resets.

it seems these are not resets of your weekly quota but resets of the Fable limit. That means you can lift your Fable max 50% usage and you'll just end up burning through your quota faster. (comment)

Anthropic's help page backs that up. A reset gives back what you already spent on Fable, your limits stay where they were, and Fable burns through usage faster than any other Claude model.

iamdelirium on Hacker News flagged how the headline coding row was run.

FrontierCode 2.0 is such a poor benchmark for Gemini since it's the only model forced to not use its own harness but Devin CLI. (comment)

Anthropic also ran Gemini 4 Carbon at medium effort on that row while everyone else got their highest setting, so the gap to Gemini there is the least trustworthy number in the table.

Jake's take

Fable's back! And it's not 5.5 as most expected, but 5.2 (although, using it, you wouldn't know that). In typical Anthropic fashion, the 0.1 bump is much bigger than it leads on. And while the price is still double Opus 5.5, the cut looks significant, especially considering that Anthropic is claiming that at high effort it matches Opus 5.5's best FrontierCode score for half the cost per task. Is Fable becoming affordable? (Also, Remote Labor Index at 41.3% against about 20% for everyone else is crazy).

At the end of the day, Fable 5.2 makes the case for using Fable for the big guns again. Opus 5.5 has been such a performant model that many have been saying Fable 5.1 is basically irrelevant. Now that's changing.