Anthropic releases Claude Fable 5.1 and Mythos 5.1: coding benchmarks double, cache reads 75% cheaper

The capability gains of this generation first show up in benchmarks. Anthropic's data shows Fable 5.1 scores 52.6% on Terminal-Bench-Science 0.1 — a test that measures the ability to complete research tasks from the command line — versus 24.7% for the previous-generation Fable 5, more than doubling the score; for comparison, Claude Opus 5 scores 29.0% and OpenAI's GPT-5.6 Sol scores 22.4% in their evaluation setup. On the more classic coding benchmark Terminal-Bench 4.0, Fable 5.1 scores 55.8% (versus 42.0% for Fable 5 and 52.3% for Opus 5), while Mythos 5.1, with more permissive safeguards, reaches 60.9% on the same coding benchmark, significantly beating OpenAI's GPT-5.6 Sol at 37.3%. On AutomationBench, which measures business workflow automation, Fable 5.1 also jumps from 17.1% to 31.4%. Anthropic also notes that Fable 5.1's knowledge cutoff has been updated to June 2026, five months later than the previous generation.

Besides performance, pricing and cost structure are another highlight of this upgrade. Anthropic cut cache read prices for the Fable line from $1.00 to $0.25 per million tokens, a 75% reduction. Because coding and agentic workloads rely heavily on cache hits, this price cut has a significant impact on real-world costs: Anthropic estimates that overall costs for typical workloads can drop by about 25%, while highly agentic workloads can see cost reductions of up to 45%. At the same time, the base pricing remains unchanged — $10 per million input tokens and $50 per million output tokens. Even more notable is the shift in cost-efficiency curves: Anthropic claims that running Fable 5.1 at 'low' or 'medium' effort produces results comparable to or better than Fable 5 at full effort, at a meaningfully lower cost. Default effort levels also vary by surface: Fable 5.1 runs at high effort by default in Claude Code, but medium in Claude Cowork and on Claude.ai.

Around this release, Anthropic also signaled progress on ecosystem and safety fronts. On the ecosystem side, SiliconANGLE reports that Anthropic signed a $35 billion cloud deal with Lambda to provide compute for model training and inference — extending Anthropic's path of multi-billion-dollar compute partnerships with Google Cloud and AWS. On the safety side, Anthropic says it invited two external security firms and Gray Swan, a specialist adversarial testing company, to attack the new models' safeguards and found no critical jailbreak vulnerabilities; Mythos 5.1 is restricted to trusted organizations precisely because it permits operations that would otherwise be blocked in sensitive domains like cybersecurity and life sciences. According to Decrypt, Fable 5.1 is not included in the Claude Pro subscription plan; users need API access or a higher subscription tier, reflecting Anthropic's increasingly granular commercialization strategy for its flagship models.

For developers, enterprises, and everyday AI users, the release of Claude Fable 5.1 sends several clear signals. First, competition among frontier models is shifting from 'who is more capable' to 'who is more economical': cutting cache costs by 75% while matching or beating rivals on capability means the price war has reached inference and agentic workloads — the most expensive part of the stack. Second, tiering the same model by safety level (general Fable + restricted Mythos) is becoming a standard move for leading labs responding to safety concerns, serving the professional needs of high-risk domains while maintaining compliance baselines for general products. Third, Anthropic's deep partnership with second-tier cloud providers like Lambda shows compute supply is spreading from hyperscalers to a more diverse cloud ecosystem, with long-term implications for the AI infrastructure market. What's worth watching next: whether Fable 5.1 performs as well in real coding projects as its benchmarks suggest, which organizations will gain access to Mythos 5.1, and how OpenAI's GPT-5.6 Sol will respond to this dual-front competition on price and performance.

Looking back at this release, the value of Claude Fable 5.1 is more than 'just another stronger model.' It combines doubled performance with dramatically lower costs in a single update: Terminal-Bench-Science 0.1 jumped from 24.7% to 52.6%, cache read prices dropped 75%, and highly agentic workloads can see up to 45% cost reductions — meaning developers running long-lived coding agents can genuinely save real money. At the same time, the 'same model, different tiers' strategy of Fable 5.1 and Mythos 5.1 offers an actionable template for AI safety governance: releasing the most powerful capabilities in controlled environments while maintaining safety baselines across the general product line. For ordinary users, a cheaper and stronger coding model means the barrier to AI-assisted programming keeps falling; for the industry, the three-in-one competition of capability, cost, and safety is pushing the large-model race into a more mature phase.