Claude Fable 5.1 Launches: Cache Reads at a Quarter of the Price, Three Breaking API Changes

Anthropic launched Claude Fable 5.1 and Claude Mythos 5.1 on September 1, 2026. The headline claim — “the world’s most advanced models for coding and knowledge work” — is marketing; the parts that actually change how you build are in the release notes: a 75% cut to cache-read pricing that restructures the economics of long-running agents, three breaking API changes that will hit untested integrations, and a two-tier safeguard split that formalizes how Anthropic sells frontier capability to regulated work.

One model, two safeguard tiers

Fable 5.1 and Mythos 5.1 are, by Anthropic’s own description, the same underlying model. What differs is the safeguard stack:

Claude Fable 5.1 is the generally available tier — claude.ai, Claude Code, and the API, with model ID claude-fable-5-1 across the Claude API, Amazon Bedrock (anthropic.claude-fable-5-1), Google Cloud Vertex AI, and Microsoft Foundry.
Claude Mythos 5.1 carries heavier safeguards tuned for cybersecurity and life-sciences work, and is available only through Anthropic’s trusted-access programs. Access is currently limited to a set of US organizations, with cyberdefense access routed through the company’s CVP portal; the platform docs tie it to Project Glasswing.

This tier split is the concrete outcome of the summer’s security incidents: capability reaches general availability through the safer channel first, while the higher-risk version stays behind vetting. It also shows up in the benchmarks — on agentic coding (Terminal-Bench 4.0), Mythos 5.1 scores 60.9% against Fable 5.1’s 55.8%, a gap Anthropic attributes to cyber safeguards intervening on Fable’s tasks, a gap it expects to shrink as the new safeguards reduce false positives.

The pricing change that matters most

Fable 5.1 keeps Fable 5’s headline rates: $10 per million input tokens, $50 per million output tokens. The real change is underneath:

Cache reads drop to $0.25/MTok — a quarter of the previous rate.
Cache writes remain $12.50/MTok (5m) and $20/MTok (1h).
Batch API keeps its 50% discount on input and output.

Why this matters more than it sounds: agentic workloads re-read enormous shared context on every turn — system prompts, tool definitions, repository state, prior reasoning. Anthropic’s own measurement across four weeks of August usage puts typical workloads at roughly 25% cheaper and highly agentic, cache-heavy work at up to 45% cheaper at default effort. That’s not a rounding error; for agents with million-token contexts in play, cache reads can dominate the bill, and Cognition’s launch-day statement says exactly this: the new cache pricing makes “a Fable-class model finally economical for the workloads we’d kept on Opus, starting with code review.”

Spec-wise the model ships with a 1M-token context window, 128K max output, always-on adaptive thinking, a June 2026 knowledge cutoff, and a default effort of high in Claude Code (medium on Claude Cowork and claude.ai). Anthropic commits to retirement “not sooner than September 1, 2027.”

The breaking changes — audit your integrations

The migration guide lists three breaking changes for anyone already calling Fable 5:

Forced tool use is not supported — requests that force a specific tool now return an error instead of being silently honored.
Thinking blocks are tied to the model that produced them — earlier models in a conversation cannot read Fable 5.1’s thinking blocks.
Editing earlier turns invalidates thinking blocks — conversation surgery that used to preserve reasoning context no longer does.

None are exotic; all three live in the exact paths agent frameworks exercise most (tool orchestration, multi-model handoffs, conversation editing). If your stack pins models dynamically or rewrites history mid-thread, these deserve a test run before you switch the default model ID.

The additive changes are more interesting than the usual fare:

Per-message effort (beta) — change the effort level mid-conversation without invalidating the prompt cache, which pairs naturally with the new cache-read pricing.
Turn-scoped system messages (beta) — system-level instructions scoped to a single turn.
Progress updates between tool calls (beta) — display: "updates" returns readable intermediate updates during long tool sequences.
Content provenance — provenance metadata on outputs.
The lower cache-read price itself.

The benchmarks, with the fine print attached

Anthropic’s table (production safeguards enabled) puts Fable 5.1 ahead of its own Fable 5 and Opus 5, and ahead of OpenAI’s GPT-5.6 Sol on the listed agentic tasks:

Benchmark	Fable 5.1	Fable 5	Opus 5	GPT-5.6 Sol
Terminal-Bench-Science 0.1	52.6%	24.7%	29.0%	22.4%
Terminal-Bench 4.0	55.8%	42.0%	52.3%	37.3%
OSWorld 2.0 (partial)	77.9%	72.9%	75.4%	—
Humanity’s Last Exam (no tools)	60.9%	57.8%	56.6%	—
AutomationBench	31.4%	17.1%	26.9%	19.6%
CursorBench 3.2.0	73.4%	70.5%	70.0%	67.2%

Read this with the caveats Anthropic itself attaches: these are the vendor’s own harness results — on Terminal-Bench-Science, the public leaderboard numbers for Opus 5 and Fable 5 differ from Anthropic’s reproduction (within stated noise, but worth knowing); OSWorld 2.0 scores use the benchmark authors’ August 2026 task release and are explicitly not comparable to earlier published results; and where safeguards intervened, Fable 5.1 and Fable 5 scored zero on some OSWorld tasks, and cyber/biology tasks were completed by other Claude models. Vendor-reported, but unusually well-documented vendor-reported. The system card is public for anyone who wants the safety evaluations behind the deployment decision.

Enterprise Frontier Safeguards: privacy without a data-retention exception

Alongside the model, Anthropic detailed Enterprise Frontier Safeguards (EFS) — a system where customer data lives in cloud infrastructure controlled entirely by the customer rather than Anthropic, delivering zero-data-retention-equivalent privacy while still applying state-of-the-art misuse prevention. EFS rolls out to enterprise customers in phases “beginning later this fall”; until then, eligible customers can use Fable 5.1 with zero data retention. This is Anthropic’s answer to a real procurement blocker — regulated buyers have long had to choose between frontier capability and data governance.

The safeguard improvements also have a measurable developer-facing effect: Anthropic reports 60% fewer false positives in cybersecurity flagging, and Fable 5.1 can now be used to discover software vulnerabilities (though not to develop exploits for them) — a meaningful line for security teams.

What developers should do now

Read the migration guide before changing model IDs — the forced-tool-use and thinking-block changes are the likely breakage points.
Re-cost your agentic workloads. If your spend is cache-read-heavy, the $0.25/MTok rate changes model-selection math; re-run the comparison against Opus 5 ($5/$25) rather than assuming old ratios.
Adopt per-message effort where you were forking conversations to vary effort — it preserves the cache now that cache economics are favorable.
If you need Mythos-tier capability for cyber or bio work, access is program-gated, not self-serve — start with the CVP portal.

Limitations and unknowns: all performance figures are Anthropic-reported; independent reproductions will take days. The EFS timeline (“later this fall”) is uncommitted. Mythos 5.1 access remains US-limited for now. And the effort-level defaults differ by surface (high in Claude Code, medium elsewhere), which will surprise anyone benchmarking against the docs’ default.

Anthropic’s summer was dominated by alignment failures and security overhauls. Fable 5.1 is the first frontier release since, and the shape of it is telling: the frontier ships, but the riskiest configuration ships to vetted customers only, with the economics tuned to pull the most expensive agent workloads into the safer tier.