Quick answer: Claude Fable 5.1 is Anthropic's frontier model, released September 1, 2026. In the API it is claude-fable-5-1: a 1M-token context window, 128K max output, a June 2026 knowledge cutoff, adaptive thinking that is always on, and pricing of $10 per 1M input / $50 per 1M output — identical to Claude Fable 5. The widely quoted "25% cheaper" figure does not come from a rate cut. It comes from one line item: cache reads, dropped 75% to $0.25 per 1M tokens. Whether you see 25%, 45%, or 0% savings depends entirely on how much of your bill was cache reads.
That distinction is the single most consequential thing to understand about this release, and it is missing from nearly every summary. This article covers what Fable 5.1 actually is, the three changes that will break existing Fable 5 integrations, and where the savings do and do not land.
Credibility note: every specification, price, and score below comes from Anthropic's own launch announcement and the Claude Platform model documentation, both read on September 6, 2026, cross-checked against OpenAI's independently published comparison table and Artificial Analysis's live leaderboard. We have not run our own benchmark suite; vendor numbers are labelled as such. Anthropic evaluated Fable 5.1 with production safeguards enabled and says this likely depressed some of its own scores — a caveat we carry through rather than quietly drop. Research and drafting were AI-assisted; every figure was checked back to the primary sources listed at the end before this went live.
What This Article Solves
The pain point is a headline that describes a mechanism, not a discount. "Fable 5.1 costs 25% less" reads like a price cut. It is not one. The list rates are unchanged to the cent, and the savings arrive only through prompt caching. If your integration does not use caching — or uses it lightly — your bill after switching to Fable 5.1 is the same bill you had before, and you will spend a week wondering where the 25% went.
By the end of this page you will know what Fable 5.1 is, exactly which pricing line moved and by how much, the three breaking changes waiting in a Fable 5 migration, what the safeguard changes mean if you work in security or life sciences, and how Claude Mythos 5.1 differs from it.
Claude Fable 5.1 at a Glance
| Field | Value |
|---|---|
| Vendor | Anthropic |
| Released | September 1, 2026 |
| Claude API model ID | claude-fable-5-1 |
| Amazon Bedrock ID | anthropic.claude-fable-5-1 |
| Context window | 1M tokens |
| Max output | 128K tokens |
| Knowledge cutoff | June 2026 (training data cutoff also June 2026) |
| Thinking | Adaptive, always on |
| Default effort | high |
| Modalities | Text + images in, text out |
| Input price | $10 / 1M tokens |
| Output price | $50 / 1M tokens |
| Cache read | $0.25 / 1M tokens |
| 5-minute cache write | $12.50 / 1M tokens |
| 1-hour cache write | $20 / 1M tokens |
| Batch API | 50% discount on input and output |
| Retirement | Not sooner than September 1, 2027 |
| Platforms | Claude API, Amazon Bedrock, Google Cloud, Microsoft Foundry, Claude Platform on AWS |
Effort is a first-class control here, not a hidden setting. Fable 5.1 supports low, medium, high, xhigh, and max, and the defaults differ by surface: High in Claude Code, Medium in Claude Cowork and on claude.ai. Anthropic's own claim is that at Low or Medium effort, Fable 5.1 reaches results "similar to or better than" Fable 5's at substantially lower cost — which makes effort the most underused cost lever in the lineup.
The 25% That Is Really a Cache Discount
Here is the mechanism, stated plainly.
Anthropic cut the price of cache reads — tokens the model re-reads from context it has already processed — by 75%, from $1.00 to $0.25 per 1M tokens. Nothing else changed. Input is still $10. Output is still $50. Cache writes are still $12.50 for the 5-minute tier and $20 for the 1-hour tier.
Why that produces a double-digit headline: in long agentic sessions, the same system prompt, tool definitions, and accumulated conversation get re-read on every single turn. After a few dozen turns, cache reads are not a line item on the bill — they are the bill. Cutting them by three quarters removes most of the cost.
Anthropic's published figures, based on four weeks of actual August 2026 usage at default effort:
| Workload shape | Indexed cost (Fable 5 = 100) | Savings |
|---|---|---|
| Typical (Claude Enterprise + Claude Code + API) | 75 | ~25% |
| Highly agentic (context-heavy, tool-heavy) | 55 | ~45% |
And the case nobody advertises: a single-turn, no-caching request costs exactly what it cost on Fable 5. Short classification calls, one-shot extraction, stateless API requests with no reused prefix — zero savings. Not a small saving. Zero. We break that arithmetic down turn by turn in the full Fable 5.1 rate-card analysis, including three worked examples that land on 0%, 26%, and 45%.
The practical read: if you are running long agent loops, this is a large and real cut and you should switch. If you are running short stateless calls, Fable 5.1 is a capability upgrade at unchanged cost, which is still good — just not the thing the headline promised you.
What the Benchmarks Show
Anthropic's published comparison (September 1, 2026):
| 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% |
| GDPval-AA v2 | 1853 | 1723 | 1824 | 1711 |
| OSWorld 2.0 (partial) | 77.9% | 72.9% | 75.4% | — |
| OSWorld 2.0 (strict) | 41.7% | 36.1% | 39.6% | — |
| Humanity's Last Exam (no tools) | 60.9% | 57.8% | 56.6% | — |
| Humanity's Last Exam (with tools) | 65.0% | 63.8% | 63.6% | — |
| AutomationBench | 31.4% | 17.1% | 26.9% | 19.6% |
| CursorBench 3.2.0 | 73.4% | 70.5% | 70.0% | 67.2% |
Terminal-Bench-Science more than doubling — 24.7% to 52.6% in one point release — is the standout, and it is the number Anthropic built its scientific-research narrative around. For where the predecessor sat against open-weight rivals, see our earlier Fable 5 benchmark comparison.
Two honesty notes Anthropic itself supplies, both worth carrying forward. First, on Terminal-Bench-Science the standard error is ±3.5–4.5 points per model, which means differences under about 8 points between any two models on that benchmark are inside the noise. Second, Fable 5.1 was evaluated with production safeguards enabled; where safeguards intervened on OSWorld 2.0 and AutomationBench tasks, the model scored zero. Anthropic states this "likely reduces the performance of Fable 5.1" on those benchmarks. A vendor volunteering that its own numbers are depressed is unusual, and it should raise rather than lower your confidence in the rest.
The cross-check that matters most. OpenAI's GPT-6 Astra announcement, published two days later, independently lists Fable 5.1 at 55.8% on Terminal-Bench 4.0 and 52.6% on Terminal-Bench Science 0.1 — identical to Anthropic's own figures. Two competing vendors reporting the same score for the same model is the strongest verification available for a five-day-old release. Meanwhile Artificial Analysis's live leaderboard, checked September 6, currently ranks Claude Fable 5.1 (max with fallback) first on its Intelligence Index at 57, ahead of GPT-6 Astra at 55.
The cost side of that same independent dashboard is less flattering: Artificial Analysis puts Fable 5.1's cost per Intelligence Index task at $6.12 against Astra's $2.57. Fable 5.1 scores marginally higher and spends considerably more tokens doing it. If you want the current top of the intelligence leaderboard, this is it; if you want frontier-adjacent results per dollar, it is not.
Three Things That Break in a Fable 5 Migration
Anthropic documents these as breaking. If you have an existing Fable 5 integration, read this list before you swap the model ID:
- Forced tool use returns an error. If your code pins a specific tool via
tool_choice, that call now fails rather than degrading. - Earlier models cannot read its thinking blocks. Any pipeline that hands a Fable 5.1 response to an older Claude model for a second pass will break on the thinking content.
- Editing earlier turns invalidates thinking blocks. Rewriting conversation history — a common pattern in agent frameworks that prune or summarise context — discards the preserved reasoning.
Five changes are purely additive: per-message effort (beta, letting you change effort mid-conversation without invalidating the prompt cache), turn-scoped system messages (beta), readable progress updates between tool calls via display: "updates" (beta), the lower cache read price, and content provenance.
That third breaking change deserves a flag for agent builders specifically. Context-window management by editing history is standard practice in most agent frameworks. On Fable 5.1 that practice now costs you the model's preserved reasoning, which is precisely the thing long-horizon work depends on. Plan for per-message effort and Anthropic's preserved-thinking documentation instead.
Safeguards: Loosened in Two Directions, Still Firm in Others
Anthropic paired the capability release with a set of safeguard changes aimed squarely at false positives — the complaint that benign work kept getting blocked.
Cybersecurity. Fable 5.1 can now be used to identify software vulnerabilities, which Fable 5 would not do. Anthropic reports Claude Code users should see roughly 60% fewer safeguard interventions per session. Still redirected to the Opus models: penetration testing, exploit generation, and binary-based vulnerability scanning. So: finding bugs, yes; weaponising them, no.
Biology. The latest safeguards fire 85% less often for benign elementary-biology and medical questions than the ones that shipped with Fable 5. Genuine life-sciences R&D queries are still routed to the Opus models.
Anti-distillation. API accounts created from September 1, 2026 onward can no longer manually edit Claude's prior context in a multi-turn conversation while preserving the transcript of its prior thinking. Existing accounts are unaffected for now, but Anthropic says the change will apply to everyone with future model releases. If your integration does exactly this, you have a deprecation clock running.
EU AI Act watermarking. Under the Code of Practice Anthropic signed in July 2026, models released after August 2, 2026 carry an invisible statistical watermark. Anthropic states it contains no information about the user, their organization, or their conversations, and has no practical impact on output quality. A detection API is in private preview for regulators, researchers, and obligated enterprises.
Claude Mythos 5.1: Same Model, Different Locks
This is the most commonly misunderstood part of the launch. Claude Mythos 5.1 is not a bigger model. Anthropic is explicit: Fable 5.1 and Mythos 5.1 are the same model with different levels of safeguards. Identical specifications, identical pricing.
Mythos 5.1 is available by invitation only, through Project Glasswing, via two trusted-access routes: the Cyber Verification Program for defensive security work, and the Life Sciences Verification Program, built in partnership with the US government. Currently it is limited to a set of US organizations. It also powers Claude Security, Anthropic's codebase vulnerability scanner.
The one place the difference shows up in benchmarks: Anthropic reports Mythos 5.1 at 60.9% on Terminal-Bench 4.0 against Fable 5.1's 55.8% — the same weights, five points apart, purely because the safeguards stop intervening. That gap is a useful estimate of how much capability the production safeguards cost on security-adjacent work.
Who Should Actually Use It
Anthropic's own guidance is unusually restrained: the documentation says to start with Claude Opus 5 for most workloads at $5/$25, and reach for Fable 5.1 "for demanding reasoning and long-horizon agentic work, or when your evals on Claude Opus 5 at higher effort still fall short."
That is a vendor telling you its most expensive model is not the default. Take the advice.
Fable 5.1 earns its price when: you run long, context-heavy agentic sessions where the cache-read cut compounds turn over turn; you need the current top of the independent intelligence leaderboard; you do multi-hour or multi-day unattended runs; or you have already tried Opus 5 at high effort and hit a ceiling.
Skip it when: your calls are short and stateless — you pay Fable prices for zero cache savings; Opus 5 already passes your evals, at half the input rate and half the output rate; or your bottleneck is latency, since Fable 5.1's comparative latency is documented as the slowest in the current Claude lineup.
The cheapest way to settle it is a head-to-head on your own hardest task. Run Claude Fable 5.1 on glm5.app next to whatever you use today — one account, one credit balance, no second vendor contract — and diff the outputs before you migrate anything.
Frequently Asked Questions
Is Fable 5.1 cheaper than Fable 5? Only if you use prompt caching. The list rates are identical: $10 input, $50 output. The savings come entirely from cache reads dropping 75% to $0.25 per 1M tokens. Anthropic measures ~25% lower cost on typical workloads and up to ~45% on highly agentic ones. Single-turn requests without caching cost exactly the same as before.
Is Claude Fable 5.1 free? No. There is no free API tier for it, and Anthropic's own documentation recommends Claude Opus 5 as the default for most workloads. It is accessible through paid Claude plans and paid API access on the Claude API, Bedrock, Google Cloud, Microsoft Foundry, and Claude Platform on AWS.
What is the difference between Claude Fable 5.1 and Claude Mythos 5.1? They are the same model with different safeguards. Mythos 5.1 has more permissive safeguards for cybersecurity and life-sciences work and is invitation-only through Project Glasswing, currently limited to a set of US organizations. Same specifications, same pricing.
When was Claude Fable 5.1 released? September 1, 2026, alongside Claude Mythos 5.1. Anthropic's documentation commits to retirement no sooner than September 1, 2027 — a one-year minimum lifetime, useful if you are planning a long-lived integration.
What is Claude Fable 5.1's context window? 1M tokens, with 128K max output. Unlike some competitors, Anthropic does not publish a surcharge for requests above a given input-token threshold.
Does Claude Fable 5.1 beat GPT-6 Astra? It depends on the benchmark, and both vendors' tables agree more than you would expect. OpenAI's own launch table shows Fable 5.1 ahead on Humanity's Last Exam with tools (65.0% vs 57.2%) and on the Artificial Analysis Intelligence Index version OpenAI cited (65.7 vs 61.2), while Astra leads on Terminal-Bench 4.0, Terminal-Bench Science, DeepSWE, and computer-use benchmarks. On cost per task, independent measurement favours Astra substantially. Our side-by-side of both vendors' tables works through where the two disagree and where they corroborate each other.
Sources
- Introducing Claude Fable 5.1 and Claude Mythos 5.1 — Anthropic — the launch announcement, September 1, 2026: cache-read price change and indexed cost figures, benchmark table with standard-error and safeguard caveats, safeguard changes, anti-distillation, EU AI Act watermarking, and Mythos access programs.
- Claude Fable 5.1 — Claude Platform documentation — model IDs per platform, the full rate card including both cache-write tiers, context and output limits, knowledge cutoff, default effort, release and retirement dates, and the breaking/additive change list.
- Models overview — Claude Platform documentation — lineup comparison and Anthropic's own model-selection guidance.
- GPT-6 Astra: A new generation of intelligence — OpenAI — an independent vendor's published Fable 5.1 scores, used here to cross-check Anthropic's own numbers.
- Artificial Analysis — model comparison — third-party Intelligence Index, output speed, and cost-per-task figures, checked September 6, 2026.
All figures verified September 6, 2026. Rate cards and safeguard behaviour change; confirm against Anthropic's pricing documentation before committing a production budget.
