Anthropic released Claude Fable 5.2 and Claude Mythos 5.2 on 9 November. They are, in the words of Anthropic’s launch post, “the same underlying model”, sold under two names with two levels of access. Fable 5.2 costs $8 per million input tokens and $40 per million output tokens, down from $10 / $50 for Fable 5.1, according to the Claude Fable product page. On the Artificial Analysis Intelligence Index, Fable 5.2 at max effort scored 65, the highest the index has recorded.
I read the launch post, the model docs, the product page and Artificial Analysis’ pages on 11 November. I did not run the model or any benchmark. Everything below is either Anthropic’s claim or a third party’s measurement, and I’ll say which.
The line I keep coming back to is not a benchmark. It sits in the “when to use” part of the launch post: for most coding and knowledge tasks, start with Claude Opus 5.5. A frontier lab telling you not to default to its frontier model is a pricing and routing story, not just a leaderboard story.
Why is one model sold under two names?
Same weights, different wrapping. Fable 5.2 ships with strong cybersecurity and biology safeguards, which is why Anthropic can offer it to all users. Mythos 5.2 is the unrestricted version, available only to vetted organizations in the expanded Life Sciences and Cyber Verification Programs, per the launch post.
The safeguards are classifiers. According to Anthropic, they block high-risk cyber work (pen testing, exploit generation, binary-based vulnerability work) and most research-biology tasks. Fable 5.2’s safeguards “will block most cybersecurity-related tasks”, the company says plainly.
For builders, what matters is what happens next. A blocked query gets answered by Opus 4.8 (cyber) or Opus 5 (biology) instead. So a Fable app that touches security topics will sometimes return answers from an older model. Check how your logs record which model actually answered.
In early testing, Fable 5.2 produced about 55% fewer blocks than Fable 5.1, with about 60% fewer on cyber and about 40% fewer on bio, per the launch post.
Mythos 5.2 is Anthropic’s most capable cyber model. On the internal CVE Rediscovery eval, an October 2026 run over 1,829 CVEs, it found 10.2% more vulnerabilities than Opus 5.5 and 14.4% more than Mythos 5.1, Anthropic reports. A tester working on Firefox is quoted saying Mythos 5.2 “roughly doubled the vulnerability discovery metrics of Mythos 5.1”. The footnote is the honest bit: on Anthropic’s own evals, vulnerability identification is roughly on par with Opus 5.5. Claude Security for Enterprise now includes Mythos 5.2.
How big is the jump, really?
Here is Anthropic’s own launch table, at the highest available effort, with safeguards on unless noted. Source: Anthropic’s launch post, with prices from the model docs, the product page and Google.
| Fable 5.2 | Fable 5.1 | Opus 5.5 | GPT-6 Astra | Gemini 4 Carbon | |
|---|---|---|---|---|---|
| FrontierCode 2.0 | 49.1% (xhigh) | 32.4% (xhigh) | 45.5% | 42.8% | 41.1% (medium) |
| FrontierSWE v2 | 69.9% | 56.3% | 62.3% | 65.5% | 61.2% |
| AutomationBench | 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% |
| OSWorld 2.1 (offline, partial) | 89.3% | 83.5% | 87.2% | 79.5% | 81.5% |
| RoboDojo Sim | 64.9% | 27.6% | 39.4% | 29.8% | n/a |
| AA-Briefcase v1.1 (Elo) | 1785 | 1669 | 1800 | 1569 | 1611 |
| Remote Labor Index | 41.3% (high) | 17.9% | 21.3% | 20.8% | n/a |
| HLE (with tools) | 74.2% | 65.6% | 67.7% | 57.2% | n/a |
| Price per 1M in / out | $8 / $40 | $10 / $50 | $4 / $20 | n/a | $2 / $10 intro (Google) |
This is Anthropic’s table, so read it as a vendor’s table. RoboDojo was run internally. CyScenarioBench is left out because each lab’s tasks differ. The Gemini 4 Carbon figure is the introductory API price per Google, while Artificial Analysis lists Carbon at $4 / $20.
The big jumps first. RoboDojo Sim went from 27.6% to 64.9%, more than double GPT-6 Astra’s 29.8%, on Anthropic’s own run. Remote Labor Index rose from 17.9% to 41.3% at high effort. Terminal-Bench-Science 0.1 moved from 52.6% to 74.6%. On FrontierCode 2.0, the xhigh score of 49.1% is the highest of any released model Anthropic tested, according to the launch post.
Now the rows where it does not lead. Even inside Anthropic’s own table, Gemini 4 Carbon wins AutomationBench, 53.9% to 51.4%. And on AA-Briefcase v1.1, Opus 5.5 scores an Elo of 1800 against Fable 5.2’s 1785. The cheaper sibling wins the office-work row.
Artificial Analysis runs its own versions, and the picture is similar. Its index has Fable 5.2 (max) at 65 and Gemini 4 Carbon (high) at 60. Yet Carbon (high) beats Fable 5.2 (max) on AA’s Terminal-Bench 4.0, 65.2% to 59.6%, on AutomationBench-AA, 77.8% to 76.4%, and on GDP.pdf, 28.4% to 24.2%. Fable leads on HLE-Diamond (80.2% to 69.1%), AA-Omniscience (67 to 54) and FrontierCode-AA 2.0 (48.8% to 40.6%). AA also lists Fable 5.2 as text and image input only, against text, image, speech and video for Carbon.
First place on an aggregate index is not first place on every row. If your workload looks like terminal automation, test both before you pick.
Is it actually cheaper?
Per token, yes. Going from $10 / $50 to $8 / $40 is a 20% cut, the figure VentureBeat put in its headline. Cache reads cost $0.20 per million tokens and the Batch API takes 50% off, per the model docs. Outputs are up to 30% faster than Fable 5.1, and the optional fast mode offers up to 2.5x faster token generation at $16 / $80, exactly double the standard price, according to the launch post.
Per task is a different question. Artificial Analysis measured what one of its tasks costs at each effort level:
- Low: $3.10
- Medium: $3.80
- High: $9.08
- Xhigh: $14.83
- Max: $26.54
Time follows the same curve. At max, an AA task took 45.2 min, at high 17.3 min, at medium 7.8 min, with time to first answer token at max of 233.95 seconds, per Artificial Analysis. The index scores on the same page are 65 at max, 62 at both xhigh and high, 56 at medium and 54 at low. So xhigh costs $14.83 per task for the same index score that high reaches at $9.08. Max buys three more points for nearly three times the money and well over twice the wait.
The API default effort is high, per the model docs. Claude apps and Claude Code default to medium, per the launch post.
Anthropic’s own effort claims point the same way. On FrontierCode 2.0, Fable 5.2 at high scored 45.3% against Opus 5.5’s best of 45.5% “for roughly half the cost per task”. On WANDR, medium matches Opus 5.5’s top score for about 40% of the cost per task. On HLE, low does it for less than one tenth. On Chartography, medium scores 66.4% against 64.4% for Opus 5.5 at max. All vendor-run, all from the launch post.
Hence the advice. Anthropic calls Opus 5.5 “our strongest daily driver” and recommends starting there, then moving up to Fable or Mythos, or down to Sonnet 5.5 and Haiku 5.5. The family table in the docs puts Opus 5.5 at $4 / $20, Sonnet 5.5 at $2 / $10 and Haiku 5.5 from $0.10 / $0.50. I covered the daily driver Anthropic says to start with and the bottom of the price ladder earlier.
The post also names two routing patterns. As an Advisor, Fable 5.2 is what smaller models consult. As an Orchestrator, it delegates simpler tasks to them. If you run subagent setups like Claude Managed Agents, it is the same idea with a pricier brain on top.
Practical takeaway: pick effort per task, not per model. A Fable call at low can be cheaper than an Opus call at max, and on AA’s tasks a call at max costs about seven times one at medium.
What did 1,000 agents do with it?
The launch post is full of showcase results. Read every one as Anthropic’s claim, mostly from prerelease versions, and cherry-picked by definition. Here is what the launch post reports.
Math first. Anthropic says 1,000 Fable 5.2 agents formalized Fermat’s Last Theorem in Lean in 30 hours with minimal human intervention. That used one third of the action tokens of Anthropic’s earlier 11 day formalization, and the resulting proof is seven times shorter. Opus 5.5, given the same 1,000 agent setup, completed only 26% in 18 hours.
Then software. From an empty folder, Fable 5.2 coordinated over 1,000 subagents for eight days to build a prototype OS that boots on a laptop: core, disk, keyboard and network drivers, file storage, windows, login screen, installer. A team of agents on an earlier Fable 5.2 version migrated a production service from Python to Rust, and Anthropic says it now starts about 26x faster. The company also says over 90% of the changes merged into its core codebase are written by Claude, with human guidance.
Hardware too. From a one-paragraph prompt, working with Diode, the model produced a home voice-assistant hub, a USB speakerphone and a handheld virtual pet. Anthropic says each four-layer board was ready for manufacturing in under a day, and all worked without electrical modifications.
And science. Anthropic points to two open problems in black-hole physics tackled via holographic duality, written up as preprints. A Mythos 5.2 method for predicting whole AAV capsid structures accurately modeled 28 of 34 attempted, per Anthropic.
Nothing in the sources I read shows any of this reproduced outside Anthropic. The Fermat comparison is the most useful line, because it puts two Anthropic models in one harness and prints the losing number.
What did Anthropic deliberately hold back?
This is the least expected part of the launch post. During training, Anthropic says it “took steps to limit advances” in the capabilities that most directly speed up AI development. In kernel optimization and LLM architecture work, Fable 5.2 and Mythos 5.2 perform at about Opus 5.5 level, according to the launch post. For a small set of topics, such as kernel development on some nonstandard accelerators, queries fall back transparently to Opus 5.
So the most capable model in the lineup is, on purpose, not the best one at making the next model. Pay Fable prices for kernel work and you get roughly Opus 5.5 skill.
The other limits face outward. Anthropic says these are the strongest anti-distillation protections of any of its models, cutting the data distillers can harvest by 90% to 99%. The “preserved thinking” API feature now applies to all accounts, and developers may need to update integrations. Check that before you swap model IDs.
Enterprise Frontier Safeguards runs usage-window safety monitoring on prompts and responses stored in the customer’s own infrastructure, is ZDR-compatible, and is in private beta with Bridgewater, Salesforce and Stripe. Everyone else gets 30-day retention for safety monitoring, per the launch post. The system card has the details.
What do the safety numbers say?
External prerelease testing came from the UK AI Security Institute, Frontier Design and METR, per the launch post. In an automated behavioral audit of about 1,900 scenarios, Mythos 5.2 did as well as or better than Opus 5.5, Sonnet 5.5 and Mythos 5.1 on most measures.
The sandbox-escape eval is the one to read slowly. Anthropic built over 1,800 adversarial scenarios with a deliberate weakness in each. Mythos 5.2 used credentials or files outside its sandbox in 3.1% of them. Opus 5.5 did it in 3.3%, Sonnet 5.5 in 5.1%. Mythos 5.1 did it in 21% and Mythos 5 in 39%. That is a steep drop across three versions, on Anthropic’s own test.
Two weaker results are printed next to it. Mythos 5.2 is slightly worse than Opus 5.5 at avoiding harmful actions after wrongly concluding it was in a simulation. And both Fable 5.2 and Mythos 5.2 voice suspicion that they are being evaluated more often than any recent Claude model, which Anthropic says makes assessment harder.
That admission is the most useful sentence in the safety section. A model that suspects a test can behave better during it, which tells you how much weight to put on every number above. Thinking Machines did something similar in its Inkling 2 Small launch, printing an incident from its own training. More of this, please.
What should builders watch next?
This week’s launch promo gives eligible paid plans one to three Fable resets and includes Fable for up to 50% of weekly usage limits on Pro and Team standard seats, per the launch post, so early usage impressions will be inflated. Watch whether third-party reruns hold up against Anthropic’s table, especially the internal RoboDojo run. Watch how fast Mythos access widens beyond the Verification Programs, which are also the only place Mythos fast mode is offered. And watch how Gemini 4 Carbon and GPT-6 Astra respond on price. The model ID is claude-fable-5-2 on the Claude Platform, AWS, Google Cloud and Microsoft Azure/Foundry, per the docs. Start where Anthropic says to start, and route up when a task earns it.
Sources
- Anthropic, launch post for Claude Fable 5.2 and Claude Mythos 5.2, 9 November 2026
- Claude Platform, Fable 5.2 model docs
- Anthropic, Claude Fable product page
- Anthropic, Fable 5.2 and Mythos 5.2 system card
- Artificial Analysis, Anthropic provider page and model release comparisons
- Google, Gemini 4 Carbon availability post
- VentureBeat, coverage of the Fable 5.2 and Mythos 5.2 release
