AI
Anthropic cuts Claude Fable 5.2 pricing by 20% while boosting coding performance and speed
The latest version of Anthropic's flagship model delivers sharper reasoning, faster output, and a significant price reduction to compete with OpenAI's GPT-6 Astra.
1/3
The New Stack
Anthropic rolled out Claude Fable 5.2 and Mythos 5.2 on Monday, bringing notable improvements across coding performance, inference speed, and cost-efficiency. The new release arrives alongside a substantial pricing reduction that narrows the gap with rival models.
Price and speed improvements#
Fable 5.2 now costs $8/$40 per million input/output tokens—down from $10/$50—bringing it 20% cheaper than OpenAI's GPT-6 Astra at standard rates. Mythos 5.2 adopts the same pricing structure, though it remains accessible primarily through Anthropic's specialized verification programs for cybersecurity and life sciences organizations.
Beyond the sticker price, both models generate responses 30% faster than their predecessors. A new Fast mode option accelerates token generation further—up to 2.5 times quicker—though it comes at premium rates ($16/$80 per million tokens), effectively doubling the standard cost. Fast mode is available through Claude Code, the Claude Platform, Google Cloud Models, and Anthropic-hosted Microsoft Foundry.
Anthropic emphasizes that while Fable's standard rates remain twice those of Opus 5.5, real-world cost-per-task comparisons show Fable 5.2 at High effort can match Opus 5.5 at Max effort for roughly half the expense.
Coding and reasoning benchmarks#
The model demonstrates substantial gains on complex coding tasks. On FrontierCode 2.0, Fable 5.2 achieved 49.1%—a dramatic leap from Fable 5.1's 32.4% and outpacing both Opus 5.5 (45.5%) and GPT-6 Astra (42.8%). It also leads on FrontierSWE v2 with 69.9%, ahead of Astra's 65.5% and Opus's 62.3%.
The company's internal tests included a striking demonstration: a single engineer asked Fable 5.2 to construct a prototype operating system from scratch. The model orchestrated over 1,000 sub-agents across eight days, producing a working OS with drivers, persistent storage, a graphical interface, and an installer.
Not every benchmark favors Fable, however. Google's Gemini 4 Carbon leads on AutomationBench (business workflow tasks) at 53.9% versus 51.4%, and Opus 5.5 still holds a slight edge on the Elo measure for knowledge-work tasks.
Safety, access, and reduced friction#
Anthropic maintained the same robust cybersecurity and biology safeguards in Fable 5.2 as in its predecessor. However, testing revealed approximately 55% fewer safety blocks overall—60% fewer in cybersecurity contexts and 40% fewer for biology queries. High-risk activities like penetration testing and exploit generation remain blocked.

