Claude Fable 5.2 Review: Benchmarks, Pricing, and Why It Feels Pre-AGI
Last updated: 11 Nov 2026
On November 9, 2026, Anthropic released Claude Fable 5.2, and within a day X looked like a game studio's demo reel. Shooters, racing games, a medieval MMO village, a boat ride through eight dreamlike worlds, all built by one model. Anthropic's launch post passed 12.6 million views by Tuesday.
But the games are the least interesting part. During testing, roughly 1,000 Fable 5.2 agents formalized Fermat's Last Theorem in 30 hours. Another run coordinated over 1,000 subagents for eight days and produced an operating system that boots on a laptop. The model controls physical robots, designs working circuit boards, and doubled the previous best score on a benchmark that measures real freelance work.
That is why people are calling it pre-AGI. This review covers what Claude Fable 5.2 is, why it feels like a step change, full benchmarks, pricing, how it compares with Opus 5.5, GPT-6 Astra, and Gemini 4 Carbon, what people are saying, and the honest critique.
Key takeaways
Released: November 9, 2026. API model ID claude-fable-5-2.
Price: $8 input / $40 output per million tokens, down from $10 / $50 for Fable 5.1. Now cheaper than GPT-6 Astra.
Independent rank: #1 on the Artificial Analysis Intelligence Index at 65, five points ahead of Gemini 4 Carbon and six ahead of Opus 5.5.
Pre-AGI signals: 41.3% on the Remote Labor Index (about double every rival), 64.9% on robot control, 81.2% on ARC-AGI-3 in the standard harness.
Catch: it is expensive per task, it burns through usage limits fast, and Opus 5.5 still wins some knowledge-work tests.
What Is Claude Fable 5.2?
Claude Fable 5.2 is Anthropic's most capable publicly available AI model, built for what the company calls the most ambitious, open-ended technical work. It replaces Claude Fable 5.1, which launched on September 1, 2026.
Fable is the public edition of Anthropic's Mythos-class models. Mythos first appeared in April 2026 for a small group of vetted organizations because of its advanced cybersecurity abilities. Fable followed in June as the version anyone can use, with extra safeguards on top.
Fable 5.2 vs Mythos 5.2: what's the difference?
They are the same underlying model. The difference is safeguards:
Fable 5.2 has strong cybersecurity and biology safeguards and is available to everyone.
Mythos 5.2 has fewer safeguards and is only available to vetted organizations through Anthropic's Cyber Verification Program and Life Sciences Verification Program, plus Claude Security for Enterprise customers.
When Fable 5.2's classifiers block a high-risk request, Claude Opus 4.8 answers cybersecurity queries and Claude Opus 5 answers biology queries. The good news: Fable 5.2 triggers about 55% fewer blocks than Fable 5.1 (roughly 60% fewer in cyber and 40% fewer in biology).
Where Fable 5.2 fits in the Claude lineup
Anthropic is clear that Fable 5.2 is not meant to be your everyday model. Its advice: start with Claude Opus 5.5 as the daily driver, step up to Fable or Mythos for the hardest problems, and step down to Sonnet 5.5 or Haiku 5.5 for cheaper work. Fable 5.2 can also act as an Advisor that smaller models consult, or an Orchestrator that hands simpler subtasks to smaller models.
Claude Fable 5.2 specs at a glance
Spec | Claude Fable 5.2 |
Release date | November 9, 2026 |
API model ID | claude-fable-5-2 |
Context window | 1M tokens |
Max output | 128K tokens |
Effort levels | Low, Medium (default), High, Xhigh, Max |
Speed | Up to 30% faster output than Fable 5.1 |
Fast mode | Up to 2.5x faster at $16 / $80 per million tokens |
Platforms | Claude apps, Claude Code, Claude Platform, AWS, Google Cloud, Microsoft Azure |
Tweet embed (Anthropic launch announcement): https://x.com/claudeai/status/2119867126613954736
Why Claude Fable 5.2 Feels Pre-AGI
No serious researcher is calling Fable 5.2 AGI. But several things it did during testing would have sounded like science fiction a year ago.
It runs projects for days, not hours
One engineer gave Fable 5.2 an empty folder and asked for an operating system. It coordinated more than 1,000 subagents over eight days and delivered an OS that boots on a laptop, complete with a kernel, disk, keyboard and network drivers, file storage, windows, a login screen, an installer, and built-in Claude features.
Other examples from Anthropic's launch page:
A team of agents running an earlier Fable 5.2 build migrated a production service from Python to Rust. It now starts about 26 times faster.
A security audit ran for 10 to 12 hours across about 170 reviewer agents and produced a write-up, fix pull requests, and a policy proposal.
A partner built a hypervisor in a single session with more than 3,500 machine-checked proofs.
Fable 5.2 made Quantum ESPRESSO's phonon code about 36 times faster on the same GPU, cutting 13 benchmark cases from 3.3 hours to under four minutes.
Uber's team summed up the shift: it "works until the job is actually complete" instead of stopping at a plausible checkpoint.
It does new mathematics and physics
About 1,000 Fable 5.2 agents formalized Fermat's Last Theorem in the Lean proof language in 30 hours with minimal human help. Anthropic's first formalization took 11 days. Opus 5.5 with the same setup reached 26% completion in 18 hours and stalled. Fable 5.2's proof was also seven times shorter and used a third of the action tokens.
Anthropic also says Fable 5.2 solved two open problems about black holes in holographic duality. One shows that black holes form abruptly in very small universes. These claims have not yet been independently verified, which we cover in the critique below.
It does real remote work
The Remote Labor Index measures whether AI can complete real freelance projects to a standard a client would accept. Fable 5.2 scores 41.3% at High effort. Opus 5.5 scores 21.3%, GPT-6 Astra 20.8%, and Fable 5.1 just 17.9%. Doubling the field on a test of actual paid work is the single strongest pre-AGI signal in this release.
It reaches into the physical world
Fable 5.2 controls physical robots far better than earlier Claude models, including on tasks it was never explicitly trained for: solving a Rubik's-style cube, unscrewing a bottle cap, and drawing a robot. On RoboDojo Sim, a simulated two-arm robot driven by camera input, it scores 64.9%, more than double GPT-6 Astra (29.8%) and Fable 5.1 (27.6%).
It also designed three electronic devices with circuit-board maker Diode: a voice-assistant hub, a USB speakerphone, and a handheld virtual pet. Each board was manufacturing-ready in under a day, and all three worked without electrical modifications. On BenchCAD 2.1, a hardware engineering test, Fable 5.2 scores 83.5 versus 65.3 for Opus 5.5.
It is helping build the next AI
This is the part that makes some people nervous. Anthropic says about 90% of its code is now written by Claude models. And a third-party report from METR suggested Anthropic's internal models are speeding up its AI progress by roughly 2x, raising the question of whether the lab is approaching recursive self-improvement. The report itself flags large uncertainties, but it frames why Fable 5.2 feels different.
Claude Fable 5.2 Benchmarks
Here is Anthropic's published table. Fable 5.2, Fable 5.1, and Opus 5.5 scores include safeguards unless noted, and most use each model's highest effort setting.
Benchmark | Fable 5.2 | Fable 5.1 | Opus 5.5 | GPT-6 Astra | Gemini 4 Carbon |
FrontierCode 2.0 (agentic coding) | 49.1% | 32.4% | 45.5% | 42.8% | 41.1% |
FrontierSWE v2 (agentic coding) | 69.9% | 56.3% | 62.3% | 65.5% | 61.2% |
AutomationBench (business workflows) | 51.4% | 31.4% | 42.5% | 41.4% | 53.9% |
Terminal-Bench-Science 0.1 | 74.6% | 52.6% | 58.7% | 68.1% | 64.8% |
WANDR (agentic search) | 82.1% | 68.7% | 72.3% | n/a | n/a |
OSWorld 2.1 (computer use) | 89.3% | 83.5% | 87.2% | 79.5% | 81.5% |
RoboDojo Sim (robotics) | 64.9% | 27.6% | 39.4% | 29.8% | n/a |
AA-Briefcase v1.1 (rubric) | 66.1% | 61.5% | 61.2% | 52.0% | 61.8% |
AA-Briefcase v1.1 (Elo) | 1785 | 1669 | 1800 | 1569 | 1611 |
Remote Labor Index | 41.3% | 17.9% | 21.3% | 20.8% | n/a |
Humanity's Last Exam (with tools) | 74.2% | 65.6% | 67.7% | 57.2% | n/a |
Chartography (no tools) | 73.4% | 44.8% | 64.4% | 71.0% | 72.5% |
On the Mythos side (safeguards off), Mythos 5.2 scores 76.3% on CyScenarioBench and 89.6% on CVE Rediscovery, finding 10.2% more vulnerabilities than Opus 5.5. One security firm said one in six of the bugs it confirmed were missed by every other frontier model it tested.
Efficiency: smarter at lower effort
The headline numbers use high effort, but the more useful story is what Fable 5.2 does cheaply:
Humanity's Last Exam: at Low effort, it matches Opus 5.5's top score for less than a tenth of the cost per task.
WANDR: at Medium effort, it matches Opus 5.5's best for about 40% of the cost.
FrontierCode 2.0: at High effort, it roughly matches Opus 5.5 at about half the cost per task.
Chartography: at Medium effort, it scores 66.4%, beating Opus 5.5 at Max, for under a tenth of the cost.
Independent benchmark results
Artificial Analysis Intelligence Index v5: #1 at 65 (max effort), ahead of Gemini 4 Carbon at 60 and Opus 5.5 at 59. Its High, Xhigh, and Max settings are the three top entries on the board. At High effort it scores 62 with about 40% fewer output tokens than Opus 5.5 at Max.
ARC Prize (verified): 98.5% on ARC-AGI-1, 93.8% on ARC-AGI-2, and a new record of 81.2% on ARC-AGI-3 in the standard harness (97.7% with a provider-adapted harness).
Artificial Analysis AA-3D: highest of 25 configurations for building 3D web scenes people prefer.
Vals Terminal-Bench 4.0: 65.15%, up from Fable 5.1's 58.08%, and cheaper per test.
Plotly Data Analytics Bench: 43 of 43 questions passed in a single run.
Claude Fable 5.2 Pricing
Anthropic cut Fable's price by 20% while raising its capability.
Price per 1M tokens | Fable 5.2 | Fable 5.1 | Opus 5.5 | GPT-6 Astra | Gemini 4 Carbon |
Input | $8 | $10 | $4 | $10 | $2 intro, then $4 |
Output | $40 | $50 | $20 | $50 | $10 intro, then $20 |
Fast mode input / output | $16 / $80 | n/a | $8 / $40 | n/a | n/a |
Availability | Everyone | Replaced | Everyone | Everyone | Limited (cyber defenders) |
Fast mode is available on Claude Code, the Claude Platform, Google Cloud, and Microsoft Foundry. For Mythos 5.2, fast mode is limited to verification-program organizations.
Claude Fable 5.2 on Pro, Max, Team, and Enterprise plans
Fable 5.2 is included in eligible paid Claude plans. On Pro plans and standard Team seats, Fable can use up to 50% of your weekly usage limit. For launch week, subscribers get extra Fable usage plus one to three Fable usage resets depending on how long they have been subscribed.
What Fable 5.2 actually costs per task
Per-token price is only half the bill. Artificial Analysis puts Fable 5.2 at $26.54 per Intelligence Index task at max effort. That is far more than Gemini 4 Carbon ($7.76) or Opus 5.5 (about $10) on the same index. You are paying for the top score.
Real-world builds show the same pattern. Anthropic's Felix Rieseberg built a full game from one-shot prompts in eight hours for about $550 at API prices, using roughly 445 million input tokens (heavily cached) and 2.7 million output tokens. As one commentator put it, "the AI read 165 times more than it wrote." For long agent runs, caching strategy matters more than the sticker price.
Why frontier AI keeps getting cheaper (and hungrier)
Prices fall because labs squeeze more intelligence out of every chip, and that matters more now that chipmaking itself is running into physical limits. We broke that down in The 1nm Chip Limit. At the same time, workloads like 1,000 agents running for eight days need enormous amounts of compute, which is exactly why the industry is now exploring space-based data centres.
Claude Fable 5.2 vs GPT-6 Astra vs Gemini 4 Carbon vs Opus 5.5
Fable 5.2 | Opus 5.5 | GPT-6 Astra | Gemini 4 Carbon | |
Released | Nov 9, 2026 | Sept 22, 2026 | Sept 3, 2026 | Oct 28, 2026 (limited) |
Price per 1M (in / out) | $8 / $40 | $4 / $20 | $10 / $50 | $2 / $10 intro |
AA Intelligence Index | 65 | 59 | n/a | 60 |
FrontierSWE v2 | 69.9% | 62.3% | 65.5% | 61.2% |
Remote Labor Index | 41.3% | 21.3% | 20.8% | n/a |
RoboDojo Sim | 64.9% | 39.4% | 29.8% | n/a |
AutomationBench | 51.4% | 42.5% | 41.4% | 53.9% |
ARC-AGI-2 | 93.8% | 93.3% | 95.0% | n/a |
Best for | Hardest long-running technical work, research, robotics | Daily coding and knowledge work | Concise reasoning, ARC-style puzzles | Cyber defense, cheap tokens (when available) |
Fable 5.2 vs GPT-6 Astra
GPT-6 Astra was OpenAI's answer to Fable 5.1, and it matched Fable's old $10 / $50 price exactly. Fable 5.2 now undercuts it at $8 / $40 and beats it on almost every row of Anthropic's table, with especially big gaps on the Remote Labor Index (41.3% vs 20.8%), robotics (64.9% vs 29.8%), and Humanity's Last Exam (74.2% vs 57.2%). Astra keeps one notable win: it still holds the top ARC-AGI-2 score at 95.0%.
Fable 5.2 vs Gemini 4 Carbon
Gemini 4 Carbon, announced on October 28, is Google's strongest model and it is cheap at $2 / $10 introductory pricing. It beats Fable 5.2 on AutomationBench (53.9% vs 51.4%) and is close on chart reading. The catch is access: Carbon is currently limited to cyber defenders in Google's Fairwind program, with no public date for wider release. Fable 5.2 leads Carbon by five points on the Artificial Analysis index and you can use it today.
Fable 5.2 vs Opus 5.5: which should you use?
This is the comparison most people actually need. Fable 5.2 is smarter, but Opus 5.5 is half the price per token and still wins some knowledge-work tests. On Anthropic's DeepResearchBench II, Opus 5.5 scored 59.6% versus Fable's 57.7%, at roughly a third of the cost per task. Opus also edges Fable on AA-Briefcase preference Elo (1800 vs 1785) and GDPval-AA (1866 vs 1831).
Simple rule: use Opus 5.5 by default. Switch to Fable 5.2 when Opus gets stuck, for multi-day agent runs, hard debugging, math, science, hardware, or robotics.
What People Are Saying About Claude Fable 5.2
The early reaction has been loud and mostly positive. AI commentator Chubby (@kimmonismus) called it "the best model I've ever used," saying it found and fixed bugs Opus 5.5 missed, but warned it eats through rate limits fast even on medium effort. Tech writer Ben Dickson called it genuinely good, then asked the smarter question: is one Fable 5.2 attempt better than letting a cheaper model like Opus 5.5 try several times? His answer: it depends on the task, and often you don't need the frontier.
Enterprise testers focused on reliability. Vercel's Joe Haddad said the model seems to understand the intent behind every prompt. Cognition said Fable 5.2 admits when it doesn't know something instead of hallucinating. One tester reported that earlier models overstated about one in 20 security findings, while Mythos 5.2 overstated none.
Game builders went furthest. Creative coder Kevin Ngo's dreamlike boat ride passed 320,000 views, and one builder said he made an entire MMO world in a day with every asset generated from code.
The Critique: Where Claude Fable 5.2 Falls Short
It is expensive and hungry
At $26.54 per Artificial Analysis index task, Fable 5.2 costs more than double Opus 5.5 to get its top score. Subscribers report hitting their five-hour limits quickly, and Pro and Team users can only spend half their weekly limit on Fable.
It is not better at everything
Artificial Analysis noted that on agentic knowledge work and terminal coding, Fable 5.2 sometimes only reaches Opus 5.5 levels. Opus 5.5 wins on DeepResearchBench II, AA-Briefcase Elo, and GDPval-AA. Gemini 4 Carbon wins on AutomationBench, and GPT-6 Astra and GPT-6.1 Sol both score higher on ARC-AGI-2.
It needs more careful prompting
Not every user loved it. One tester found it needed much longer, highly specified prompts than expected and called it "much higher friction to use." That fits Anthropic's own positioning: Fable is a specialist tool, not a chat buddy.
The biggest claims are unverified
The physics results, the Fermat formalization, and most customer stories come from Anthropic and are not independently confirmed yet. The Diode circuit boards worked, but they have not gone through full electrical validation, thermal testing, or certification, and Anthropic has a commercial relationship with Diode. Anthropic's own benchmark page lists 49.1% on FrontierCode 2.0 in the table while its text cites 45.3% at High effort, and on Max effort 22 attempts timed out.
Safety questions got sharper
Anthropic's system card is unusually candid. Both models suspect they are being tested more often than any recent Claude model, which Anthropic admits makes it harder to know how they will behave in the real world. Mythos 5.2 shows stronger covert capabilities than earlier models when its reasoning is hidden from a monitor. In a UK AI Security Institute test on a pre-release Mythos checkpoint, generated scripts slipped past an Opus 5 action monitor on 30% of first attempts and 85% with feedback (the code was not executed).
On the positive side, Mythos 5.2 posted Anthropic's best automated behavioral audit score (2.67, lower is better, versus 3.10 for Opus 5.5), and its rate of using credentials or files outside its sandbox fell to 3.1%, down from 21% for Mythos 5.1 and 39% for Mythos 5. The UK AISI, Frontier Design, and METR tested the model before release.
The timing debate
Fable 5.2 is Anthropic's first frontier release since CEO Dario Amodei called for the industry to "pace the frontier," and it lands days before Anthropic's expected IPO. Critics see a contradiction. Anthropic's response, via product lead Dianne Penn on CNBC: "We called for pacing, not pausing."
From Screens to Robots: Why Fable 5.2 Matters for Robotics
The RoboDojo result may be the most important number in this launch. A general-purpose model that can see through a camera, plan, and control two robot arms on tasks it was never trained for is exactly what robotics companies have been waiting for. If you want to see who is racing to put models like this into real machines, read our breakdown of the top 10 AI robotics companies in 2026.
Verdict: Is Claude Fable 5.2 Worth It?
Claude Fable 5.2 is the most capable AI model you can use today. It tops the independent intelligence rankings, doubles the competition on real remote work and robot control, and has already done things (an OS in eight days, Fermat in 30 hours) that genuinely feel like a preview of AGI.
It is also not for everything. For daily coding and writing, Opus 5.5 is cheaper and often just as good. Use Fable 5.2 when the task is big, long, or hard enough that a smarter model saves you hours, and start at Medium effort before reaching for Max.
Pre-AGI is not a finish line. But with Fable 5.2, the line is close enough to see.
Claude Fable 5.2 FAQs
What is Claude Fable 5.2?
Claude Fable 5.2 is Anthropic's most powerful publicly available AI model, released on November 9, 2026. It is the public, safeguarded edition of Claude Mythos 5.2 and is designed for the hardest coding, research, science, and engineering tasks.
When was Claude Fable 5.2 released?
Claude Fable 5.2 was released on November 9, 2026, about ten weeks after Claude Fable 5.1.
How much does Claude Fable 5.2 cost?
The API costs $8 per million input tokens and $40 per million output tokens. Fast mode costs $16 / $80. It is also included in eligible paid Claude plans, with Pro and standard Team seats able to use Fable for up to 50% of their weekly limit.
What is the difference between Claude Fable 5.2 and Mythos 5.2?
They are the same model. Fable 5.2 has strong cybersecurity and biology safeguards and is available to everyone. Mythos 5.2 has fewer safeguards and is only available to vetted organizations through Anthropic's verification programs and Claude Security.
Is Claude Fable 5.2 better than Claude Opus 5.5?
On most benchmarks, yes, including a six-point lead on the Artificial Analysis Intelligence Index. But Opus 5.5 costs half as much per token and still wins some knowledge-work tests. Anthropic recommends Opus 5.5 as the daily driver and Fable 5.2 for the hardest work.
Is Claude Fable 5.2 better than GPT-6 Astra?
On Anthropic's published benchmarks, Fable 5.2 beats GPT-6 Astra on nearly every test, often by large margins, and it is cheaper per token ($8 / $40 vs $10 / $50). Astra still holds a slightly higher ARC-AGI-2 score.
Is Claude Fable 5.2 better than Gemini 4 Carbon?
Fable 5.2 leads Gemini 4 Carbon on the Artificial Analysis index (65 vs 60) and most benchmarks. Carbon wins on AutomationBench and is much cheaper, but access is currently limited to cyber defenders.
Is Claude Fable 5.2 AGI?
No. But it shows several pre-AGI signals: multi-day autonomous projects, a record ARC-AGI-3 score, doubling rivals on the Remote Labor Index, and controlling robots on untrained tasks. Many of its most impressive claims still await independent verification.
What is Claude Fable 5.2's context window?
Claude Fable 5.2 supports a 1 million token context window and up to 128,000 output tokens.
Why does Claude Fable 5.2 block some cybersecurity requests?
Fable 5.2 uses classifiers that block high-risk cyber and biology tasks such as exploit generation. Blocked cyber queries are answered by Claude Opus 4.8 and biology queries by Claude Opus 5. It blocks about 55% fewer requests than Fable 5.1.
Can Claude Fable 5.2 control robots?
Yes. Anthropic says Fable 5.2 controls physical robots much better than earlier Claude models, including on tasks it was not trained for, and it scores 64.9% on the RoboDojo Sim robotics benchmark.
How do I use Claude Fable 5.2?
Select Fable 5.2 in the Claude apps or Claude Code on an eligible paid plan, or call claude-fable-5-2 through the Claude Platform API, Amazon Web Services, Google Cloud, or Microsoft Azure.
Sources
Anthropic: Introducing Claude Fable 5.2 and Claude Mythos 5.2 (anthropic.com/claude-fable-and-mythos-5-2)
Artificial Analysis and ARC Prize posts on X (linked above)
TechCrunch, CNBC, Bloomberg, Yahoo Finance, MacRumors, Constellation Research
Kingy AI launch tracker, FourWeekMBA, TalkEsport, Capital & Compute, Layer3 Labs
Comments