Claude Fable 5.2 and Mythos 5.2 Explained: Anthropic’s New Coding Models
Key Takeaways
- Anthropic released Claude Fable 5.2 and Claude Mythos 5.2 on 9 November 2026, describing them as built for the hardest technical challenges, with Fable 5.2 positioned as a new bar in coding and finance.
- The two names describe the same underlying model with two different safeguard levels: Fable 5.2 is available to everyone with strong cybersecurity and biology safeguards, while Mythos 5.2 removes some of those filters and is limited to organisations that pass Anthropic’s verification programmes.
- Pricing fell 20 percent from the 5.1 generation: 8 dollars per million input tokens and 40 dollars per million output tokens, against 10 and 50 dollars before, with output up to 30 percent faster.
- On Anthropic’s own reported benchmarks, Fable 5.2 scored 49.1 percent on FrontierCode 2.0 at its highest effort level, ahead of Claude Opus 5.5 at 45.5 percent, and 69.9 percent on FrontierSWE v2 against 62.3 percent for Opus 5.5.
- The gains are widest on long, hard tasks: 74.6 percent on Terminal-Bench-Science against 58.7 percent for Opus 5.5, and 64.9 percent on the RoboDojo robotics simulation against 39.4 percent.
- Anthropic deliberately capped the model’s skill at the research tasks that could most directly speed up AI development itself, so on accelerator kernel optimisation and language-model architecture work it performs at about the level of Opus 5.5 by design.
- The benchmark figures in this article are Anthropic’s own reported numbers and have not been independently reproduced; where an independent check exists, it is reported separately.
What Launched on 9 November
Anthropic released Claude Fable 5.2 and Claude Mythos 5.2 on Sunday 9 November 2026. The launch post presented them as the company’s frontier models for the hardest technical challenges, and the timing was conspicuous: the release landed in the same week Anthropic opened the roadshow for its initial public offering, which we covered in our report on Anthropic’s IPO being oversubscribed. Product news and financing news reached the same audience in the same week, and the models were the product half of that story.
The 5.2 models follow Fable 5.1 and Mythos 5.1, which were released in September 2026. The version number is worth a pause, because it confuses almost everyone who reads it. The everyday Claude line, Opus, Sonnet and Haiku, had just shipped as version 5.5, so a flagship numbered 5.2 looks older than the models it is meant to beat. Anthropic has never published a numbering policy that explains this, and the honest answer is that the number does not encode a ranking. Fable and Mythos are a separate line from Opus, Sonnet and Haiku, with their own numbering, and the 5.2 badge sits on what Anthropic’s own benchmark table presents as its strongest model.
The basics are unchanged from the 5.1 generation. Both models keep a one-million-token context window and a maximum output of 128,000 tokens. Adaptive thinking is always on, with five effort levels from low to max, so the model decides how much reasoning a task deserves and the user can push it further when a task warrants it. On the API the model is called claude-fable-5-2, and it was available from launch day on the Claude API, Amazon Bedrock, Google Cloud and Microsoft Foundry, as well as in the Claude apps on paid plans. According to Anthropic’s model documentation, as reported by independent coverage, both models have a retirement date not sooner than 9 November 2027, which means Anthropic commits to keeping them available for at least a year.
One more detail from the launch matters for anyone planning around these models. Anthropic describes Fable 5.2 as the model for the hardest technical challenges, and the launch materials lean heavily on coding, finance, scientific research and robotics. Those four areas are where the benchmark gaps over the previous generation are largest, and they are also where the pricing and speed changes matter most, because long agentic tasks are where token bills accumulate.
One Model, Two Names
The most important thing to understand about this launch is that Fable 5.2 and Mythos 5.2 are the same model. Anthropic’s own description is that they are the same underlying model, and the two names describe two safeguard profiles applied to identical weights. Fable is the version with strong cybersecurity and biology safeguards, and it is available to every API customer. Mythos is the same model without some of those safeguards, and it is reserved for organisations that pass Anthropic’s verification programmes, the Cyber Verification Program and the Life Sciences Verification Program.
The split is a regulatory and safety decision, not a technical one. Fable’s safeguards exist so that the model can be offered broadly without handing over capabilities that could be misused for cyberattacks or biological harm. Mythos exists because some legitimate work, defensive cybersecurity research and life-sciences research among it, genuinely needs the unfiltered capability, and the organisations doing that work can be vetted. The price is the same for both, and the benchmark scores on general tasks are the same; the difference is only where the safeguards intervene.
That difference shows up in the numbers. Anthropic reports that Fable 5.2 triggers about 55 percent fewer safety blocks than Fable 5.1 in total, with roughly 60 percent fewer blocks in cybersecurity and 40 percent fewer in biology. In other words, the safeguarded model got less trigger-happy, which matters for legitimate users who were previously hitting refusals on borderline requests. Mythos 5.2 is reported separately on the security and biology tasks where its unfiltered capability is the point: 76.3 percent on CyScenarioBench against 61.7 percent for Mythos 5.1, 89.6 percent on CVE Rediscovery against 78.3 percent, and 85.1 percent on a 24-hour de novo protein binder design task. Those are the tasks where the safeguard split is supposed to matter, and the gaps are large.
There is a practical consequence for buyers. Most teams will only ever meet Fable 5.2, and for them the Mythos numbers are interesting but irrelevant. Organisations doing defensive security work or life-sciences research should check whether they qualify for the verification programmes, because Mythos is the version built for them. And anyone writing about these models should be careful not to quote Mythos scores as if they describe the model most people can use; the general-purpose scores are the Fable scores, and the Mythos scores belong to a restricted variant.
Pricing: 20 Percent Cheaper Per Token
The price list is straightforward. Fable 5.2 and Mythos 5.2 cost 8 dollars per million input tokens and 40 dollars per million output tokens on the Claude API, down from 10 and 50 dollars for the 5.1 generation. That is a 20 percent cut. Batch processing halves both figures, to 4 and 20 dollars. There is also an optional fast mode that runs up to two and a half times faster at double the price, 16 dollars in and 80 dollars out.
The price cut is real, but it does not automatically mean a lower bill, and Anthropic’s own system-card data makes that point clearly. On FrontierCode 2.0, an agentic coding benchmark, Fable 5.2 at low effort solves 34.2 percent of tasks at an average cost of 2.10 dollars per task. At medium effort it solves 37.6 percent at 2.48 dollars. At high effort, 45.3 percent at 6.72 dollars. At xhigh effort, 49.1 percent at 14.19 dollars. At max effort, 45.7 percent at 44.70 dollars. Read that last line twice: the most expensive setting solves fewer tasks than the second-most expensive one, at more than three times the cost per task.
The lesson is that the effort level matters more than the price list. A team that runs everything at max effort will pay far more than a team that matches effort to task difficulty, and will not necessarily get better results. Anthropic’s own guidance points the same way: the company still recommends starting with Opus 5.5 and moving up to Fable 5.2 only when a task needs it. The per-token price is still twice that of Opus 5.5, so the cheaper list price does not make Fable 5.2 the cheap option. It makes it the expensive option that got somewhat less expensive.
Independent cost checks reinforce the caution. On Terminal-Bench 4.0, Vals AI measured 14.56 dollars per test with Fable 5.2 against 12.04 dollars with Opus 5.5. On CursorBench 4.0, at maximum effort, Cursor reported 32.79 dollars against 13.43 dollars for the comparison model. Those are third-party measurements of cost per task, not Anthropic’s figures, and they show the same pattern: Fable 5.2 can cost substantially more per completed task than the model below it, especially at high effort. The right question for a team is not what the model costs per token, but what it costs per task completed, at the effort level they actually use.
Speed: Faster, With a Fast Mode
Output is up to 30 percent faster than the 5.1 generation, according to Anthropic, and the fast mode pushes that to up to two and a half times faster at double the price. Speed matters here for a different reason than it usually does. These models are built for long agentic tasks that can run for many minutes or hours, and a 30 percent speed improvement on a task that runs for an hour is eighteen minutes saved. For interactive use, where a developer is waiting on a response, the difference is felt immediately.
The fast mode is a trade-off worth understanding. It doubles the price for up to two and a half times the speed, which means the cost per token doubles while the time per token falls. For a team running overnight batch jobs, speed is worth little and the fast mode is a bad deal. For a team running interactive sessions where a developer is blocked waiting for output, the fast mode may be worth every cent. The choice depends on whether the scarce resource is money or human waiting time, and different teams will answer that differently.
The Benchmarks, In Detail
Every benchmark figure in this section comes from Anthropic’s own launch materials and system card, as reported by independent coverage, and none of it has been independently reproduced. That is the standard caveat for launch-day benchmarks, and it applies here with full force. Anthropic tested its own model, its rivals’ models, and published the table. The figures are useful as a first signal, not as a final verdict.
Start with coding, because that is where Anthropic makes its strongest claim. On FrontierCode 2.0, an agentic coding benchmark, Fable 5.2 scored 49.1 percent at xhigh effort, against 45.5 percent for Opus 5.5 and 32.4 percent for Fable 5.1. That is the headline number, and it is the largest single jump in the table: a 16.7-point improvement over the previous generation. On FrontierSWE v2, a software-engineering benchmark, Fable 5.2 scored 69.9 percent against 62.3 percent for Opus 5.5 and 56.3 percent for Fable 5.1. On OSWorld 2.1, which measures computer use, Fable 5.2 scored 89.3 percent against 87.2 percent for Opus 5.5 and 83.5 percent for Fable 5.1.
The pattern in the coding numbers is consistent: a clear lead over the previous generation, and a smaller but real lead over Opus 5.5. The lead over Opus is a few points on the coding and computer-use benchmarks, which is meaningful but not dramatic. The lead over Fable 5.1 is much larger, which is what you would expect from a generational step.
The gaps widen sharply on scientific research and robotics. On Terminal-Bench-Science, an agentic scientific-research benchmark, Fable 5.2 scored 74.6 percent against 58.7 percent for Opus 5.5 and 52.6 percent for Fable 5.1. That is a 22-point jump over the previous generation and a nearly 16-point lead over Opus. On Humanity’s Last Exam with tools, Fable 5.2 scored 74.2 percent against 67.7 percent for Opus 5.5 and 65.6 percent for Fable 5.1. On the RoboDojo robotics simulation, Fable 5.2 scored 64.9 percent against 39.4 percent for Opus 5.5 and 27.6 percent for Fable 5.1, more than doubling the previous generation’s score.
There is one benchmark where Fable 5.2 does not lead, and it is worth naming because it corrects the impression that the model wins everywhere. On AutomationBench, which measures business workflows, Fable 5.2 scored 51.4 percent against 42.5 percent for Opus 5.5 and 31.4 percent for Fable 5.1, a large improvement, but Google’s Gemini 4 Carbon scored 53.9 percent and leads the table. On the AA-Briefcase knowledge-work Elo rating, Opus 5.5 at 1800 edges out Fable 5.2 at 1785. The model is the strongest Anthropic has shipped, but it is not the best at everything, and the two benchmarks where it trails are both knowledge-work benchmarks rather than coding ones.
The comparison with rival labs is also in Anthropic’s table, and it should be read with the same vendor-reported caveat. On FrontierCode 2.0, Anthropic lists GPT-6 Astra at 42.8 percent and Gemini 4 Carbon at 41.1 percent, both behind Fable 5.2 at 49.1 percent. On FrontierSWE v2, GPT-6 Astra is listed at 65.5 percent and Gemini 4 Carbon at 61.2 percent, behind Fable 5.2 at 69.9 percent. On Terminal-Bench-Science, GPT-6 Astra is listed at 68.1 percent and Gemini 4 Carbon at 64.8 percent, behind Fable 5.2 at 74.6 percent. Those are Anthropic’s numbers for its competitors, run in Anthropic’s harness, and the rival labs may well publish different numbers. Treat them as Anthropic’s claim, not as a settled ranking.
The Finance Claim
The research behind this article describes Fable 5.2 as billed as a new bar in coding and finance, and the launch materials do present the model as built for the hardest technical challenges, with finance among them. The specific finance evidence is thinner than the coding evidence. Anthropic’s published benchmark table, as covered by independent outlets, does not include a standalone finance benchmark with a headline score in the way it includes FrontierCode for coding or Terminal-Bench-Science for research. The AutomationBench result, which covers business workflows, is the closest published proxy, and on that benchmark Fable 5.2 trails Gemini 4 Carbon.
That does not mean the finance claim is false. It means the public evidence for it is weaker than the evidence for coding, and readers should not treat the two claims as equally supported. A model that is strong at long-horizon reasoning, tool use and multi-step verification is plausibly strong at financial analysis, and Anthropic clearly believes its finance performance is a selling point. But until Anthropic or an independent tester publishes finance-specific results, the finance claim rests on the general reasoning results rather than on a finance benchmark. We flag it here rather than repeating it as established fact.
The Deliberate Cap
The most unusual detail in this launch is the one Anthropic volunteered. The company says it deliberately limited the model’s skill at the research tasks that could most directly speed up AI development itself. On accelerator kernel optimisation and language-model architecture work, Fable 5.2 performs at about the level of Opus 5.5 by design, not by accident. The cap is a safety choice, and it connects to a position Anthropic’s chief executive Dario Amodei took in September 2026, when he called for a slower pace of AI development.
This is a rare thing in the AI industry: a lab shipping its strongest model with a disclosed, deliberate limitation on the capability that would most accelerate its own field. It is also a choice with practical consequences. Teams doing the specific kind of research that the cap covers will not see the gains that coding or science teams see, and they should know that before they pay Fable 5.2 prices for Opus 5.5-level results on those tasks. For everyone else, the cap is irrelevant, and the model performs as the benchmark table describes.
Whether the cap is the right call is a debate for the safety community, not for this article. What matters for buyers is that it exists, that it is disclosed, and that it is specific. It covers accelerator kernel optimisation and language-model architecture work. It does not cover coding, science, robotics, finance or general knowledge work, where the model performs at the level the benchmarks show.
The Thousand-Agent Proof
One launch detail has travelled further than any benchmark: Anthropic says about 1,000 Fable 5.2 agents produced a formalisation of Fermat’s Last Theorem in the Lean proof system in about 30 hours. It is a striking demonstration, and it is worth understanding what it does and does not prove.
What it proves is that the model can sustain a very large, long, coordinated agentic task, with many sub-agents working in parallel on a problem that requires precision, persistence and verification. Formalising a mathematical proof is unforgiving work: a single error invalidates the result, and the task requires thousands of small steps to be correct and consistent. Completing it in 30 hours with a thousand agents is a genuine capability demonstration.
What it does not prove is that a thousand agents are a practical tool for most teams, or that the same approach works as well on messier, less formal problems. A mathematical proof has a clear success criterion and a verification mechanism built in. Most real-world work does not. The demonstration is best read as evidence about the model’s ability to handle long-horizon, multi-agent coordination, not as a promise that throwing a thousand agents at any problem will solve it.
Who Should Upgrade, And Who Should Not
Anthropic’s own guidance is the right starting point: start with Opus 5.5 and move up only when a task needs it. That advice is worth taking seriously, because it comes from the company selling the more expensive model, and because the cost data supports it.
Upgrade to Fable 5.2 when the work is long, hard and agentic. Multi-hour coding tasks, complex scientific research, robotics simulation, and work where the model needs to sustain reasoning across many steps are the cases where the benchmark gaps are largest and the extra cost is most likely to pay for itself. The FrontierCode cost data shows that high effort on a hard task costs about 6.72 dollars per task for a 45.3 percent solve rate, against 2.48 dollars at medium effort for 37.6 percent. That is a real trade-off, and on a hard task the extra spend may be worth it.
Stay on Opus 5.5 when the work is short, routine or cost-sensitive. Opus 5.5 is cheaper per token, it edges out Fable 5.2 on the AA-Briefcase knowledge-work Elo, and for many everyday tasks the difference in output quality will be small or invisible. Teams running high-volume, low-complexity workloads should be especially cautious: the per-token price is twice Opus, and at high effort the cost per task can be several times higher.
Match effort to difficulty, whatever you choose. The FrontierCode data shows that max effort can cost more and solve less than xhigh effort, and the general principle is that more reasoning is not always better reasoning. Start low, measure the solve rate on your own tasks, and increase effort only where the results improve. The effort level is a dial, not a badge of seriousness.
The IPO-Week Context
The launch cannot be separated from the week it happened in. Anthropic opened its IPO roadshow on 10 November, the day after the models shipped, and by 12 November Bloomberg was reporting that the offering had drawn orders for about six times the shares on offer. The models and the IPO are separate stories, but they reached the same audience in the same week, and the product launch was, intentionally or not, part of the investment case. Investors were being asked to buy into a company that was simultaneously demonstrating that its flagship model had taken a large step forward.
That context cuts both ways. A strong product launch supports the investment case, and the benchmark table is genuinely impressive. But it also means the launch materials were produced in a week when Anthropic had every incentive to present its best numbers in the best light. That is not an accusation of dishonesty; it is a reason to treat the vendor-reported benchmarks as a first signal and to wait for independent evaluations before drawing final conclusions. The history of model launches is full of benchmark leads that narrowed or reversed under independent testing.
For more on the financing side, our earlier coverage explains the IPO terms and what the reported demand does and does not prove. Our analysis of OpenAI and Anthropic’s trillion-dollar compute bill explains why the money Anthropic raises matters for its ability to keep training models at this scale, and our report on Claude Code secret leaks covers a security question that coding teams using Claude should understand alongside any upgrade decision.
What We Could Not Verify
- The claim that Fable 5.2 sets a new bar in finance is not supported by a standalone finance benchmark in the published materials we reviewed. The closest published proxy is AutomationBench, where Gemini 4 Carbon leads. We report the finance claim as Anthropic’s positioning, not as a verified benchmark result.
- All benchmark figures in this article are Anthropic’s own reported numbers. They have not been independently reproduced, and Anthropic notes that some scores, including cybersecurity scenario scores, use different tasks for each lab. Independent leaderboards were updating within hours of release, and early results diverge from the first-party picture on coding.
- A report that Epoch AI recorded a new record score of 174 for Fable 5.2 appeared in truncated search results and we could not verify the benchmark, the scale or the context. We do not rely on that figure.
- The claim, in the research behind this article, that Fable 5.2 took the top spot on an AI index is consistent with reporting that Fable 5.2 scored 62 on the Artificial Analysis index against 60 for Gemini 4 Carbon, but we have not verified the index directly and the ranking depends on the settings and effort levels used. Treat it as reported, not confirmed.
- The 1,000-agent Fermat’s Last Theorem formalisation is Anthropic’s account of its own demonstration. We have not independently verified the agent count, the time taken, or the completeness of the formalisation.
- The retirement date of not sooner than 9 November 2027 comes from Anthropic’s model documentation as reported by independent coverage, not from a document we inspected directly.
Sources
- Maverick AI, “Claude Fable 5.2: what’s new, pricing and benchmarks (Nov 2026)”: https://www.maverickconsulting.it/en/risorse/claude-fable-5-2
- i10x, “Claude Fable 5.2 and Mythos 5.2: Benchmarks, Price, Limits”: http://i10x.ai/news/anthropic-claude-fable-5-2-mythos-5-2-launch
- DataStudios, “Claude Fable 5.2 launches at 8 dollars and 40 dollars per 1M tokens, 20 percent below Fable 5.1”: https://www.datastudios.org/post/anthropic-claude-fable-5-2-mythos-5-2-launch-pricing-plans
- StackFutures, “Claude Fable 5.2 and Mythos 5.2 Launch at 8 dollars and 40 dollars per Million Tokens, With 74.2 percent on HLE and a Prototype OS Built by 1,000 Subagents”: https://stackfutures.com/blog/claude-fable-5-2-mythos-5-2-launch-price-cut-8-40-frontiercode-49-nov-2026/