Anthropic’s Fable 5.2 Takes Aim at the AI Frontier

Anthropic has answered OpenAI’s GPT-6 Astra with two new models that aim higher while charging less: Claude Fable 5.2 and Claude Mythos 5.2. Released on November 9, 2026, the models target coding, reasoning, vision, cybersecurity, and research work that Anthropic says once looked impossible.

The launch delivers a clear message from Anthropic: “We called for pacing, not pausing.” Fable 5.2 brings the broad release, while Mythos 5.2 focuses on vetted cybersecurity professionals and life sciences researchers. Together, they mark Anthropic’s first major model release since the assessment group METR raised questions about recursive self-improvement, or RSI.

More Capability, Lower Cost

Anthropic says Fable 5.2 and Mythos 5.2 show improvements in coding, reasoning, and vision, while communicating more clearly with users. The company claims the new models perform better than competitors on coding evaluation benchmarks and offer higher performance than GPT-6 Astra at a lower price point.

Both models cost $8 per million uncached input tokens and $40 per million output tokens. That cuts Fable’s input price from $10 to $8 and its output price from $50 to $40 compared with the 5.1 generation. The rates sit 20% below GPT-6 Astra’s standard short-context pricing, while cached tokens cost 20 cents per million compared with one dollar per million for Astra.

Fable 5.2 also returns responses 30% faster across effort levels. Faster inference serving is available for double the price, giving teams another way to trade cost against speed. Fable supports zero data retention policies through Anthropic’s Enterprise Frontier Safeguards system.

Anthropic recommends Opus 5.5 for daily tasks and positions Fable 5.2 for the “most ambitious, open-ended technical work.” That distinction matters because Fable still costs twice as much as Opus 5.5 at standard rates, yet its stronger reasoning can make the higher price useful for difficult engineering jobs.

The Numbers Behind the Launch

Fable 5.2 debuted at No. 1 on Artificial Analysis’ Intelligence Index leaderboard, edging out Google’s still unreleased Gemini 4 Carbon by 5 points. In Anthropic’s launch comparison, one of the two new Claude models leads on 13 of 15 reported score lines.

The clearest gains appear in software development. FrontierCode 2.0 tests whether a patch meets the standard for merging into a real open-source repository, and Fable’s score rose across every effort level:

• Medium: 37.6% at $2.48 per task
• High: 45.3% at $6.72 per task
• Xhigh: 49.1% at $14.19 per task

Fable at High scored 2.5 points above GPT-6 Astra at Max while costing almost the same per task. Astra spent $6.68 at Max, while Fable averaged $6.72 at High. Opus 5.5 reached a nearly identical score at Max, but averaged $12.82 per task.

The higher setting does not always deliver the best value. Fable at Max averaged $44.70 and scored below its cheaper Xhigh setting, while Fable’s Max result would fall from 45.7% to 44.4% if timed-out runs received zero credit. Medium remains the default in Claude apps and Claude Code.

Fable 5.2 also scored 69.9% on FrontierSWE v2, ahead of GPT-6 Astra at 65.5% and Carbon at 61.2%. On BenchCAD 2.1 with tools, Fable reached 83.5 out of 100 at Max, compared with 65.3 for Opus 5.5. On AA-Briefcase, Fable led the rubric pass rate 66.1% to 61.2%, though Opus held the head-to-head Elo lead, 1,800 to 1,785.

From Software Migrations to Mathematical Proofs

Anthropic’s strongest claims reach beyond benchmark tables. The company says the new models helped engineers complete otherwise-infeasible software migrations, designed three working electronic devices from scratch, solved two open problems in the physics of black holes, and made substantial progress on robotics evaluations.

A team of 1,000 agents powered by Fable 5.2 formalized the proof of Fermat’s Last Theorem in 30 hours. The result translated a mathematical proof into a format that a computer can check step by step, compared with 11 days for a previous, unreleased research version of Claude. Anthropic also says teams of 1,000 Fable agents built an operating system in 8 days.

Fable 5.2 was available to select parties before launch, and Anthropic says it already “solved two open problems in theoretical physics, formalized Fermat’s Last Theorem, and supported numerous other research projects.” Physical Superintelligence also announced a new proof of the Aaronson‑Ambainis conjecture, saying it used “a non-public Anthropic model” inside its own system. Anthropic did not confirm or deny whether Fable 5.2 was that model.

Access, Guardrails, and the RSI Question

Fable 5.2 is available on paid Claude plans, Claude Code, the Claude Platform, AWS, Google Cloud, and Microsoft Azure. Its model ID is claude-fable-5-2. Mythos 5.2 uses the same underlying model with different safeguards and remains application-only for vetted cybersecurity professionals and life sciences researchers in Anthropic’s Project Glasswing program.

Claude Enterprise customers can access Mythos through Anthropic’s Claude Security product. Anthropic says safeguards block “high-risk cyber activities including penetration testing, exploit generation, and binary-based vulnerability, as well as most tasks in research biology.”

Safety remains central because METR found internal models were speeding up Anthropic’s AI progress by roughly 2x and suggested the lab might be approaching recursive self-improvement. Anthropic’s release notes also acknowledge ongoing challenges in testing model capabilities accurately.

Anthropic maintains that Mythos 5.2 cannot act covertly against its user, but its system card says the model “shows stronger covert capabilities than prior Claude models on several of our evaluations, particularly when its reasoning is hidden from the monitor.” The same document says low absolute performance on new covert evaluations supports claims based on detecting harmful actions in high-stakes situations.

That tension now sits at the center of the Fable 5.2 story. Anthropic has lowered prices, increased speed, and pushed performance into software engineering, science, and autonomous agent work. The next test is whether these gains can scale while the company keeps its promise to pace the technology without pausing its progress.

Stay connected via Google News

Follow us for the latest AI News