Claude Fable 5.1: A Cache Discount, Three Breaking Changes, and the Same Data Terms
Fable 5.1 shipped on 1 September at Fable 5's list price, with cache reads cut from $1.00 to $0.25 per million tokens. The saving exists only for cache-heavy agent traffic, three API changes break working code, and the 30-day retention requirement has not moved. Why Opus 5 stays the ceiling for most EU routes and Fable 5.1 is an exception you justify in writing.
This week Anthropic released Claude Fable 5.1 and Claude Mythos 5.1. They arrived on Tuesday 1 September. Fable 5.1 keeps Fable 5's list price, $10 in and $50 out per million tokens, and changes one line: a cache read drops from $1.00 to $0.25. VentureBeat's headline called it a 75% cost reduction, which is true of that line and of nothing else on the invoice. Two days later OpenAI shipped GPT-6 Astra, and the comparison table Anthropic had just published was about the previous model.
Five days in, what can be said with confidence is who the discount is for, what the release breaks, and why the data terms still decide whether an EU team can use it at all.
A discount on one line of the invoice
The model is claude-fable-5-1. Input is $10 per million tokens and output $50, "the same as Claude Fable 5". Cache writes are unchanged at $12.50 for five minutes and $20 for an hour. The Batch API still halves input and output. The context window is 1M tokens at the standard rate across the whole window, with 128K tokens of output.
The change is in one footnote of the pricing page: "Cache hits and refreshes on Claude Fable 5.1 and Claude Mythos 5.1 are priced at 0.025x the base input price. All other models use the standard 0.1x multiplier." That is 2.5% of input where every other Claude model pays 10%.
Anthropic's savings claim is that "typical workloads, costs are reduced by around 25% relative to Fable 5", and that for "complex coding and highly agentic tasks, the savings could be up to around 45%". The footnote says how it was measured: "Indexed cost of running the same workloads on Fable 5 and Fable 5.1, at usage-based pricing measured at default effort over four weeks of actual usage in August 2026." So it is Anthropic's own traffic, at the default effort level, over one month, and nobody outside has replicated it. It may well be right. It is not your number.
Your number depends on one ratio: how much of your input is cache reads. The docs say it plainly: "Long agentic sessions that re-read a cached prefix pay a quarter of the Claude Fable 5 rate." Here is what that means on list prices, and this arithmetic is ours. Take one agent turn that re-reads a 900K-token cached prefix and adds 10K fresh tokens. On Fable 5 the input side costs $0.90 plus $0.10. On Fable 5.1 it costs $0.225 plus $0.10. Now take a cold 100K-token prompt with no cache at all: $1.00 on either model.
That example leaves out the one-off cache write and all output tokens, both unchanged. The direction is what matters. A long agent session with a stable prefix gets most of the cut. A document-in, answer-out workload with a fresh prompt every time gets none of it.
There is a commercial reason for the shape of this discount. VentureBeat, citing the Financial Times, reported that Fable 5 accounted for only about 11% of Anthropic spending among 70,000 companies, and cited The Information on "growing concern among enterprise customers about unpredictable AI bills". We have those two claims second-hand.
Three changes that break working code
The docs list three breaking changes for anyone already calling Fable 5. Each one is a check a team can run this week.
- Forced tool use is rejected. A tool_choice of {"type": "any"} or {"type": "tool", "name": "..."} now returns a 400: "tool_choice: type "tool" and "any" are not supported for this model." Anthropic's reason is that thinking is always on and a forced call would skip it. Check: grep your code for tool_choice. Move schema enforcement to strict tool use with auto, or to structured outputs, and say in the prompt when the tool applies.
- Editing earlier turns invalidates thinking blocks. Change anything in front of a Fable 5.1 thinking block, whether the system prompt, the tools array or an earlier message, and the next request fails with "The block is bound to a different conversation". The check is "enforced for new accounts created on or after August 31, 2026"; older accounts opt in. The patterns that trip it are common: injecting a per-request reminder and deleting it next turn, rebuilding system or tools between requests, trimming history client-side. Check: run a session with prefix_mismatch_behavior: "drop_block" under the thinking-binding-controls-2026-08-01 beta header and log input_transformations. If anything shows up, your harness edits history. The fix is to treat the conversation as append-only and use turn-scoped system messages, mid-conversation tool changes and server-side compaction.
- Thinking blocks are bound to the model that produced them. "Claude Fable 5.1 reads earlier models' thinking blocks, and no earlier model reads Claude Fable 5.1's." When a router or a fallback moves a conversation from Fable 5.1 to Opus 5 or Opus 4.8, "the API drops the block before the model sees it", and without the beta header "the drop is silent". Check: if your routing layer can switch models mid-conversation, log those switches and score them in your evals as a reasoning reset.
Read the second and third items next to § 01 and the release has a single design behind it. The price cut rewards a cached, stable prefix. The thinking-block rules punish anything that disturbs one. Anthropic is pricing and enforcing the same pattern: long, append-only agent sessions on one model.
The list price did not move. Anthropic cut the one line that only a long, cached, append-only agent session ever sees, and then made append-only a rule.
Some behaviour also shifts with no code change. The docs note more single tool calls per turn where Fable 5 batched several, fewer progress updates, more whole-file rewrites for small edits, and a model "more likely to reproduce passages of the source without marking them as quotations" when summarising. Multilingual performance is "on par with Claude Fable 5", so expect no gain in Slovak. And every text output now "carries Anthropic's statistical text watermark on every platform", which belongs in your transparency documentation.
A quieter filter, the same data terms
The classifier improved, and Anthropic put numbers on it. "Claude Code users can expect an average of around 60% fewer interventions per session", and the biology safeguards trigger "85% less often for benign requests related to elementary biology and medical questions". The policy moved too: like Opus 5, Fable 5.1 now allows vulnerability discovery in source code at every access level.
The system card is careful about the limits. Fable 5.1 produces fewer false positives "than Fable 5 did at launch, though they are still likelier to trigger than Opus 5's safeguards", and the classifiers "will continue to block some benign or borderline uses out of an abundance of caution". For security work the card is blunter still: because the classifiers fire across all tested cyber evaluations, "Fable 5.1's performance on cyber tasks is nearly identical to that of Opus 4.8", the model that flagged requests fall back to. On the API the permitted fallback targets are Opus 4.8 and Opus 5, a refusal before any output is not billed, and a fallback credit refunds the prompt-cache cost of switching.
For an EU buyer the harder constraints are these:
- Retention"Claude Fable 5.1 and Claude Mythos 5.1 carry 30-day data retention and aren't available under zero data retention unless expressly authorized by Anthropic." The requirement "applies wherever Covered Models are offered"; on Bedrock and Google Cloud the retained data stays in your cloud provider's environment.
- Enterprise Frontier SafeguardsThe promised way out: customers "store their data on their own cloud infrastructure, rather than on Anthropic's systems", with monitoring that keeps "the privacy of a zero data retention agreement". It is "rolling out in phases, starting this fall". No dates, no region list, no eligibility criteria and no EU terms have been published.
- RegionsBedrock docs: "For Claude Fable 5.1, regional endpoints are currently available in us-east-1 only." On Microsoft Foundry it is hosted on Anthropic's infrastructure only, Global Standard only. Google Cloud offers the global endpoint or the eu multi-region at a 10% premium. The first-party API has no EU inference geo.
- Mythos 5.1"Currently, it is only available to a set of US organizations". Anthropic says it is "coordinating with the US government to expand access". On 3 September the Commission's spokesperson Thomas Regnier answered: "Europe is not a security risk; we're an economic opportunity".
We could not confirm whether an EU cross-region inference profile exists for Fable 5.1 on Bedrock. The docs sentence above is all we have, and it points the other way.
Our read
For most EU production routes, Opus 5 remains the ceiling you can actually deploy under your data terms. It runs under zero data retention, has an Azure-hosted option and EU profiles on Bedrock, and its classifier leaves technical work alone. Anthropic's own docs agree on the ordering: start with Opus 5, and reach for Fable 5.1 "when your evals on Claude Opus 5 at higher effort still fall short".
That makes Fable 5.1 a per-route exception, and an exception should be justified in writing. We would want four lines on the page before a route names claude-fable-5-1. The route carries no personal data, or the DPO has accepted 30-day retention for it. The workload is cache-heavy and the harness is append-only, because otherwise you pay Fable 5 prices. Your own eval shows Opus 5 at xhigh failing the task. And the log records every model switch, because a fallback now costs the conversation its reasoning as well as its model.
The honest limit on this advice is that it is two weeks old and rests on vendor numbers for the capability gap. If Enterprise Frontier Safeguards arrives with EU regions and customer-held monitoring data, the retention objection weakens and the calculation changes. Until Anthropic publishes dates, plan as if it has not shipped, because it has not.
Sebrona writes routing policies where an exception like this is a reviewed entry with a reason, an eval result and an owner, inside an EU data boundary. If you are deciding whether Fable 5.1 belongs on any of your routes, write to info@sebrona.com.
Reading
Where the prices, API changes, scores and quotes come from. The two worked examples in § 01 are our arithmetic on Anthropic's list prices.
- Fable 5.1 and Mythos 5.1 announcementThe cache-read cut, the 25% and 45% claims and their measurement footnote, the 60% and 85% intervention figures, Enterprise Frontier Safeguards, Mythos 5.1 access, and the benchmark table.
- Fable 5.1 and Mythos 5.1 system cardThe false-positive comparison with Fable 5 and Opus 5, the cyber-task sentence, the Opus 4.8 fallback, Table 8.1.A, Cursor's and Artificial Analysis's independent rows, and the CursorBench cost per task.
- Anthropic pricing page and Fable 5.1 docsEvery price in Fig. 01, the 0.025x multiplier footnote, batch rates, model ID, context and output limits, the three breaking changes with their error strings and beta headers, behaviour differences, the watermark, fallback targets and billing, and the retention requirement.
- Platform docsBedrock regional endpoints for Fable 5.1, Foundry hosting options, Google Cloud endpoint types and premium, first-party inference_geo values, and the models overview's "start with Claude Opus 5" guidance.
- VentureBeat, Carl FranzenThe headline framing, the vendor-reported caveat, the detail on customer-held monitoring data, and the relayed Financial Times and The Information claims, which we did not read at source.
- EuropeRegnier on Mythos 5.1 (EU Today, 3 Sep).
- OpenAI, GPT-6 Astra announcementThe 3 and 4 September rollout dates and the list price. Astra itself is outside the scope of this post.
- Not verifiedAny date, region or eligibility rule for Enterprise Frontier Safeguards; an EU inference profile for Fable 5.1 on Bedrock; any third-party replication of the savings claim.