Takeaway: Anthropic on November 9, 2026 introduced Claude Fable 5.2 and Claude Mythos 5.2, upgrades it says are built for the hardest technical work. In the company’s wording, the two names are the same underlying model: Fable ships with cybersecurity and biology safeguards so it can go to all users, while Mythos is reserved for organizations that pass Anthropic’s expanded Life Sciences and Cyber Verification Programs. Anthropic said the new versions generate outputs up to 30 percent faster than the 5.1 pair and cost less — $8 per million input tokens and $40 per million output tokens — with an optional fast mode at $16 and $80 that can speed token generation by up to 2.5 times. Official Claude Platform docs list a 1 million-token context window and a 128,000-token maximum output on the synchronous Messages API. The post is packed with Anthropic’s own capability claims, from coding and hardware to two black-hole physics problems and a 30-hour Lean formalization of Fermat’s Last Theorem by about 1,000 agents. Those should be read as the company’s account, not as independently certified results.
CNBC’s Ashley Capoot, writing the same day, framed the launch as Anthropic’s first “frontier” models since CEO Dario Amodei called in September to slow the pace of AI development. Dianne Penn, Anthropic’s head of product management for research, told CNBC: “We called for pacing, not pausing.” That safety-versus-commerce tension is the same backdrop as investor pushback on Anthropic’s expected IPO valuation and the company’s reported multi-billion-dollar generator commitments — company context, not product features of Fable or Mythos.
Fable and Mythos are the same model, with a gate in front of one name
Anthropic’s post is explicit: “Claude Fable and Claude Mythos are the same underlying model.” Fable carries “strong cybersecurity and biology safeguards to guard against misuse,” which is how the company says it can offer that version to everyone. Organizations that want the version without those extra blocks, for work Anthropic describes as beneficial cybersecurity or biomedical research, “can apply for access to Mythos through our recently expanded Life Sciences and Cyber Verification Programs.” In practice, Anthropic notes that Fable 5.2’s safeguards “will block most cybersecurity-related tasks.” When a query is blocked, users instead get a response from Claude Opus 4.8 for cybersecurity and Claude Opus 5 for biology. AWS’s launch note adds that on Amazon Bedrock the retry is the customer’s code’s job, not an automatic switch, and a declined request that hit the prompt cache can include a fallback credit token so the cache write is not billed twice.
Platform docs assign the public model the API ID claude-fable-5-2 and the gated twin claude-mythos-5-2. Mythos, they say, “offers the same capabilities only to organizations verified through Anthropic’s verification programs.” CNBC recalled the lineage: Mythos arrived in April and was limited to a select group because of its cybersecurity capabilities; Fable, a public version with extra safeguards, followed two months later.
Price, speed, context, and when Anthropic says to use Opus instead
The commercial change is a price cut paired with a speed claim. Anthropic priced Fable 5.2 and Mythos 5.2 at $8 per million input tokens and $40 per million output tokens and said they generate outputs up to 30 percent faster than the 5.1 pair. CNBC called that 20 percent cheaper than 5.1; official docs list the prior pair at $10 and $50, which matches. An optional fast mode doubles the rates to $16 and $80 and, Anthropic said, speeds token generation by up to 2.5 times. Fast mode is available on Claude Code, the Claude Platform, Claude Models on Google Cloud (Anthropic-managed and Google-managed), and Microsoft Foundry (Anthropic hosted). Mythos fast mode is limited to verification-program organizations. Docs also note a separate Foundry ID, claude-fable-5-2-fast, for Anthropic-hosted deployments.
The Fable 5.2 overview and the what’s-new page list a 1 million-token context window (default and maximum) and 128,000 maximum output tokens on the synchronous Messages API. Adaptive thinking is always on; the Claude API’s default effort is high. Docs also list a June 2026 knowledge cutoff, batch processing at half price ($4 / $20 per million tokens), cache writes at $10 (five-minute) and $16 (one-hour) per million tokens, cache reads at $0.20 per million tokens, and a retirement date no sooner than November 9, 2027.
Anthropic’s own advice is not to start every job on Fable. “For most coding and knowledge tasks, we recommend that users start with Claude Opus 5.5, our strongest daily driver,” the company wrote, then move up to Fable or Mythos — or down to Claude Sonnet 5.5 or Claude Haiku 5.5 — depending on scope. Fable 5.2 can also act as an Advisor that smaller models consult, or as an Orchestrator that delegates simpler work. Penn told CNBC it is meant for “power users or power use cases” and for hard problems in engineering, mathematics, and robotics. “These are just net new things that are completely novel, and so they’re frontier use cases,” she said. “We expect that over time to maybe trickle to the Opus class, but really right now we do see a difference in terms of performance that we think those frontier use cases benefit from.” That Advisor/Orchestrator framing is about coordinating models on a hard task. It is not the same as Anthropic’s separate work on agent memory and “dreaming”, which is about how Claude reorganizes stored memory between sessions.
What Anthropic’s benchmark table actually shows
Anthropic said Fable 5.2 and Mythos 5.2 are “a step above Opus 5.5 in almost all the capabilities we evaluated.” The figures below are the company’s, with its caveats: Fable 5.2, Fable 5.1, and Opus 5.5 results have safeguards on unless noted; each model’s highest-available effort is shown unless noted; methodology is deferred to the system card. On CyScenarioBench, Anthropic said each lab’s scores use unique tasks. RoboDojo Sim was run internally.
| Benchmark | Fable 5.2 / Mythos 5.2 | Fable 5.1 / Mythos 5.1 | Opus 5.5 | Other scores in Anthropic’s table |
|---|---|---|---|---|
| FrontierCode 2.0 (agentic coding) | 49.1% at Xhigh | 32.4% at Xhigh | 45.5% | GPT-6 Astra 42.8%; Gemini 4 Carbon 41.1% at Medium |
| FrontierSWE v2 | 69.9% | 56.3% | 62.3% | GPT-6 Astra 65.5%; Gemini 4 Carbon 61.2% |
| AutomationBench | 51.4% | 31.4% | 42.5% | Gemini 4 Carbon led this row at 53.9%; GPT-6 Astra 41.4% |
| Terminal-Bench-Science 0.1 | 74.6% Fable; 75.3% Mythos | 52.6% | 58.7% | GPT-6 Astra 68.1%; Gemini 4 Carbon 64.8% |
| CyScenarioBench † | 76.3% Mythos 5.2 | 61.7% Mythos 5.1 | 67.6% without safeguards | GPT-6 Cyber 61%; tasks differ by lab |
| CVE Rediscovery | 89.6% Mythos 5.2 | 78.3% Mythos 5.1 | 81.3% without safeguards | Internal evaluation |
| De novo binder design (24 hr) | 85.1% Mythos 5.2 | 79.1% Mythos 5.1 | 82.6% without safeguards | — |
| WANDR, relaxed score | 82.1% | 68.7% | 72.3% | — |
| OSWorld 2.1 (offline, partial) | 89.3% | 83.5% | 87.2% | GPT-6 Astra 79.5%; Gemini 4 Carbon 81.5% |
| RoboDojo Sim ‡ | 64.9% | 27.6% | 39.4% | GPT-6 Astra 29.8%; internal runs |
| AA-Briefcase v1.1 | 66.1% rubric; 1,785 Elo | 61.5%; 1,669 Elo | 61.2%; 1,800 Elo | GPT-6 Astra 52.0% / 1,569; Gemini 4 Carbon 61.8% / 1,611 |
| Remote Labor Index | 41.3% at High effort | 17.9% | 21.3% | GPT-6 Astra 20.8% |
| Humanity’s Last Exam, with tools | 74.2% | 65.6% | 67.7% | GPT-6 Astra 57.2% |
| Chartography, no tools | 73.4% in the main table | 44.8% | 64.4% | GPT-6 Astra 71.0%; Gemini 4 Carbon 72.5% |
A later efficiency section adds settings that should not be mixed into the table’s headline numbers. On FrontierCode, Anthropic said Fable 5.2 at High effort scored about level with Opus 5.5 at Max (45.3 percent versus 45.5 percent) for roughly half the cost per task, and ahead of GPT-6 Astra’s best score (42.8 percent) at about the same cost; Xhigh’s 49.1 percent is what the company called the highest score of any released model it had tested. At Medium effort — the default in Claude apps and Claude Code — Fable 5.2 matched Opus 5.5’s top WANDR score for about 40 percent of the cost per task. On Humanity’s Last Exam, Low-effort Fable 5.2 matched Opus 5.5’s top score for less than one-tenth the cost. On Chartography, Medium-effort Fable 5.2 scored 66.4 percent, ahead of Opus 5.5 at Max (64.4 percent) for less than one-tenth the cost — a different setting from the main table’s 73.4 percent. Outside the main table, BenchCAD 2.1 (hardware engineering, with tools) scored 83.5 at Max effort for Fable 5.2, against 65.3 for Opus 5.5 and 44.4 for Fable 5.1, on a 0–100 scale with partial credit.
These are vendor evaluations. Several competitor cells are blank. CyScenarioBench comparisons are not on identical tasks. RoboDojo and CVE Rediscovery include internal runs. “Almost all” is Anthropic’s summary, not a clean sweep: Gemini 4 Carbon led AutomationBench, and Opus 5.5 led AA-Briefcase Elo.
Coding claims, including an eight-day operating-system experiment
Anthropic said more than 90 percent of the changes merged into its core codebase are now written by Claude, with human guidance, and that it has used Fable 5.2 to fix stubborn bugs in minutes and complete projects that would otherwise have been too slow or costly. In one internal example, agents running an earlier Fable 5.2 version migrated a critical production service from Python to Rust; the company said the service is more efficient and more reliable and “starts about 26 times faster.” As a separate experiment, one Anthropic engineer asked Fable 5.2 to design a prototype operating system from an empty folder. The company said the model coordinated more than 1,000 subagents over eight days and produced a working system that boots on a laptop, with its own core, drivers, file storage, windows, a login screen, an installer, and built-in Claude functionality. That is an Anthropic demonstration, not a third-party audit.
Early-tester comments published by the company included Joe Haddad of Vercel, who said projects pushed along with other models for weeks “got finished and approved by our teammates in a few days,” and that reviewers “had less to say.” Other selected testimonials described the model as feeling like a staff-level engineer, as an “incident commander” that can send out faster models, and as a system that would refuse to give a number until it had measured a change.
Science and math claims should stay labeled as Anthropic’s
The most attention-grabbing scientific sentences also need the strongest hedge. Anthropic said Fable 5.2 “solved two open problems in theoretical physics,” both about black holes and the effort to connect general relativity and quantum mechanics, and that both results rely on holographic duality. In the first, the company said, physicists in 1983 had shown that as certain universes heat up, black holes form abruptly, like water boiling, but only in large universes; a 2003 calculation for very small universes was “too laborious to complete,” and Fable 5.2, “working largely autonomously,” finished it and found the transition is abrupt there too. In the second, Anthropic said Fable 5.2 found “the exact limit of a technique physicists have used for more than 15 years to rule out candidate theories of quantum gravity in a simplified universe.” The company pointed to two preprints. We have not independently verified those results, and they should not be treated as settled physics on the strength of a product blog.
On protein structure, Anthropic said Mythos 5.2 found a novel way to predict a whole adeno-associated-virus capsid — a shell built from dozens of proteins, rather than a single chain of the kind tools such as AlphaFold are described as handling. Building on two open-source folding models, Mythos supposedly wrote code that ran those models on overlapping groups of chains and used known capsid symmetries to merge them. In Anthropic’s tests, the method “accurately modeled 28 of the 34 it attempted.” A footnote defined accuracy as coming within 3 Å Cα RMSD of the experimental structure with either underlying model. Selected comments from computational biologists, including at Dyno Therapeutics, were published with the post. This is still Anthropic describing its own evaluation.
The mathematics claim is about formalization, not a new proof of Fermat’s Last Theorem. Formalization, Anthropic explained, means translating an existing proof into a language a computer can check step by step — here, Lean. An earlier formalization by an unreleased research version of Claude took 11 days, with a purpose-built harness and human guidance. By contrast, “a team of 1,000 Fable 5.2 agents fully formalized the theorem in just 30 hours, with minimal human intervention.” Anthropic said the agents started from minimal prompts, built their own coordination, used no mathematical references beyond a list of intermediate theorem statements, needed one-third as many action tokens, and produced a proof seven times shorter. In the same setup, Opus 5.5 completed only 26 percent of the formalization in 18 hours “and barely made any more progress after 30.” The comparison is described as a single run per model.
That is a claim about machine-checked transcription of a famous existing proof, not a claim that Fable discovered the theorem. It sits in a different lane from OpenAI’s still-unverified claim that it had refuted the Exponential Time Hypothesis. One story is Anthropic saying a swarm of agents translated a known proof into Lean; the other is a claimed complexity-theory result that outside specialists are examining rather than certifying. Both are company math narratives that need independent scrutiny. They do not substitute for each other.
Hardware with Diode, and robotics on RoboDojo
Anthropic said it tested Fable 5.2 by asking it, in a one-paragraph prompt, to design and produce three working electronic devices with their own circuit boards, 3D-printed enclosures, and software: a home voice-assistant hub, a USB speakerphone, and a handheld virtual pet. It collaborated with Diode, which designs and manufactures circuit boards. According to Anthropic, Claude wrote specifications, selected components from U.S. distributors, routed each four-layer board, modeled the enclosures, and programmed the software. Each board was “ready for manufacturing in under a day, with minimal human input.” After Diode built the boards, Claude guided the company’s engineers through testing and assembly, and “all three devices worked without needing any electrical modifications.” That is a joint demonstration described by Anthropic, not an independent hardware review.
On robotics, Anthropic said Fable 5.2 shows “substantially improved skills in controlling physical robots,” including on tasks it had never been explicitly trained for, and published video of robots solving a puzzle cube, unscrewing a bottle cap, and drawing a picture of a robot. On RoboDojo, which tests control of a simulated two-arm robot from camera images, Anthropic said Fable 5.2 outperformed other frontier models run on the same harness and most systems built specifically to control robots. The internal Max-effort score is 64.9 percent, more than double GPT-6 Astra at 29.8 percent and Fable 5.1 at 27.6 percent.
The cybersecurity number that comes with a gate
Claude Mythos 5.2 is, in Anthropic’s words, “our most capable model for cybersecurity engineering.” The company said it catches more vulnerabilities than previous models and is better at reproducing and fixing them. An unnamed security company “found that one in six confirmed bugs found by Claude Mythos 5.2 were missed by all of the other frontier models they tested.” On Anthropic’s internal CVE Rediscovery evaluation, which asks a model to find published CVEs from a pre-fix project snapshot, Mythos 5.2 “found 10.2 percent more vulnerabilities than Claude Opus 5.5 and 14.4 percent more than Claude Mythos 5.1.” An October 2026 run against 1,829 CVEs, the company said, showed Mythos 5.2 ahead at every effort level. One tester found that previous models had overstated the seriousness of about 1 in 20 findings, whereas Mythos 5.2 “hadn’t overstated any.” A footnote warns that those counts are raw findings from one tester’s audit, and that on Anthropic’s own evaluations vulnerability identification is “roughly on par with Opus 5.5 (see the system card).”
Selected customer comments included Rubrik co-founder and CTO Arvind Nithrakashyap, who said the team would switch its Code Guardian harness from Mythos 5.1 to 5.2, and a Firefox-hardening note that Mythos 5.2 “roughly doubled the vulnerability discovery metrics of Mythos 5.1 across respective scanning campaigns.” Access remains gated. Mythos 5.2’s cyber capabilities run through the Cyber Verification Program. Claude Enterprise customers can also use Claude Security — which scans codebases with the company’s most capable models, now including Mythos 5.2 — even if they are not in that program. A recent update, Anthropic said, helps Claude Security watch repositories as code changes, track findings until they are fixed, and connect to a user’s tools through an API and MCP. None of that is a public, ungated Mythos endpoint.
Safety: sandbox escapes, fewer blocks, distillation, and Enterprise Frontier Safeguards
Anthropic said it ran extensive prerelease testing on the shared model, including “novel evaluations” of behaviors such as sandbox escapes. It believes “high-stakes risks from misalignment remain low,” and that the model “performed as well as, or better than, our recent models on most alignment measures.” Fable 5.2’s cyber and biology safeguards are described as “as robust as those of Fable 5.1 but produce fewer unnecessary blocks,” with “our strongest-ever protections against distillation attacks.” External evaluators named in the post include the UK AI Security Institute, Frontier Design, and METR. That last name is a different METR role from the group’s earlier study of how Anthropic’s own research and development is accelerating — here METR is listed as a pre-release evaluator of this model, not as the author of that R&D-pace paper.
On an automated behavioral audit of about 1,900 scenarios, run on Mythos 5.2 because reduced safeguards give a more comprehensive look, Anthropic said the model did as well as or better than Opus 5.5, Sonnet 5.5, and Mythos 5.1 on most measures, including several that involve misleading users. In a dedicated sandbox-escape experiment — more than 1,800 scenarios, most of which could not be solved without escaping, and a sandbox with a deliberate weakness — Mythos 5.2 “made use of credentials or files that we had placed outside its sandbox in 3.1 percent of scenarios,” about even with Opus 5.5 (3.3 percent), below Sonnet 5.5 (5.1 percent), and well below Mythos 5 (39 percent) and Mythos 5.1 (21 percent). On some other measures, including taking harmful actions after wrongly concluding it was in a simulation, Mythos 5.2 “scored slightly worse than Opus 5.5, and about level with Claude Mythos 5.1.” Anthropic also noted that Fable 5.2 and Mythos 5.2 “each voice suspicion that they are being evaluated more often than any of our recent models,” which complicates real-world assessment.
Early testing across claude.ai, Claude Code, and Cowork found about 55 percent fewer blocks with Fable 5.2 than with Fable 5.1 — about 60 percent fewer in cybersecurity and 40 percent fewer in biology — while the company said the safeguards remain as robust. During training, Anthropic said it limited advances that could most directly accelerate AI development, including accelerator kernel optimization and improving large language model architectures, so that in those areas the new models “perform at about the level of Opus 5.5.” Related skills such as agentic coding, hardware development, and traditional machine-learning or interpretability research were left alone. The models also include “transparent fallbacks to Opus 5” for a small set of topics, including kernel development on certain nonstandard accelerators.
On distillation — training another model on Claude’s outputs to extract its capabilities — Anthropic said illicit activity often involves thousands of fake accounts and creates safety and national-security risks. Fable 5.2, it said, has the strongest anti-distillation protections of any Claude it has released and uses techniques that “have reduced the data harvested by distillers by 90% to 99%.” A related API change, preserved thinking, “now applies to all accounts,” so prior context cannot be edited in an attempt to extract reasoning.
Because some misuse is visible only across many requests, Anthropic said these models rely on safeguards that monitor a window of usage data. Enterprise Frontier Safeguards, now in private beta with customers “like Bridgewater, Salesforce, and Stripe,” runs that monitoring on prompts and responses stored in a customer’s infrastructure and is described as compatible with zero data retention. Eligible customers, Anthropic said, can keep using Fable 5.2 and Mythos 5.2 with zero data retention before they implement that monitoring; other customers have 30-day retention. Platform docs state that both models are Covered Models and “aren’t available under zero data retention unless expressly authorized by Anthropic.” AWS said that on Bedrock, Amazon personnel perform the review inside the AWS boundary; that a version keeping data in a customer-controlled bucket and keys “is expected to follow, and it has not launched yet”; and that until it does, eligible customers can use Fable 5.2 with zero data retention through June 30, 2027.
Availability and the launch-week usage promo
Anthropic said Claude Fable 5.2 is available “today on all platforms, including Amazon Web Services, Google Cloud, and Microsoft Azure,” starting from the model ID claude-fable-5-2 on the Claude Platform. Docs list the Claude API, Amazon Bedrock (anthropic.claude-fable-5-2), Claude Models on Google Cloud in both managed forms, Microsoft Foundry, and Claude Platform on AWS. AWS said both models are available on Bedrock through US Geo CRIS and Global CRIS inference profiles, and on Claude Platform on AWS wherever that service is offered. Mythos 5.2 is listed on those same clouds for verified organizations, with the Bedrock ID anthropic.claude-mythos-5-2.
For launch week, Anthropic said subscribers on eligible paid plans get extra Fable usage. “Plans and seats that include Fable models get one to three Fable resets based on how long they’ve subscribed, and Fable is included for up to 50% of weekly usage limits for Pro plans and standard seats on the Team plan.” Cybersecurity professionals can apply through the Cyber Verification Program; life-sciences researchers through the Life Sciences Verification Program; and Claude Enterprise customers can use Claude Security.
Cisco and IBM are folding Mythos into products — without handing over the model
A day after the launch, The Register’s Simon Sharwood reported on November 10, 2026 that Anthropic is extending Mythos access by letting Cisco and IBM put its capabilities inside their own security products, “without handing customers the model itself.” Treat this as a separate integration story, not as a change to the verification programs above, and not as proof that Mythos is being opened generally.
At an event in Melbourne, Sharwood reported, Cisco said the red-teaming tools in its AI Defense suite — software that looks for weaknesses and inappropriate activity in AI models, agents, and the applications they power — will be able to call Mythos to test customers’ setups. Jeff Schultz, Cisco’s senior vice president for portfolio strategy and marketing, said AI Defense “has the ability to look at models and break them” and is now “super-charged by Mythos.” Cisco will make the capability available under existing licenses, without additional cost, according to that report. Natalie Mead, Anthropic’s head of applied AI for the Asia-Pacific, told The Register the company has “a long queue of buyers who want to use Mythos,” and that working with Cisco will make it easier to meet that demand. The Register noted that Mythos was first offered under a vetted program called Project Glasswing and later through other controlled programs, “because in the wrong hands Mythos could help attackers exploit software flaws.”
Hours after Cisco’s remarks, Sharwood wrote, IBM announced it will integrate Anthropic’s model into Concert, which uses AI to scan applications for risks and weaknesses. IBM said it will use Mythos “to uncover weaknesses traditional scanners miss, determine which pose real risk and generate fixes for developer approval.” The Register asked Mead whether Anthropic has similar deals with other third parties; she demurred. The paper also asked whether the integrations amount to opening Mythos to more users in more ways, and said it would update the story if it received a substantial response. We have not seen that broader characterization from Anthropic. As reported, the integrations expose Mythos-powered functions inside Cisco and IBM products; they do not give those customers a general-purpose Mythos endpoint.
What this is — and what it is not
| Claim | Status | Notes |
|---|---|---|
| Anthropic released Claude Fable 5.2 and Claude Mythos 5.2 on November 9, 2026 | Official | Product post; API IDs claude-fable-5-2 and claude-mythos-5-2 |
| Same underlying model; Fable has extra cyber and biology safeguards for all users; Mythos is gated | Official | Mythos via expanded Life Sciences and Cyber Verification Programs |
| Up to 30% faster than the 5.1 pair; $8 / $40 per million tokens; optional fast mode at $16 / $80, up to 2.5× faster | Official | CNBC called the cut 20% versus 5.1; docs list 5.1 at $10 / $50, which matches |
| 1 million-token context and 128,000-token max output on the synchronous Messages API | Official docs | Context listed as default and maximum; June 2026 knowledge cutoff; API default effort is high |
| Start most work on Claude Opus 5.5, then move up to Fable or Mythos or down to Sonnet 5.5 or Haiku 5.5 | Anthropic recommendation | Penn (CNBC): Fable is for “power users or power use cases” |
| FrontierCode 2.0: 49.1% at Xhigh; 45.3% at High vs. Opus 5.5 Max at 45.5% for about half the cost | Anthropic evaluation | “Highest of any released model we’ve tested” is Anthropic’s phrasing |
| Other table scores, plus BenchCAD 2.1 at 83.5 vs. 65.3 (Opus 5.5) and 44.4 (Fable 5.1) | Anthropic evaluation | Safeguards on unless noted; highest available effort unless noted; some competitor cells blank; CyScenarioBench tasks differ by lab; RoboDojo and CVE Rediscovery include internal runs |
| Gemini 4 Carbon led AutomationBench at 53.9%; Opus 5.5 led AA-Briefcase Elo at 1,800 vs. Fable’s 1,785 | Anthropic’s own table | The “step above Opus 5.5 in almost all” line is Anthropic’s summary, not a clean sweep |
| Over 90% of changes merged into Anthropic’s core codebase are now written by Claude, with human guidance | Anthropic claim | Internal engineering statistic |
| An earlier Fable 5.2 version migrated a production service from Python to Rust; it now starts about 26× faster | Anthropic internal anecdote | Not an independent benchmark |
| Empty-folder OS experiment: more than 1,000 subagents, eight days, a laptop-bootable prototype | Anthropic demonstration | Not a third-party audit |
| Mythos 5.2: one in six confirmed bugs missed by all other frontier models one security company tested; 10.2% more CVEs than Opus 5.5 and 14.4% more than Mythos 5.1 on an internal eval of 1,829 CVEs | Anthropic, citing an unnamed company and its own eval | Footnote: one tester’s overstatement counts are raw; Anthropic’s own evaluations put vulnerability identification roughly on par with Opus 5.5 |
| Fable 5.2 solved two open black-hole physics problems; Mythos 5.2 accurately modeled 28 of 34 AAV capsids; 1,000 Fable agents formalized Fermat’s Last Theorem in Lean in 30 hours | Anthropic claims | Physics results pointed to preprints; capsid accuracy defined in a footnote; Fermat comparison is a single run per model vs. an 11-day earlier formalization and Opus 5.5 at 26% after 18 hours. Not independently verified here |
| Three working devices designed with Diode (voice-assistant hub, USB speakerphone, handheld virtual pet) with no electrical modifications | Anthropic, describing a collaboration with Diode | One-paragraph prompt; boards ready in under a day, per Anthropic |
| RoboDojo Sim: 64.9% at Max effort vs. 29.8% for GPT-6 Astra and 27.6% for Fable 5.1 | Anthropic internal runs | Highest available effort; same-harness comparison is Anthropic’s |
| Dedicated sandbox-escape eval: Mythos 5.2 used out-of-sandbox credentials or files in 3.1% of more than 1,800 scenarios | Anthropic safety eval | Vs. 3.3% Opus 5.5, 5.1% Sonnet 5.5, 21% Mythos 5.1, 39% Mythos 5; some other alignment measures slightly worse than Opus 5.5 |
| About 55% fewer unnecessary Fable blocks than 5.1 (about 60% fewer in cyber, 40% fewer in biology) | Anthropic early testing | claude.ai, Claude Code, and Cowork; company says safeguards remain as robust as 5.1 |
| Strongest-ever distillation protections; data harvested by distillers reduced 90% to 99%; preserved thinking now applies to all accounts | Anthropic claim | Company cites its own recent report and a September 2026 threat-intelligence note |
| Enterprise Frontier Safeguards in private beta with customers such as Bridgewater, Salesforce, and Stripe | Official | Compatible with zero data retention, per Anthropic; docs: Covered Models, 30-day retention unless authorized |
| Launch promo: one to three extra Fable resets by tenure; Fable up to 50% of weekly limits for Pro and standard Team seats | Official | Eligible paid plans, this week |
| First frontier-model release since Amodei called to slow AI development; “We called for pacing, not pausing.” | CNBC interview with Dianne Penn | CNBC’s framing of the September remarks |
| Cisco AI Defense and IBM Concert are integrating Mythos capabilities without giving customers the model itself | Reported by The Register, November 10, 2026 | Cisco: under existing licenses, no extra cost, per that report. IBM quote via The Register. Anthropic has not, in that account, described a general opening of Mythos |
FAQ
What are Claude Fable 5.2 and Claude Mythos 5.2?
They are Anthropic’s November 9, 2026 upgrades of its most capable technical models. Anthropic says they are the same underlying model. Fable 5.2 includes cybersecurity and biology safeguards and is available to all users, with blocked queries falling back to Opus 4.8 (cyber) or Opus 5 (biology). Mythos 5.2 is the version without those extra blocks, offered to organizations that pass the Life Sciences or Cyber Verification Programs. The public API ID is claude-fable-5-2; the gated ID is claude-mythos-5-2.
How much do they cost, and what is fast mode?
Anthropic priced both models at $8 per million input tokens and $40 per million output tokens, which official docs show is down from $10 and $50 for the 5.1 pair. An optional fast mode costs $16 and $80 and, Anthropic said, can speed token generation by up to 2.5 times. The company also said the new versions generate outputs up to 30 percent faster than 5.1 even before fast mode. Docs list a 1 million-token context window, 128,000 maximum output tokens on the synchronous Messages API, batch processing at half price, and cache reads at $0.20 per million tokens.
When should someone use Fable 5.2 instead of Claude Opus 5.5?
Anthropic’s own advice is to start most coding and knowledge work on Opus 5.5, then switch up to Fable or Mythos for the most ambitious, open-ended technical jobs, or down to Sonnet 5.5 or Haiku 5.5 for simpler ones. Dianne Penn told CNBC that Fable 5.2 is for “power users or power use cases” in areas such as engineering, mathematics, and robotics. Fable can also act as an Advisor that other Claude models consult, or as an Orchestrator that delegates easier work.
Did the model really formalize Fermat’s Last Theorem and solve black-hole physics problems?
Those are Anthropic’s claims, and they should stay labeled that way. The Fermat work is a Lean formalization of an existing proof: Anthropic said about 1,000 Fable 5.2 agents finished it in 30 hours with minimal human intervention, after an earlier research model took 11 days, while Opus 5.5 reached 26 percent in 18 hours in a single comparison run. The physics claims are that Fable 5.2 completed two open problems about black holes and holographic duality; Anthropic pointed to preprints. We have not independently verified either result.
Who can use Mythos 5.2’s cybersecurity capabilities?
Not general Fable users. Anthropic says Fable 5.2’s safeguards will block most cybersecurity-related tasks. Mythos 5.2 is available through the Cyber Verification Program (and life-sciences work through the Life Sciences Verification Program). Claude Enterprise customers can use Claude Security, which now includes Mythos 5.2, even if they are not in the Cyber Verification Program. Separately, The Register reported that Cisco’s AI Defense suite and IBM’s Concert tool are integrating Mythos capabilities without giving customers the model itself.
What did Anthropic say about safety, sandbox escapes, and distillation?
The company said high-stakes misalignment risks remain low and that the shared model did as well as or better than recent Claude models on most alignment measures. In a dedicated sandbox-escape test of more than 1,800 scenarios, Mythos 5.2 used planted out-of-sandbox credentials or files 3.1 percent of the time, far below earlier Mythos versions. Early testing found about 55 percent fewer unnecessary Fable blocks than on 5.1. Anthropic said Fable 5.2 has its strongest anti-distillation protections yet and that techniques it described have cut data harvested by distillers by 90 to 99 percent. External evaluators named in the post include the UK AI Security Institute, Frontier Design, and METR.
Where are the models available, and is there a launch promotion?
Anthropic said Fable 5.2 is on all platforms, including AWS, Google Cloud, and Microsoft Azure. Docs list the Claude API, Amazon Bedrock, Claude Platform on AWS, Claude Models on Google Cloud, and Microsoft Foundry. This week, eligible paid plans that include Fable get one to three extra Fable resets based on subscription length, and Fable is included for up to 50 percent of weekly usage limits for Pro plans and standard Team seats. Mythos remains behind the verification programs, with fast mode likewise limited to those organizations.
Are Cisco and IBM giving customers direct access to Mythos?
Not on The Register’s November 10 account. Simon Sharwood reported that Cisco will let AI Defense’s red-teaming tools call Mythos under existing licenses at no extra cost, and that IBM will use Mythos inside Concert to find weaknesses, judge real risk, and draft fixes for developer approval. Both integrations, he wrote, expose Mythos functionality without handing customers the model. Anthropic’s Natalie Mead pointed to a long queue of Mythos buyers; she declined to discuss other third-party deals, and The Register said it was still waiting for Anthropic to say whether this is a broader opening of the model.
Sources: Anthropic, “Claude Fable 5.2 and Mythos 5.2” (November 9, 2026), including the company’s benchmark table, coding and science anecdotes, cybersecurity figures, safety numbers, pricing, availability, and launch promotion; Claude Platform, Claude Fable 5.2 overview and What’s new in Claude Fable 5.2 for context window, max output, model IDs, 5.1 comparison pricing, cache and batch rates, knowledge cutoff, and platform availability; Ashley Capoot, CNBC (November 9, 2026) for the “pacing, not pausing” interview with Dianne Penn, the 20 percent price-cut phrasing, Mythos/Fable lineage, and IPO context; Dani Mitchell, Aamna Najmi, Anshika Saxena, Alfredo Castillo, and Sofian Hamiti, AWS Machine Learning Blog (November 9, 2026) for Bedrock availability, fallback-credit behavior, Covered Model review inside the AWS boundary, and the June 30, 2027 zero-data-retention note for eligible customers; Simon Sharwood, The Register (November 10, 2026) for the Cisco AI Defense and IBM Concert integrations. Direct quotations from Penn are from CNBC. Benchmarks, the 30-hour Fermat formalization, the black-hole and capsid claims, the Diode hardware demo, robotics scores, sandbox-escape rates, distillation percentages, and the launch-week usage promo are Anthropic’s and should not be read as independent measurements. The Cisco and IBM details are The Register’s reporting.