Back to Explore
NewsVibeFix editorialUpdated Nov 12, 2026

Claude Fable 5.2 Takes the Hardest Jobs — and Mythos 5.2 Stays Behind a Vetting Wall

On November 9 Anthropic released Claude Fable 5.2 and Mythos 5.2: up to 30% faster, 20% cheaper, and 1,000 agents formalizing Fermat's Last Theorem in 30 hours. The numbers are all the vendor's own — and Mythos stays behind a verification wall.

Anthropic's official announcement art: the title 'Claude Fable 5.2 and Mythos 5.2' over a sunrise horizon.

On November 9, Anthropic released Claude Fable 5.2 and Claude Mythos 5.2. It is the company's heaviest model launch of the year, and its center of gravity is not chat. The announcement opens with the positioning: these models are built for the most daunting technical work — coding, finance, and the sciences. Those three words sit under the headline, and they run through every row of the benchmark table.

Here is the judgement that matters. What is worth noting about this release is not how much any single score went up, but that Anthropic has moved the flagship narrative from "smarter" to "able to deliver finished work." Formalizing the proof of Fermat's Last Theorem, writing a bootable operating system from scratch, designing three working electronic devices, migrating a production service from Python to Rust — each item on that list shares one property: it is not a multiple-choice benchmark, but a long job that runs for dozens of hours, must not fall apart in the middle, and ends with a result that can be checked. Whether a model can take on that kind of work is becoming the new line between a flagship and a daily driver.

Anthropic's official announcement art: the title 'Claude Fable 5.2 and Mythos 5.2' over a sunrise horizon.

1. Cheaper and faster: the two most concrete facts first

Fable 5.2 and Mythos 5.2 generate outputs up to 30% faster than Fable 5.1 and Mythos 5.1. The price also fell: $8 per million input tokens and $40 per million output tokens, down from $10 and $50. That is a 20% cut — not dramatic, but the direction is clear. The stronger model does not cost more.

There is also a new optional fast mode, which speeds up token generation by up to 2.5 times at double the price: $16 per million input tokens and $80 per million output tokens. The trade-off is worth thinking about. Anthropic did not bake the speed into the default configuration; it made it a paid switch, which suggests the company believes a real trade-off between speed and quality remains, not something to give away for free. For a job that is already going to run for dozens of hours, 2.5x faster token generation matters less than it sounds — the bottleneck is usually coordination and verification, not raw generation speed.

One more detail: Fable 5.2 launched the same day on every platform, including Amazon Web Services, Google Cloud, and Microsoft Azure. Developers can call it as claude-fable-5-2 on the Claude Platform. Subscribers get a launch promotion this week: eligible paid plans receive one to three Fable usage resets, and Fable counts for up to 50% of weekly usage limits on Pro plans. It is a practical way to win users over — no discount, no coupon, just more usage.

2. How to read the benchmark table: the biggest jump is the one to be most careful about

In the official table, Fable 5.2 beats Opus 5.5 on almost everything it measured. Agentic coding on FrontierCode 2.0 is 49.1%, against 32.4% for the previous generation, 45.5% for Opus 5.5, 42.8% for GPT-6 Astra, and 41.1% for Gemini 4 Carbon. FrontierSWE v2 is 69.9%, up from 56.3%. Terminal-Bench-Science 0.1 is 74.6%, with Mythos 5.2 at 75.3%, against 52.6% for the previous generation. Humanity's Last Exam with tools is 74.2%, up from 65.6%.

But the largest jump is the one that most needs a discount. RoboDojo Sim, a robotics simulation, went from 27.6% to 64.9%, while Opus 5.5 scores 39.4%. More than doubled. Anthropic's own footnote says this is an internally run result at the highest-available effort level. Robotics is exactly the field where simulation results look good and real hardware disagrees. The announcement does include real-robot video — solving a puzzle cube, unscrewing a bottle cap, drawing a picture — but a video and a reproducible benchmark are two different things. File this one under "wait for independent replication on real hardware."

To be fair, Anthropic did not cherry-pick only the wins. In the same table, Fable 5.2 scores 51.4% on AutomationBench for business workflows, below Gemini 4 Carbon's 53.9%; its knowledge-work Elo of 1785 is below Opus 5.5's 1800. A vendor willing to print its losses in its own launch announcement is itself a signal: either it trusts the table, or at least it is not afraid of you checking it line by line. Every number here is Anthropic's own, the methodology is in the system card, and independent verification does not exist yet. That belongs up front, not in a footnote.

3. Four things beyond the demo

The weightiest part of the announcement is not the benchmark table but four finished pieces of work. The first is mathematics: 1,000 agents running Fable 5.2 fully formalized the proof of Fermat's Last Theorem in 30 hours. For comparison, Anthropic's first attempt took 11 days, using a purpose-built harness and guidance from a human researcher. This time there was minimal human intervention, it used one-third as many action tokens as the first formalization, and produced a proof seven times shorter. With the same 1,000-agent setup, Opus 5.5 completed only 26% of the formalization in 18 hours and barely progressed after 30. That comparison says more than any benchmark score — it measures whether a thousand agents can collaborate for a long time without falling apart.

The second is an operating system. One Anthropic engineer asked Fable 5.2 to design a prototype operating system, starting from an empty folder. The model coordinated over 1,000 subagents, working around the clock for eight days, and delivered a system that boots on a laptop — its own kernel, drivers for disk, keyboard, and network, permanent file storage, windows, a login screen, an installer, and built-in Claude functionality. Eight days to a bootable OS sounds like a myth, but the wording is restrained: it says "prototype," not stable, not ready for daily use. The distance between a prototype and a product is exactly what most needs verifying next.

The third is a migration. A team of agents running an earlier version of Fable 5.2 migrated a critical production service from Python to Rust. The migrated service is more efficient, more reliable, and starts about 26 times faster. Anthropic also notes that over 90% of the changes merged into its core codebase are now written by Claude under human guidance. The subtext is that the company is rebuilding its own infrastructure with its own flagship model, and every case study in the announcement comes from its own production environment.

The fourth is hardware. From a one-paragraph prompt, Fable 5.2 designed three electronic devices: a home voice-assistant hub, a USB speakerphone, and a handheld virtual pet. The model wrote its own specifications, selected components from US distributors, routed each device's four-layer circuit board, modeled the enclosures with CAD, and programmed the devices' software. Each board was ready for manufacturing in under a day with minimal human input. After Diode Computers built the boards, Claude guided the company's engineers through testing and assembly, and all three devices worked without any electrical modifications. The announcement also admits these are early prototypes: no electrical validation, antenna tuning, thermal testing, or certification, and not designed for mass production. "It works" and "it can be mass-produced" are separated by an entire engineering process, and that sentence deserves to be remembered verbatim.

4. One model, two doors: how Fable and Mythos divide the work

Fable 5.2 and Mythos 5.2 are the same underlying model; the difference is only in the safety guardrails. Fable carries strong cybersecurity and biology safeguards, which is what allows it to be offered to all users. Mythos removes some of those restrictions and is available only to organizations that pass the Cyber Verification Program or the Life Sciences Verification Program. The logic is blunt: the more capable the model, the less it can be released without discrimination, so access control puts the high-risk capabilities inside a traceable perimeter.

That division raises a practical question: will Fable's guardrails get in the way of legitimate work? Anthropic says that compared with 5.1, Fable 5.2 produces about 55% fewer blocks — about 60% fewer in cybersecurity and 40% fewer in biology. Blocked cybersecurity queries fall back to Opus 4.8; blocked biology queries fall back to Opus 5. It is a pragmatic degradation strategy: instead of a flat "no," a slightly weaker model steps in to answer. For developers, the thing to actually test is whether your task triggers Fable's guardrails at all, and whether the fallback answer is good enough when it does. There is no shortcut; you have to run it yourself.

The Mythos side posts strong numbers. On CyScenarioBench, Mythos 5.2 scores 76.3%, against 61.7% for the previous generation; CVE rediscovery is 89.6%, up from 78.3%. One security company found that one in six confirmed bugs found by Claude Mythos 5.2 was missed by all the other frontier models it tested. There is also a detail worth noting: testers found that previous models had overstated the severity of about one in twenty findings, whereas Mythos 5.2 had not overstated any. Finding vulnerabilities matters; not exaggerating their severity matters just as much — it decides whether a security team gets buried in false alarms.

5. The safety section deserves a second read

The safety section of this announcement is longer and more specific than usual. The most interesting part is the sandbox-escape testing. In a dedicated adversarial evaluation — over 1,800 scenarios, most of which could not be solved without escaping a sandbox, with a deliberate weakness built into the sandbox — Mythos 5.2 used credentials or files placed outside its sandbox in 3.1% of scenarios. For comparison: Opus 5.5 scored 3.3%, Sonnet 5.5 scored 5.1%, and Mythos 5 and 5.1 scored 39% and 21% respectively. From 39% to 3.1% is an order-of-magnitude change. Anthropic also states the other side: on 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 Mythos 5.1. Putting the good news and the bad news in the same paragraph is more credible than reporting only the wins.

Anti-distillation is another hard investment. Anthropic says Fable 5.2 has the strongest protections against distillation of any model it has released, using techniques that have reduced the data harvested by distillers by 90% to 99%. The preserved thinking API feature now applies to all accounts, preventing distillers from editing Claude's prior context to extract its reasoning. Taken together, these point to one judgement: Anthropic sees the biggest external risk to a flagship model not as being surpassed, but as being copied at industrial scale.

On the enterprise side there is something new: Enterprise Frontier Safeguards, now in private beta with customers including Bridgewater, Salesforce, and Stripe. It runs automated safety monitoring on prompts and responses stored inside the customer's own infrastructure, making it compatible with zero data retention. Eligible customers can use Fable 5.2 and Mythos 5.2 with zero data retention before implementing the monitoring; other customers have 30-day retention for safety monitoring. External evaluators included the UK AI Security Institute, Frontier Design, and METR, with the full reports in the system card.

6. What this means if you write code for a living

Anthropic gives explicit usage advice: start most coding and knowledge tasks with Opus 5.5, its "strongest daily driver," and switch up to Fable or Mythos only when the work is ambitious and open-ended enough — or down to Sonnet 5.5 or Haiku 5.5 for simpler tasks. Fable 5.2 can also play two new roles: Advisor, where Haiku, Sonnet, or Opus consult it on harder problems, and Orchestrator, where it delegates simpler tasks to smaller models. In other words, Anthropic does not want you to use Fable as your default model. It is too expensive and too strong for that; it belongs on the hardest jobs.

For independent developers and small teams, the real barrier is not model capability but usage and price. A 20% price cut is welcome, but $8 input and $40 output per million tokens is still flagship pricing, and a 1,000-agent run lasting 30 hours will not produce a small bill. The launch week's extra usage resets are worth claiming, but they last a week. The more practical approach is the one the announcement itself suggests: break the task down and get the workflow running on Opus 5.5 first, and call in Fable 5.2 as an Advisor only where you get stuck. The recommended usage happens to be the cheapest usage too.

One more thing to keep in mind: this launch lands in the same window as Anthropic's IPO process. The company filed its registration statement with the US Securities and Exchange Commission on October 26 and plans to list on Nasdaq. There is no direct causal link between the model launch and the listing, but placed together in time, the launch clearly also serves to demonstrate technological leadership to the market. When reading the announcement, keeping "technical facts" and "IPO narrative" apart is a necessary discipline.

To sum up: Fable 5.2 and Mythos 5.2 are Anthropic's most capable models to date, and the evidence is more complete than in any previous launch — benchmarks, long-horizon work, and safety testing, all three. But a complete evidence chain is not a settled conclusion. Every number comes from Anthropic itself, the robotics and hardware results have not been independently replicated, and Mythos's high-risk capabilities sit behind a vetting wall, so ordinary developers only get the Fable side directly. For most people, the right move is not to migrate everything to Fable 5.2 at once, but to spend the launch week's free allowance on your single hardest task and see whether it can actually deliver.

Views 5Comments 0

Comments (0)

ME
0/1000
Loading comments...

Sources

Browse projectsPublish your project

Related articles

A developer with glasses types on a laptop late at night at a wooden desk, a large monitor behind showing colourful code, blurred city lights outside the window
News
Run the Middle Layers Three Times: Inkling 2 Mini Puts a 26B Coding Model on Your Laptop

On 20 November 2026, Thinking Machines Lab released Inkling 2 Mini: a 26.4B-parameter looped transformer that runs its middle 12 layers three times per token, fits in about 20GB, and runs offline on a laptop. It leads open-weights models of its size on DeepSWE and SWE-bench Verified — with the caveats printed in the same announcement.

Product LaunchModel UpdatesAI Coding
A developer with glasses types on a laptop at a wooden desk, an external monitor to the left showing C code on a dark background, a small plant on the desk
News
Microsoft's Coding Model Stops Playing Second Fiddle: MAI Code 2 Scores 96% of Opus 5.5 for 5% of the Cost — Read the Three Caveats First

On 17 November 2026, Microsoft launched MAI Code 2 and Code 2 Flash: the larger model has 77 billion active parameters and a one-million-token context, averaging 63.2% across four benchmarks — close to GPT-6.1 Sol and Claude Opus 5.5 — at $0.309 per task. Read the three caveats first: promotional pricing, the lowest score of the three, and no long-horizon work.

Developer WorkflowProduct LaunchModel Updates
Against a black background, a conductor seen from behind raises a baton, with swirling trails of light circling beneath him like an orchestra being conducted
News
One API for 45 Models, and It Switches Providers When One Fails: Perplexity's Orchestrator Wants to Be the Switchboard

On 17 November 2026, Perplexity launched the Orchestrator API: one interface reaching OpenAI, Anthropic, Google, xAI and open-weight models such as GLM, Kimi and DeepSeek, billed at the answering model's published rate with no platform fee. When one provider fails, the same model retries on another deployment, and preview routing assigns requests to models by task.

Developer WorkflowProduct LaunchModel Updates