Claude Fable 5.1: The Cache Cut That Changes Agent Costs

Anthropic shipped Claude Fable 5.1 on September 1, cutting cache read prices by 75% from $1.00 to $0.25 per million tokens, which reduces the real cost of highly agentic workloads by up to 45%. The model also more than doubled its Terminal-Bench-Science 0.1 score from 24.7% to 52.6%, but introduces three breaking API changes, including the removal of forced tool use and restrictions on reading thinking blocks from earlier models.

Anthropic shipped Claude Fable 5.1 on September 1 with identical sticker prices to Fable 5 — $10 per million input tokens, $50 per million output. The number that moved was cache reads: down 75%, from $1.00 to $0.25 per million tokens. For anyone running agents that repeatedly re-read large system prompts or conversation context across long sessions, that single change cuts the real cost of production by up to 45%. The benchmarks improved too, but the economics story is the one worth understanding first. Why Cache Reads Are the Whole Ballgame Every other current Claude model — Opus 5 included — charges cache reads at 10% of the base input price. Fable 5.1 charges 2.5%. That gap is not subtle. A 200,000-token cached prefix read 100 times used to cost $20 on Fable 5. On Fable 5.1 it costs $5. For short one-shot queries the savings are minimal. For an agent that maintains a large static system prompt, a growing conversation history, or a codebase context window re-read on every turn, the savings compound fast. | Component | Fable 5.1 | Fable 5 | Change | |---|---|---|---| | Input | $10 / MTok | $10 / MTok | — | | Output | $50 / MTok | $50 / MTok | — | | Cache reads | $0.25 / MTok | $1.00 / MTok | -75% | | Agentic workload | ~45% cheaper | baseline | — | Anthropic estimates typical workloads save around 25%. Highly agentic workloads — the ones with long sessions, many tool calls, and repeated context — save up to 45%. That puts Fable 5.1 in a counterintuitive position: despite base input prices five times higher than Opus 5, its cache read rate $0.25 is half of Opus 5’s $0.50 . The practical takeaway is that routing expensive tasks to “cheaper” models now requires checking whether those models actually are cheaper once cache is factored in. The full pricing breakdown is in Anthropic’s official release notes. https://platform.claude.com/docs/en/models/fable-5-1/whats-new-fable-5-1 Three Breaking Changes to Audit Before Migrating The model ID change is one line: claude-fable-5 → claude-fable-5-1 . But three API changes will break existing integrations, and they are not equally loud. Forced tool use is gone. Setting tool choice to {"type": "any"} or a named tool now returns a 400 error. Adaptive thinking is always on for Fable 5.1, and forcing a tool call skips the reasoning step, degrading argument quality. The fix is to switch to tool choice: {"type": "auto"} with strict: true on the tool schema, or move to structured outputs. This surfaces in minutes of testing — it’s a loud failure. Thinking blocks from Fable 5.1 cannot be read by earlier models. The reverse works: Fable 5.1 reads blocks from older models. But if a fallback router drops from Fable 5.1 to Opus 5 mid-session, thinking blocks are silently discarded. The model re-plans without that reasoning, increasing cost and latency. Add the thinking-binding-controls-2026-08-01 beta header to surface dropped blocks in the input transformations field. Editing earlier conversation turns invalidates thinking blocks. This is the quiet one. Modifying a previous system prompt, injecting a reminder into an earlier message, or rebuilding the tools array between requests breaks all subsequent thinking blocks. The enforcement is immediate for accounts created August 31 or later. Older accounts are affected but the API only records the mismatch unless you explicitly set prefix mismatch behavior . Teams that pass migration today can hit this in the next sprint if they touch history-editing code. Forced tool use returning a 400 is loud and will be found in minutes. The thinking-block rules are quiet: a fallback router loses reasoning without saying so, and the prefix check only bites accounts created from the end of August. The Agentic Benchmark Gains Are Real Fable 5.1’s score on Terminal-Bench-Science 0.1 went from 24.7% to 52.6% — more than doubled. AutomationBench jumped from 17.1% to 31.4%. Apply standard skepticism to Anthropic’s own benchmarks, but early-access partners reported agents completing 38-hour unattended machine-learning runs and 82% task completion on browser-agent benchmarks. The model also generates around 60% fewer safety interventions per Claude Code session compared to Fable 5 — which for anyone who has had an agent grind to a halt mid-task matters more than any synthetic benchmark. New Features Worth Enabling Three new beta features change how agents are structured. Per-message effort lets you change reasoning depth mid-conversation without busting the prompt cache — raise it for hard planning steps, lower it for routine retrieval. Turn-scoped system messages let you send per-turn reminders that expire after one turn without touching conversation history or invalidating thinking blocks. Progress updates between tool calls with display: "updates" surfaces the model’s between-tool reasoning as readable text, solving the problem of long agentic turns looking completely silent to users. How to Migrate This Week Start with the model ID swap, then run three audits: remove forced tool use, verify fallback routers handle thinking block drops, and confirm your code treats conversation history as append-only. Anthropic’s migration guide https://platform.claude.com/docs/en/models/fable-5-1/migration-guide and its companion prompting guide cover each behavioral regression with specific fixes — a level of candor about model trade-offs that most release notes skip. One shift worth internalizing: Fable 5.1 at low effort is now often competitive with Claude Opus and Sonnet models on cost per task while scoring higher on benchmarks. That inverts prior routing logic. Before Fable 5.1, the sensible path was routing cheap tasks to cheaper models. Now: start at Fable 5.1 low effort, escalate only where evals demand it. VentureBeat’s coverage https://venturebeat.com/technology/anthropics-claude-fable-5-1-and-mythos-5-1-arrive-with-a-75-cost-reduction-for-fable-cache-reads has more on how early adopters are restructuring their model routing. The 75% cache cut is not generosity — it is Anthropic making more agentic runs economically rational. More agentic runs means more API calls, more data, and more Anthropic. The developers who internalize the new math first will route their infrastructure accordingly before the rest of the market catches up. See the full developer breakdown including all five new API features at ClaudeFast. https://claudefa.st/blog/models/claude-fable-5-1