Anthropic shipped Claude Fable 5.1 this afternoon (GA today on the API, AWS, Google Cloud, and Azure), with the gated Mythos 5.1 — the same underlying model under stricter safeguards — held to trusted-access programs for cyber and life-sciences work. Base prices don’t move: $10 in / $50 out per million tokens. Cache reads do: $1.00 down to $0.25 per million, a 75 percent cut that lands roughly 25 percent off typical workloads and up to ~45 percent off context-heavy, tool-heavy agent loops. This is the closed frontier answering the open layer that shipped unhedged on cost, and the lever it picked is the one long-horizon agents actually pay on.
Cache reads are the line item that decides whether agent loops are economical
What happened. Fable 5.1 holds the $10/$50 rate card (Opus 5 sits half that, at $5/$25) but cuts cache reads from $1.00 to $0.25 per million tokens — dropping the model’s cache-read multiplier to 0.025x base input, a quarter of the 0.1x every other Claude carries. Vendor benchmarks put 5.1 ahead of both Fable 5 and GPT-5.6 Sol across the board: 55.8 vs 42.0 vs 37.3 on Terminal-Bench 4.0, 73.4 vs 70.5 vs 67.2 on CursorBench 3.2.0, 60.9 vs 57.8 vs 56.6 on Humanity’s Last Exam without tools, and Artificial Analysis has it top of its intelligence index at 66. In the same release: Enterprise Frontier Safeguards (customer-owned storage, phased from later this fall instead of the existing zero-data-retention program), cyber-safeguard false positives down 60 percent, and vulnerability discovery now allowed — exploit development still isn’t. Everything except cache reads is unchanged, so treat the savings and benchmark claims as vendor numbers until independent re-runs land.
Why it matters. For anyone running Claude Code or a tool-heavy agent, the marginal cost of a multi-hour run is cache reads — re-scanning a stable system prompt and context prefix on every tool call. Cutting that 75 percent hits the exact line item that made Fable-class economics hostile to agentic workloads, in a column that has been arguing the premium tier repels adoption on price. The direction is the same one the open layer has been pushing from below: the frontier now competes on cost per task, not capability claims — last week OpenAI cut Sol, this week Anthropic cuts the cache bill. The operator catch: the 45 percent number is only real if your sessions actually re-read a long prefix. Short, cold, stateless traffic gets a benchmark story, not a cheaper bill.
What I’m watching
Whether the 25–45 percent agentic savings survive independent re-runs, and whether the cache-read multiplier migrates down to Opus and Sonnet — Anthropic says it is working to bring Fable 5.1’s improvements to the rest of the family. If the repriced cache rate spreads, the frontier’s variable economics quietly reset for everyone running agents on Claude, and the open-weights cost answer has to close a smaller gap.
Source: anthropic.com, the-decoder.com, platform.claude.com