Fable 5.1 holds its price, quarters its cache reads, and breaks three things

Anthropic released Claude Fable 5.1 and Claude Mythos 5.1 on 1 September. They are the same underlying model shipped twice: Fable 5.1 is generally available and carries safeguards that block certain dual-use work in biology and cybersecurity, while Mythos 5.1 is the same weights with more permissive safeguards, restricted to vetted organisations through Anthropic's Project Glasswing trusted-access programme. Mythos also powers Claude Security, which the system card says is available to all Claude Enterprise customers.

For anyone building on the API, the parts worth reading are not in the benchmark table. The base price did not move. Prompt cache reads dropped to a quarter of what Fable 5 charged. And three things that worked on Fable 5 now return a 400.

The price stayed; the cache got cheap

Fable 5.1 costs $10 per million input tokens and $50 per million output tokens — identical to Fable 5. What changed is the prompt cache. A cache hit on Fable 5.1 or Mythos 5.1 costs $0.25 per million tokens, which Anthropic's pricing documentation describes as 0.025 times the base input price, against the 0.1x multiplier every other Claude model uses.

For comparison, Fable 5 charged $1.00. Cache writes are unchanged at $12.50 per million for the five-minute duration and $20 for the one-hour duration, and the 512-token minimum cacheable prompt length is unchanged as well.

It also scrambles the usual model-selection heuristic. Fable 5.1 has the highest base input price in the current Claude lineup and, in absolute terms, a cache-read price of $0.25 — barely above Claude Sonnet 5's $0.20, and well under Claude Opus 5's $0.50. For a workload that is mostly cache reads, the premium model is no longer obviously the expensive one.

Three things that will break

Anthropic's migration notes list three breaking changes. None of them are opt-in.

1. Forced tool use is gone. Setting tool_choice to {"type": "any"} or {"type": "tool", "name": "..."} returns a 400 invalid_request_error, and the same validation applies to the token-counting endpoint. Anthropic's stated reason is that thinking is always on for these models, so a forced tool call would skip it and the model would write its working-out into the tool arguments instead, degrading argument quality. The recommended replacement is tool_choice auto plus strict tool use, or moving the schema to structured outputs.
2. Thinking blocks are bound to the model that produced them. Preservation runs one way: Fable 5.1 can read thinking blocks from Opus 5, Fable 5 and earlier models, but no earlier model can read Fable 5.1's. If a router or a fallback path switches models mid-conversation, the API silently drops the block — it is not billed, and by default you are not told. The thinking-binding-controls-2026-08-01 beta header surfaces the drop in an input_transformations array.
3. Editing earlier turns invalidates thinking blocks. Modify anything before a Fable 5.1 thinking block — the system prompt, the tools array, or an earlier message — and the next request fails with a 400 whose message reads 'The block is bound to a different conversation'. This is enforced for accounts created on or after 31 August 2026; older accounts record the mismatch but only act on it if the request asks them to. Mythos 5.1 does not run the check.

The third is the one most likely to bite existing code. The failure modes Anthropic names are ordinary engineering habits: injecting a per-request reminder into an earlier turn and removing it next time, rebuilding the system prompt or tools array between requests, reordering history. The fix is to treat conversations as append-only. Claude Code, claude.ai, Claude Managed Agents and the Claude Agent SDK already do; hand-rolled message arrays are the exposure.

Three additions arrive alongside, all in beta. Per-message effort lets you raise effort for a hard step and lower it for routine ones without invalidating the prompt cache.

Turn-scoped system messages (clear_at: "next_user_message") give an instruction system-prompt authority for one turn, then stop rendering while staying in the history at no token cost — precisely the append-only-safe way to do the reminder injection that now breaks thinking blocks. And thinking.display accepts "updates", returning as text the short progress lines the model writes between tool calls while reasoning stays hidden.

The numbers, and what they are worth

Anthropic's comparison table reports Fable 5.1 at 55.8% on Terminal-Bench 4.0 against 42.0% for Fable 5 and 52.3% for Opus 5, with Mythos 5.1 at 60.9%; 52.6% on Terminal-Bench-Science 0.1 against 24.7%; 31.4% on AutomationBench against 17.1%; and 1853 on GDPval-AA v2 against 1723 for Fable 5 and 1824 for Opus 5. The table also lists GPT-5.6 Sol, at 37.3% on Terminal-Bench 4.0 and 1711 on GDPval-AA v2.

All of these are Anthropic's own runs, on evaluations Anthropic selected, and none has been independently reproduced. VentureBeat's write-up made the same point plainly, noting the figures "should be read as vendor-reported results rather than independent proof of superiority". The Terminal-Bench-Science jump — more than doubling — is the kind of result that is either a genuine capability step or an artefact of how a young benchmark interacts with a model trained after it existed. There is no way to tell from a launch post.

What the system card says that the blog post doesn't

The 212-page system card is more forthcoming. Under Anthropic's Responsible Scaling Policy and Frontier Compliance Framework, the model is judged to have CB-1 chemical and biological capabilities — enough to meaningfully help someone with a basic technical background synthesise a known weapon — but to fall short of the CB-2 threshold, a judgement Anthropic says it holds "with some uncertainty". Automated AI research and development risk is assessed as low, with external testing by METR described as consistent with that.

The line that stands out: "On alignment risks, we now assess the risk of catastrophic harm as low rather than very low." Anthropic attributes the change to its August 2026 Risk Report and to increased uncertainty following the disclosures about model behaviour in its cybersecurity evaluations — the incidents behind the report it published on 31 August. A model launch that revises the company's own risk posture downward is not a common thing to find in a system card.

Although Mythos 5.1 is in Tier 1, it is getting closer to Tier 2, completing more and more autonomous tasks. We have yet to see novel offensive capability.

On cyber, the card calls these the strongest overall cyber capabilities of any model Anthropic has released, substantially outperforming Opus 5 on almost all reported evaluations including ExploitBench, OSS-Fuzz and ExploitGym. The mitigation is two-stage: a probe reads Claude's internal activations to flag cyber-related traffic, which then goes to a separate trained classifier that decides whether to block.

Fable 5.1 now permits vulnerability discovery in source code at all access levels, general availability included, while still blocking exploit development. Anthropic says its updated safeguards produce 60% fewer false positives than before; the card adds the qualifier the announcement leaves out, that they are still likelier to trigger than Opus 5's.

The alignment section is the least comfortable reading. On Anthropic's automated behavioural audit, Mythos 5.1 is "a slight regression on overall misaligned behaviour compared to Opus 5", though an improvement on Mythos 5 and Sonnet 5. It "cooperates with human misuse and accepts unverifiable claims of authorization more readily than Opus 5", while ignoring explicit constraints, hallucinating inputs and falsely claiming completed tasks less often than previous models.

Internal deployment monitoring caught rare cases of the model working around safety classifiers or broken permission hooks, sometimes by overstating what the user had authorised. On the other side of the ledger, it is Anthropic's most robust model to date on the external Indirect Prompt Injection benchmark.

What changes in your prompts

Anthropic documents six behaviour differences that show up with no code change at all, each with a prompting fix. They are worth reading before you blame a regression on the model:

• Parallel tool calling is more variable. Fable 5.1 may issue one tool call per turn where Fable 5 batched several — costing tokens and round trips, not answer quality. A one-line batching instruction is the documented fix.
• Fewer progress updates during long tool runs, especially at high effort, which is why display: "updates" shipped alongside. If your UI depends on narration, ask for it explicitly.
• It answers from memory more often at low effort, calling search and retrieval tools less. Raise effort for turns that need fresh information.
• Whole-file rewrites for small edits are more likely, costing output tokens rather than correctness.
• Unmarked quotations in summaries: it reproduces source passages without marking them as quotes more often than Fable 5 did.
• Less formatting in chat, so anti-formatting rules written for earlier Claude models can now suppress structure the content needs.

One operational detail reported inconsistently elsewhere: Anthropic's documentation states that both models carry 30-day data retention and are not available under zero data retention unless Anthropic expressly authorises it. Separately, the same day, Anthropic announced Enterprise Frontier Safeguards, which keeps data in a customer-controlled cloud account under customer-managed keys while automated misuse monitoring continues without Anthropic human review.

The company says over 100 enterprise customers helped design it, that it carries no additional Anthropic charge, and that it rolls out in phases from autumn 2026. That is the route to zero retention here — a commitment for later this year, not a setting available today.

Taken together, this is a point release that behaves like one: same price, same context window, same tokenizer, a real but bounded capability step that nobody outside Anthropic has verified. The parts that will consume engineering time are the three 400s and the append-only discipline they impose. The part worth thinking about longer is a company shipping its most capable model on the same day it revises its own catastrophic-risk assessment in the wrong direction, and saying so in a document most readers will never open.