Cache reads get 75% cheaper on Claude Fable 5.1

Anthropic on September 1 put Claude Fable 5.1 on the Claude API, AWS, Google Cloud and Azure as claude-fable-5-1. The list price stays at Fable 5’s $10 per million input tokens and $50 per million output tokens, and only cache reads change: they now cost $0.25 per million tokens, 75% less than before. Alongside it comes Claude Mythos 5.1, which Anthropic’s announcement describes as the same model with looser safeguards, available only to vetted organizations.

In the Claude Platform docs, Fable 5 is now labeled Legacy and the recommendation is to migrate. The knowledge cutoff moves from January 2026 to June 2026.

Where the savings come from

A cache read is what you pay when the model re-reads context it has already processed. Long agent sessions generate a lot of them. Before this release they cost $1 per million tokens. Now they cost a fortieth of the input price. Anthropic says typical workloads will cost about 25% less than on Fable 5, and context-heavy, tool-heavy work, where cache reads make up most of the bill, up to about 45% less. Both figures come from four weeks of real usage in August 2026 at default effort, covering Claude Enterprise, Claude Code and the API, and they apply wherever usage is billed by token. US-only inference is priced at 1.1x for input and output tokens, according to Anthropic’s Fable page.

Subscribers may not see much of this. Testing Catalog reported that Fable 5.1 on Claude and Claude Code requires extra usage credits. On the API, the cheaper cache also changes when you should compact. Anthropic’s prompting guide says compacting early to save money “may no longer be the right cost-intelligence tradeoff” on Fable 5.1 and suggests trying later compaction points.

Benchmarks with the filters on

The largest jump is on Terminal-Bench-Science 0.1. Fable 5.1 scores 52.6%, against 24.7% for Fable 5, 29.0% for Opus 5 and 22.4% for GPT-5.6 Sol. Anthropic gives a standard error of ±3.5-4.5 points per model. The public leaderboard lists Opus 5 at 30.0% and Fable 5 at 21.4%, and Anthropic’s own runs reproduced both within noise. On Terminal-Bench 4.0, Fable 5.1 scores 55.8%, against 42.0% for Fable 5, 52.3% for Opus 5 and 37.3% for GPT-5.6 Sol. Mythos 5.1, the same weights with looser safeguards, scores 60.9% on that benchmark.

That gap comes from how the tests were run. Fable 5.1 was evaluated with its production safeguards on. Where they stepped in on OSWorld 2.0, the task scored zero. In other interventions, cybersecurity tasks were finished by Claude Opus 4.8 and biology tasks by Claude Opus 5. Anthropic says this likely lowers the Fable scores. The other results are mixed. GDPval-AA v2 rises from 1723 to 1853 and AutomationBench from 17.1% to 31.4%, but CursorBench 3.2.0 goes only from 70.5% to 73.4%. The OSWorld 2.0 numbers use the benchmark authors’ August 2026 task release, so they can’t be compared with earlier published results.

Effort is now the main lever. At Medium, Anthropic says Fable 5.1 matches or beats Fable 5 at much lower cost. The prompting guide adds that at Low it often competes with Opus and Sonnet models on cost per task while scoring higher. The default is High in Claude Code and Medium in Claude Cowork and on Claude.ai. Effort names don’t map to the same amount of thinking across models, so the guide asks teams to re-run their effort sweep even if they already did one on Fable 5. At Low, the model also calls search and retrieval tools less often. One concrete case from the launch: in testing by the investment firm Millennium, Fable 5.1 found the cause of a rare crash in its internal systems that its engineers had failed to explain after several years.

The research results come mostly from Mythos 5.1. Its protein binder designs had a hit rate of nearly 50% across 12 targets, where 10-15% is typical. Custom GPU kernels it wrote made seven open-source genomics and protein models 1.4-2.5x faster on an NVIDIA H100 with identical outputs. Anthropic estimates that cuts GPU costs by 30-60% on large analyses and says it will open-source the optimizations.

Fewer interventions, same lines

On Fable 5, a tripped classifier meant a refusal that fell back to Opus 4.8. Anthropic said that happened in under 5% of sessions, and on the raw API the fallback was opt-in. Fable 5.1 can still return stop_reason: "refusal", but the cyber filter is narrower. The model may now be used to find software vulnerabilities, though not to write exploits for them, and Anthropic expects Claude Code users to see about 60% fewer cyber interventions per session than under Fable 5’s previous safeguards. Penetration testing, exploit generation and binary-based vulnerability scanning are still sent to Opus models. On the biology side, the update announced on August 6 makes the safeguards fire 85% less often on benign questions about elementary biology and medicine. Life-sciences research and development queries still go to Opus.

Enterprise Frontier Safeguards (EFS) addresses the data-retention complaint. Customer data stays in cloud infrastructure the customer controls, and by default the customer does any human review, while Anthropic keeps its ability to detect misuse. Anthropic says EFS offers the same privacy as a zero data retention agreement. It was built with more than 100 customers and will be supported on Claude Code, Claude Enterprise, the Claude Platform, Amazon Bedrock, Claude Platform on AWS, Google’s Agent Platform and Microsoft Foundry. The phased rollout begins later this fall, and until then eligible customers can run Fable 5.1 and Fable 5 with zero data retention.

Mythos 5.1 is offered through two programs. The Cyber Verification Program will add Mythos-class models “in the near future”. The Life Sciences Verification Program, run in partnership with the US government, has enrolled its first participants. For now, access is limited to a set of US organizations. Claude Security, which switched to Mythos 5 in late August, has already moved to Mythos 5.1. Anthropic’s own testing also found that the model can still sometimes get around approvals and auto-mode classifiers.

What to change in the harness

The change most likely to break existing code is a new anti-distillation rule for thinking blocks. The announcement applies it to API accounts created “from today onwards”. The prompting guide gives the cutoff as August 31, 2026. For those accounts, a thinking block is valid only in the exact conversation that produced it. If you replay it after the system prompt, the tool list or any earlier message has changed, the request returns a 400. The alternative, behind the thinking-binding-controls-2026-08-01 beta header, is prefix_mismatch_behavior: "drop_block", which drops the affected blocks instead. Existing accounts are exempt for now, but future models will enforce the check for everyone. The edits that trigger it are the same ones that restart the prompt cache: adding and removing per-turn reminders, summarizing older turns in place, and rewriting the system prompt mid-session. The guide recommends turn-scoped system messages with clear_at: "next_user_message" (beta header mid-conversation-system-clear-at-2026-08-21) and leaving trimming to server-side compaction.

Fable 5.1 also posts fewer updates during long tool-calling turns, so users may see the agent go quiet for minutes. Under the default thinking.display of "omitted", the progress notes arrive as empty blocks. Setting "updates" with the thinking-display-updates-2026-08-18 header returns them. Craig Falls, Head of Quantitative Research at Jane Street, saw the upside of the quieter style: “While prior models became hard to follow the longer they worked, Fable 5.1 remains readable over long, multi-step tasks.” In coding and computer-use loops, the model may issue independent tool calls one per turn. A one-sentence nudge, appended as a turn-scoped message, brings back parallel calls.

Fable 5.1 is also released after August 2, 2026, so under the EU AI Act Code of Practice its output carries an invisible watermark. A detection API is in private preview for regulators, media, researchers and enterprises with compliance obligations.

To check whether your own integration is affected before enforcement reaches every account, the guide suggests running a session with prefix_mismatch_behavior: "drop_block" and logging input_transformations, or capturing a few normal turns and confirming that consecutive requests are byte-identical up to the newly appended turn.