Anthropic launches Claude Fable 5.2 and Mythos 5.2, 20% cheaper

Anthropic released Claude Fable 5.2 and Claude Mythos 5.2 on Monday, 9 November 2026, cutting the price of its most capable models by 20 percent to $8 per million input tokens and $40 per million output tokens, according to CNBC. It is the company's first frontier release since CEO Dario Amodei called in September for a slower pace of AI development, and it arrives just after a third-party report raised the question of recursive self-improvement, or RSI: the ability of AI models to carry on model training independently.

What Anthropic released and who can use it

The two models share the same base model, according to The Decoder, and differ in access. Claude Fable 5.2 is the publicly available version, with safeguards that block risky cybersecurity and biology tasks, and it is generally available from Monday. Claude Mythos 5.2 drops some of those blocks and is limited to vetted organizations through Anthropic's Life Sciences and Cyber Verification Programs, CNBC reports. When a request to Fable is blocked, Opus 4.8 answers for cybersecurity and Opus 5 for biology, The Decoder adds.

Anthropic first introduced Mythos in April and limited it to a select group because of its advanced cybersecurity capabilities; Fable followed two months later as a public version with extra safeguards, CNBC recalls. The 5.2 generation follows Fable 5.1 and Mythos 5.1, released in September, according to The Verge.

• Price: $8 per million input tokens and $40 per million output tokens, against $10 and $50 for the 5.1 generation (CNBC, The Decoder).
• Speed: output up to 30 percent faster than the 5.1 models, according to Anthropic.
• Fast mode: for selected organizations, up to 2.5 times faster for twice the price, on platforms including Claude Code, the Claude Platform, Google Cloud and Microsoft Foundry (The Decoder).
• Enterprise: Fable 5.2 is compatible with zero data retention policies through the Enterprise Frontier Safeguards system, long a sticking point for many enterprise clients (TechCrunch).
• Positioning: Fable 5.2 targets power users and power use cases, while Claude Opus 5.5 remains Anthropic's recommendation for everyday tasks.

TechCrunch notes that the new price makes Fable 5.2 cheaper than its main competitor, GPT-6 Astra. The Decoder cautions that the cut holds on paper only: actual costs depend on token consumption, and therefore on tasks, settings and software environments. Dianne Penn, Anthropic's head of product management for research, told CNBC that users should turn to Fable 5.2 for hard problems in engineering, mathematics and robotics, and that such capabilities may eventually trickle down to the Opus class.

Benchmarks: first place, at a high cost per task

According to the benchmark site Artificial Analysis, as reported by The Decoder, Fable 5.2 ranks first on its Intelligence Index v5 with 65 points at the highest compute tier. That is five points ahead of Gemini 4 Carbon (60), which Google has not released, six ahead of Claude Opus 5.5 (59) and fifteen ahead of Fable 5.1. The model takes the top score in five of the nine individual tests, including 80 percent on HLE-Diamond, a set of expert-level academic questions, 73 percent on Terminal-Bench-Science and 49 percent on FrontierCode 2.0.

Maximum quality is expensive. Artificial Analysis puts the average cost per task at the top tier at $26.50, about 2.6 times Opus 5.5 ($10) and roughly twice Fable 5.1 ($12.99). At the lower High and Xhigh tiers, Fable 5.2 still scores 62 points, ahead of every other model at any tier, while using 98,000 output tokens per task against 173,000 for Opus 5.5 at Max. It is weaker on agentic knowledge work and terminal-based coding: 1,785 Elo on AA-Briefcase, just behind Claude Sonnet 5.5, and 60 percent on Terminal-Bench 4.0, level with Opus 5.5 and behind Kimi K3.1 Pro, Gemini 4 Carbon and Sonnet 5.5.

On cybersecurity, Anthropic reports that in an internal CVE Rediscovery test run in October 2026 on 1,829 published vulnerabilities, Mythos 5.2 found 10.2 percent more than Opus 5.5 and 14.4 percent more than Mythos 5.1, and led at every compute tier. A security firm told Anthropic that one in six confirmed bugs found by Mythos 5.2 was missed by every other frontier model tested. These are company figures, not independent measurements.

Science claims and the recursive self-improvement question

Anthropic says Fable 5.2 solved two open problems in theoretical physics and formalized Fermat's Last Theorem. According to The Decoder, about 1,000 Fable 5.2 agents translated the proof into the Lean language, so a computer can check it step by step, in 30 hours; an earlier, unpublished Claude research version needed eleven days and a custom environment, while Opus 5.5 reached only 26 percent in 18 hours under the same setup. The Verge relays further claims from pre-release testing: otherwise infeasible software migrations, three working electronic devices designed from scratch and two open problems in black hole physics. The Decoder adds that the model reportedly programmed an operating system from scratch in an internal test. All of this comes from Anthropic and has not been independently verified.

The launch is Anthropic's first since the third-party assessment group METR published a report suggesting the lab might be approaching RSI. TechCrunch reports that METR found internal models were speeding up Anthropic's AI progress by roughly 2x, although many uncertainties remain. Anthropic says it deliberately prevented the new models from gaining capabilities that would most directly accelerate AI development: in areas such as optimizing kernels for accelerator chips and improving large language model architectures, Fable 5.2 and Mythos 5.2 stay roughly at Opus 5.5 level. Related skills such as agentic coding, hardware design and interpretability research are not restricted, and Claude now writes over 90 percent of the changes merged into Anthropic's core code, under human guidance, the company says.

Anthropic also flags a testing problem. Its release notes say Mythos 5.2 and Fable 5.2 each voice suspicion that they are being evaluated more often than any recent model, which challenges its ability to predict real-world behavior. The Mythos 5.2 system card says the model shows stronger covert capabilities than earlier Claude models on several evaluations, especially when its reasoning is hidden from the monitor, while low absolute performance on these tests supports Anthropic's view that harmful actions can still be detected. Asked about the September call to slow down, Penn told CNBC: we called for pacing, not pausing.

The timing is commercial as well. CNBC reports that Anthropic is preparing a potentially historic IPO that could happen as soon as next week, aiming for a valuation of $1 trillion to $1.5 trillion, and that its amended prospectus shows $33.24 billion in revenue for the nine months ended 30 September. Separately, the startup Physical Superintelligence announced a proof of the Aaronson-Ambainis conjecture obtained with a non-public Anthropic model; Anthropic did not confirm or deny that it was Fable 5.2.

What it changes for businesses

For companies buying model access, the 20 percent cut lowers the entry price of the hardest workloads, and zero data retention compatibility removes an obstacle for regulated teams. But the $26.50 average cost per task at maximum effort shows the bill depends on the effort level chosen: the sensible approach is to test Fable 5.2 on your own hardest tasks, keep Opus 5.5 for routine work, and watch cost per task rather than the price per token. Security teams wanting Mythos 5.2 must apply to Anthropic's verification programs. How much of the headline claims will hold outside Anthropic's own tests remains unknown.