Anthropic released Claude Fable 5.1 and Claude Mythos 5.1 on Tuesday, September 1, 2026, and the headline is cost, not capability. Base prices stay at $10 per million input tokens and $50 per million output, but cache reads drop 75%, from $1.00 to $0.25 per million. Anthropic estimates typical workloads get about 25% cheaper and highly agentic ones save up to roughly 45%. The quieter headline sits in the developer docs: Fable 5.1's reasoning is now bound to the conversation and model that produced it, and no other model can read it.
The release also cuts false refusals sharply. Anthropic reports about 60% fewer cybersecurity false positives per Claude Code session, and biology safeguards fire 85% less often on benign requests. Fable 5.1 did not replace Fable 5: both stay selectable, and Anthropic still tells developers to start with Opus 5 and reach for Fable only when Opus at higher effort falls short. Here is what actually changed, what it costs on each plan, and the fine print that decides whether the upgrade matters for you.
What Anthropic shipped on September 1
Fable 5.1 and Mythos 5.1 are the same underlying model with different safeguard levels, shipped as a pair roughly three months after Fable 5. Fable sits above Opus, Sonnet, and Haiku as Anthropic's tier for long-horizon agentic work: multi-hour coding sessions, deep research runs, and document, spreadsheet, and slide production. Mythos 5.1 loosens cybersecurity and life-science restrictions and stays invite-only through Project Glasswing, Anthropic's trusted-access program for vetted cyberdefense and life-science organizations, currently limited to US organizations. The specs, straight from Anthropic's model documentation:
- Context and output: a 1 million token context window (both default and maximum, with flat per-token pricing across the whole window) and 128K max output tokens.
- Thinking: adaptive thinking is always on, with effort defaulting to high. Forced tool use now returns an error instead of skipping the thinking step.
- Inputs and knowledge: text and images in, text out, with a June 2026 knowledge cutoff, the freshest in the current Claude lineup.
- Model IDs: claude-fable-5-1 on the Claude API, Google Cloud, and Microsoft Foundry, and anthropic.claude-fable-5-1 on Amazon Bedrock.
- Lifecycle: Anthropic commits to keeping Fable 5.1 available until at least September 1, 2027.
The pricing change that actually matters
The sticker price did not move. Input stays at $10 per million tokens, output at $50, and the Batch API still takes 50% off both. Cache writes are unchanged too: $12.50 per million for five-minute entries, $20 for one-hour entries. The entire cut lands on cache reads, now $0.25 per million tokens, which works out to 2.5% of the base input price. Every other Claude model charges 10%.
That one lever matters because cache reads dominate agent bills. A long coding or research session re-reads its own growing history on every turn: the system prompt, the tool definitions, every earlier exchange. On Fable 5, each of those re-reads billed at $1.00 per million tokens, so a session looping hundreds of times paid more for its own past than for anything new. Cutting that line item by 75% is why Anthropic's estimate reaches 45% savings for highly agentic workloads while a typical workload sees closer to 25%. How caching works, and how the three big providers price it differently, is covered in the prompt caching explainer.
Even after the cut, Fable 5.1 is premium-priced. OpenAI's GPT-5.6 Sol lists $4 per million input tokens and $20 output under a promotional cut running through November 21, 2026, though it doubles the input rate on prompts past 272K tokens. Google's Gemini 3.7 Flash lists $0.75 and $3.75 on introductory pricing through the end of the year. Anthropic is not competing on uncached rates and does not pretend to. The pitch is that the workloads Fable exists for, long-running agents, push most of their volume through cache reads, and that is exactly the line Anthropic chose to cut.
What the vendor-reported benchmarks say
Anthropic's own numbers show the biggest jumps on scientific computing and long-horizon automation. Every figure below is vendor-reported, from Anthropic's own launch materials, not an independent evaluation.
- Terminal-Bench-Science: 52.6%, up from Fable 5's 24.7%, more than doubling the score on agentic scientific tasks.
- AutomationBench: 31.4% versus 17.1% for Fable 5 on long-horizon automation work.
- GDPval-AA v2: a rating of 1853 against Fable 5's 1723 on economically valuable knowledge work.
- Humanity's Last Exam: 65.0% with tools.
- CursorBench: 73.4% on real-world coding tasks.
Anthropic flags two caveats itself, and they deserve repeating. On runs where safeguards intervened, the affected model scored zero: both models on OSWorld 2.0, and Fable 5 on AutomationBench, so part of some gaps reflects fewer safeguard trips rather than raw capability. And its OSWorld 2.0 numbers use the benchmark's August 2026 task release: Anthropic re-ran Fable 5 and Opus 5 under the same conditions, and says the results are not directly comparable to previously published OSWorld scores.
The launch stories lean the same direction, toward endurance rather than single clever answers. Ramp described an unattended 38-hour machine learning run that diagnosed a prior result as a label artifact, kicked off six parallel experiments, and came back with findings and proposed next steps. Millennium says Fable 5.1 was the first to find the root cause of a crash that had gone unexplained for four to five years, disassembling an external vendor library and matching it against the core dump to get there. Browserbase reports 82% completion on its hardest browser-agent benchmark at about ten minutes per task, against 74% for Opus 5 and 57% for Fable 5.
Fewer false refusals, and a cheaper path when one fires
The refusal changes target the complaint that safety filters block legitimate work. Anthropic measures the roughly 60% drop in cybersecurity false positives per Claude Code session, and says biology safeguards now fire 85% less often on benign requests such as elementary biology and medical questions. When Fable 5.1 does refuse, the mechanics carry over from Fable 5: the API returns a refusal stop reason and can retry the request on a fallback model, now Opus 4.8 or Opus 5. Fallback credit refunds the prompt-cache cost of switching models, so a mid-session refusal does not add a cache rebuild to the bill. The full tier and fallback system is mapped in the Fable model tiers explainer.
Looser safeguards cut both ways. TechCrunch, quoting the system card, reports that Mythos 5.1, the same model under lighter safeguards, is a "slight regression on overall misaligned behavior compared to Opus 5" and "accepts unverifiable claims of authorization somewhat more readily," meaning a user asserting permissions they cannot prove gets further than they would elsewhere. The same assessment finds it less likely to ignore explicit constraints or falsely claim completed work than previous models. Provenance is the counterweight: every text output from Fable 5.1 and Mythos 5.1 carries Anthropic's statistical watermark, and supported image, video, and audio files the models produce carry C2PA Content Credentials, the system already covered in the Claude watermark explainer.
What each Claude plan actually gets
The consumer gating is specific enough to misread. Max subscribers and premium seats on Team and Enterprise plans have Fable models in the box: up to 50% of the weekly usage limit can go to Fable 5 or 5.1 at no extra cost, with pay-as-you-go rates after that. Pro and standard Team seats sit outside that bundle. Fable usage there does not draw on plan limits at all; it runs on purchased pay-as-you-go usage credits from the first token. The distinction stings a little more this time because Fable 5's move to paid access in July 2026 came with a one-time promotional credit, and Anthropic's support documentation states plainly that no equivalent credit exists for Fable 5.1.
Plan check in one line: Max and premium Team or Enterprise seats get Fable 5.1 inside weekly limits, up to half of them. Pro and standard Team seats can use it too, but every token bills against usage credits, and there is no launch promo this round.
Fable 5 versus Fable 5.1 at a glance
| What changed | Claude Fable 5 | Claude Fable 5.1 |
|---|---|---|
| Base price | $10 in, $50 out per MTok | Unchanged |
| Cache reads | $1.00 per MTok | $0.25 per MTok, a 75% cut |
| Terminal-Bench-Science (Anthropic-reported) | 24.7% | 52.6% |
| Cyber safeguard false positives | Baseline | About 60% fewer per session |
| Thinking block portability | Readable by newer Claude models | Readable by Fable 5.1 only |
| Editing earlier turns | Permitted | Invalidates later thinking blocks |
| Pro launch credit | One-time credit in July 2026 | None |
| Data retention | Mandatory 30-day (Covered Model) | Unchanged |
The new lock-in: reasoning only Fable 5.1 can read
Under the pricing news sits the most consequential developer change. Every Fable 5.1 thinking block records the model that produced it, and readability runs in one direction only. Move a conversation from Opus 5 or Fable 5 onto Fable 5.1 and the reasoning survives. Move a Fable 5.1 conversation to any other model, including a refusal fallback landing on Opus 5, and the API drops its thinking blocks before the target model ever sees them. Without an opt-in beta header, the drop is silent: no error, no flag, the accumulated reasoning simply stops existing from that point on.
Editing history breaks it too. Change anything before a Fable 5.1 thinking block, the system prompt, the tool list, or an earlier message, and every later block is invalidated. For accounts created on or after August 31, 2026, the API enforces this with a 400 error stating "The block is bound to a different conversation." Anthropic's prescribed pattern is to treat conversations as append-only, adding instructions through mid-conversation system messages instead of rewriting history. Claude Code and claude.ai already keep the prefix intact. Custom integrations that inject and remove status lines, rotate system prompts, or reorder turns need rework before migrating.
Two more behavior notes belong on any migration checklist. At low effort, Anthropic says Fable 5.1 answers from memory more often, calling search and retrieval tools less, and recommends raising effort or adding a verification nudge for turns that need fresh information. And Anthropic concedes in the launch announcement that its automated behavioral audit currently provides less visibility into very long-context work and multi-agent settings, exactly the territory a 1 million token agentic model is sold for. Why very long context windows degrade in practice is a known pattern, examined in the 1M token context explainer.
Cheap cache reads are still not memory
The 75% cache cut invites a tempting mental model: the assistant now remembers your session at a quarter of the price. That conflates two different things. A prompt cache is a billing optimization with a five-minute or one-hour lifespan, not a store of knowledge. It saves money while a session is alive and vanishes when the session ends. Nothing in Fable 5.1 changes what happens after the last message: the context, the reasoning, and the accumulated decisions all evaporate together. The release notes list new pricing, new betas, and new binding rules, and no new memory feature of any kind.
Stack the constraints and the shape of the trade gets clear. Reasoning readable by exactly one model. Histories that must stay append-only or break. Conversations held under mandatory 30-day retention, excluded from zero-data-retention agreements unless Anthropic expressly authorizes an exception, with Enterprise Frontier Safeguards, which moves safety-monitoring data into the customer's own cloud, rolling out in phases starting this fall. And availability itself has precedent for vanishing: a government order took Fable 5 and Mythos 5 offline in June 2026, an episode dissected in the model shutdown post-mortem. Everything a long session builds now lives inside one vendor's one model, on one clock.
The part of your working context worth keeping should not be welded to a model version. An external memory layer such as MemX (memx.app) stores your decisions, preferences, and project facts outside any single vendor, so a refusal fallback to Opus, a migration off Fable 5.1, or a model retirement does not reset what your AI knows about your work. MemX is private by architecture: per-user isolation, customer-managed encryption keys, encryption at rest, on-device processing where possible, and no training on your data.
Who should switch, and who should wait
- API teams running long agent loops: the clearest winners. Re-run cost models with $0.25 cache reads; Anthropic's own estimate is up to 45% savings on highly agentic workloads.
- Opus 5 users with passing evals: stay put. Anthropic's guidance has not changed: start with Opus 5 at $5 and $25 per million tokens and step up only when higher effort still falls short.
- Pro subscribers: trying Fable 5.1 means buying usage credits, with no promotional credit this round. For everyday chat, the models bundled in the plan remain the sensible default.
- Multi-model routers and fallback pipelines: audit thinking-block handling first. Code that switches models mid-conversation or edits history will silently lose reasoning or start throwing 400 errors.
- Compliance-sensitive teams: confirm the mandatory 30-day retention fits your data policy before routing regulated work through a Covered Model.
Fable 5.1 lands as a disciplined point release: the same model family made meaningfully cheaper to run in loops, politer about refusing legitimate work, and stricter about who owns its reasoning. Developers with agentic workloads get real savings. What nobody gets, on any plan, is continuity that outlives the session, and that remains a problem you solve outside the model. For the cheaper tiers this release leaves untouched, the Opus 4.8 breakdown covers how Anthropic's mid-range stacks up.
01Is Claude Fable 5.1 included in the Claude Pro plan?
Pro and standard Team seats can use Fable 5.1 only through pay-as-you-go usage credits; it does not draw on plan usage limits. Max and premium Team or Enterprise seats include it for up to 50% of weekly limits, then pay-as-you-go. Unlike Fable 5's paid rollout, there is no promotional credit.
02How much does Claude Fable 5.1 cost per million tokens?
The API price is $10 per million input tokens and $50 per million output tokens, the same as Fable 5. Cache reads cost $0.25 per million, a 75% cut, cache writes run $12.50 to $20, and the Batch API halves input and output prices.
03What is the difference between Claude Fable 5.1 and Claude Mythos 5.1?
They are the same underlying model with identical specs and pricing but different safeguard levels. Mythos 5.1 relaxes cybersecurity and life-science restrictions and is invite-only through Project Glasswing, Anthropic's trusted-access program for vetted organizations. Fable 5.1 is the generally available version.
04Did Claude Fable 5.1 replace Claude Fable 5?
No. Both models remain available and selectable, and Fable 5.1 extends Fable 5 rather than retiring it. Anthropic commits to keeping Fable 5.1 available until at least September 1, 2027. Migrating means updating the model ID and reviewing three breaking changes, mostly around thinking blocks.
05Why does moving a Fable 5.1 chat to another model lose its reasoning?
Fable 5.1 thinking blocks are readable only by Fable 5.1. Switch the conversation to any other model, including Opus 5 as a refusal fallback, and the API drops those blocks, silently unless a beta header reports it. Editing earlier turns also invalidates later blocks, so Anthropic recommends append-only conversations.
