Anthropic released Claude Fable 5.1 and Claude Mythos 5.1 on Tuesday and cut cache-read pricing on the flagship 75 percent, dropping the rate from $1.00 per million tokens on Fable 5 to $0.25 on the new model. That single line item is the story. Two benchmark leaps and a $35 billion cloud contract with Lambda follow from a repricing designed to make persistent agent workloads viable at the low end of the market.

The math is unusual. A cached input token on Fable 5.1 costs 2.5 percent of the uncached input price, versus the roughly 10 percent multiplier that has defined the Claude line. Anthropic estimates typical workloads get about 25 percent cheaper, and highly agentic workloads, the kind that reread long context windows on every step, fall as much as 45 percent. Uncached input stays at $10 per million tokens, output at $50. Five-minute cache writes are $12.50, one-hour $20. Batch processing halves both to $5 and $25. A 1.1x multiplier applies to U.S.-only inference; web search runs $10 per thousand queries.

Fable’s cache-read now sits 25 percent above Sonnet 5’s $0.20, though uncached Fable input still costs five times Sonnet’s. The structure pushes developers toward long-lived, context-heavy agents rather than one-shot calls.

Benchmarks moved with the price. Fable 5.1 posts 31.4 percent on AutomationBench, nearly doubling Fable 5’s 17.1 and clearing Opus 5’s 26.9. On Terminal-Bench-Science 0.1 it scores 52.6 percent against 24.7 for Fable 5, 29.0 for Opus 5, and 22.4 for GPT-5.6 Sol in Anthropic’s setup, with a ±3.5–4.5 point standard error. GDPval-AA v2 lands at 1,853, ahead of Opus 5’s 1,824 and Fable 5’s 1,723. Browserbase reported Fable 5.1 finishing 82 percent of tasks on its hardest browser-agent benchmark, versus 74 for Opus 5 and 57 for Fable 5. Ramp described an unattended 38-hour machine-learning run in which the model re-evaluated a previous result and launched six experiments. Millennium said it traced a crash bug that had resisted explanation for four to five years. Cybersecurity safeguards produce about 60 percent fewer false positives per Claude Code session.

Competitive context matters. OpenAI’s GPT-5.6 Sol carries promotional pricing of $4 in and $0.40 cached through at least Nov. 21, and Google’s Gemini 3.7 Flash sits at $0.75 in and $3.75 out through the end of 2026. Fable 5 reached only about 11 percent of Anthropic model spending more than two months after launch, a figure the 5.1 pricing is built to change.

The infrastructure buildout confirms the direction. Anthropic signed a $35 billion agreement with Lambda anchored on Hut 8’s one-gigawatt campus, expected to reach the grid in Q1 2027 with the first data hall online roughly six months later, alongside a $45 billion, 460-megawatt contract with Nscale. A cache-read cut of this size is a supply-side commitment: you don’t reprice the cheap path unless you’re planning to sell substantially more of it. The 2016 AWS Lambda debut followed the same pattern, cutting the marginal cost of an idle worker to near zero and forcing an architectural shift. Cache-read economics on frontier models are now inside that neighborhood.

Sources