Anthropic releases Claude Fable 5.2 and Claude Mythos 5.2, with Fable 5.2 priced at $8/$40

Anthropic introduced Claude Fable 5.2 and Claude Mythos 5.2, which it says are a step above Opus 5.5 in almost all the capabilities it evaluated, with improvements in coding, reasoning and vision. It says they generate output up to 30% faster than Fable 5.1 and Mythos 5.1.

Fable and Mythos. Anthropic says the two are the same underlying model. Fable has cybersecurity and biology safeguards, which is why it can be offered to all users; Mythos is for organizations that apply through its expanded Life Sciences and Cyber Verification Programs. It says Fable 5.2's safeguards are as robust as Fable 5.1's but produce fewer unnecessary blocks, and that Claude Security (Claude Enterprise) now uses Mythos 5.2. For most coding and knowledge tasks Anthropic still recommends starting with Opus 5.5, and says Fable 5.2 can also act as an advisor to smaller models or as an orchestrator that delegates to them.

Availability and price. Anthropic says Fable 5.2 is available today on all paid plans, the Claude Platform, Claude Code and all major cloud platforms. It costs $8 per million input tokens and $40 per million output tokens, down from $10 and $50. Anthropic's API pricing page lists Mythos 5.2 at the same $8 and $40, with cache reads at $0.20 and cache writes at $10 per million tokens for both models. OpenRouter lists Fable 5.2 with a 1M-token context at the same prices, and Cognition added it to Devin Desktop and Devin CLI. Cursor says it is available in Cursor too.

Effort and fast mode. Effort defaults to high in the Claude Platform and medium in Claude apps and Claude Code; Anthropic advises saving xhigh for the hardest long-running agentic work and says max is not recommended for everyday use. Fast mode is available for Fable 5.2 at 2.5x faster output for twice the price: set "speed": "fast" in the API, or toggle it in Claude Code's model picker, where it bills through extra usage credits.

API change. To help prevent distillation, preserved thinking now covers all API accounts on Fable 5.2. Anthropic says editing prior context can return a 400 error; it recommends compaction, context editing or mid-conversation system messages. Claude Code, Managed Agents and the Agent SDK handle this automatically.

Usage allowances. Ultra and Team Premium subscribers get Fable resets by how long they have subscribed: one for under 6 months, two for 6 to 12 months, three for over 12 months, usable until March 8. Pro and Team Standard subscribers get a week to try Fable 5.2, up to half their weekly limit, plus one reset.

Benchmarks.

• Anthropic says that on FrontierCode 2.0 in early testing, Fable 5.2 at High effort scored about level with Opus 5.5 at Max effort for roughly half the cost per task, and at Xhigh scored the highest of any released model tested.
• Cognition reports Fable 5.2 at #1 on FrontierCode 2.0 with 49.1.
• Cursor reports Fable 5.2 at 60% on CursorBench at max effort, the top model on it, and 56.9% at default effort, up from 49.2% for Fable 5.1.
• Anthropic says it outperformed other models on RoboDojo, a simulated two-arm robot benchmark.
• In Anthropic's own table, Fable 5.2 scores 69.9% on FrontierSWE v2 (Opus 5.5: 62.3%, GPT-6 Astra: 65.5%) and 89.3% on OSWorld 2.1 (offline set, partial). Mythos 5.2 scores 89.6% on Anthropic's internal CVE Rediscovery evaluation, which it says is 14.4% more vulnerabilities than Mythos 5.1.
• Anthropic says that, in prerelease testing, a team of 1,000 Fable 5.2 agents formalized the proof of Fermat's Last Theorem in 30 hours, and that it solved two open problems on black holes in theoretical physics.
• Artificial Analysis scores it 65 at max effort on its Intelligence Index v5, #1, ahead of Gemini 4 Carbon (60, high) and Claude Opus 5.5 (59, max). It says high effort scores 62 with about 40% fewer output tokens than Opus 5.5 at max, and that Fable 5.2 leads 5 of the index's 9 evaluations, including 80% on HLE-Diamond. Its model page lists a cost of $26.54 and 44.1 minutes per Intelligence Index task at max effort, and 275M output tokens across the index, which it calls very verbose.