Release confirmed: Fable 5.1 launched on September 1, 2026
Anthropic released Claude Fable 5.1 and Claude Mythos 5.1 on September 1, 2026. The official model overview marks Fable 5.1 as Active (latest), gives a release date of September 1, 2026, and states retirement will be no sooner than September 1, 2027. [1][2]
The model is available through the Claude API as `claude-fable-5-1`, as well as Amazon Bedrock, Claude Platform on AWS, Google Cloud, and Microsoft Foundry. AWS and Google Cloud published same-day availability material. [2][8][10]
- Claude API ID: `claude-fable-5-1`
- Amazon Bedrock ID: `anthropic.claude-fable-5-1`
- Google Cloud ID: `claude-fable-5-1`
- Status: Active (latest)
- Release: September 1, 2026
Fable 5.1 and Mythos 5.1: same model, different safeguards
Anthropic states that Claude Fable 5.1 and Claude Mythos 5.1 are the same underlying model. The difference is the safeguard layer and access model. Fable 5.1 is generally available, while Mythos 5.1 is offered only to approved organizations through trusted-access programs including Project Glasswing. [1][3][9]
This distinction matters when reading security and cyber benchmarks. Anthropic reports 60.9% for Mythos 5.1 on Terminal-Bench 4.0 versus 55.8% for Fable 5.1 and attributes the gap to tasks where Fable safeguards intervened. [1]
Claude Fable 5.1 technical specifications
| Specification | Claude Fable 5.1 |
|---|---|
| Context | 1M tokens |
| Max output | 128K tokens |
| Thinking | Adaptive, always on |
| Default API effort | high |
| Input | text + images |
| Output | text |
| Reliable knowledge cutoff | June 2026 |
| Training cutoff | June 2026 |
Fable 5.1 has a 1M-token context window as both the default and maximum, with up to 128K output tokens. It accepts text and images and produces text. Its reliable knowledge cutoff and training data cutoff are both June 2026. [2][3]
Adaptive thinking is always on. The API model overview lists `high` as the default effort. Anthropic separately says Fable 5.1 defaults to High in Claude Code and Medium in Claude.ai and Cowork. [1][2][3]
Pricing: base rates stay flat, cache reads drop 75%
| Model | Input / MTok | Cache read / MTok | Output / MTok |
|---|---|---|---|
| Fable 5.1 | $10 | $0.25 | $50 |
| Fable 5 | $10 | $1.00 | $50 |
| Opus 5 | $5 | $0.50 | $25 |
| Sonnet 5 | $2 | $0.20 | $10 |
Base pricing is unchanged from Fable 5: $10 per million input tokens and $50 per million output tokens. A five-minute cache write costs $12.50/MTok and a one-hour write costs $20/MTok. The big change is cache read pricing: $0.25/MTok instead of $1.00/MTok on Fable 5. [3][4]
For Fable 5.1, a cache hit costs 0.025× the base input price, compared with the standard 0.1× multiplier used by other current Claude models. Batch API pricing is 50% off base input and output, or $5/MTok input and $25/MTok output for Fable 5.1. [4]
Where do Anthropic’s 25–45% cost-savings claims come from?
Anthropic estimates Fable 5.1 costs around 25% less than Fable 5 for typical token-billed workloads, while savings for highly agentic, cache-heavy workloads can reach roughly 45%. [1]
This is not a blanket 25–45% list-price reduction. Anthropic says its chart uses four weeks of actual August 2026 usage at default effort. The real savings depend on how much of a workload is served from prompt-cache reads. [1]
- Base input/output pricing: unchanged from Fable 5
- Cache read: 75% cheaper
- Typical workload: about 25% lower cost according to Anthropic
- Highly agentic workload: up to about 45% lower cost according to Anthropic
- Production economics should be measured per successful task, not only per token
Benchmarks: strong launch results, still vendor-reported
| 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% | — |
| HLE no tools | 60.9% | 57.8% | 56.6% | — |
| HLE 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% |
Anthropic published a launch comparison against Fable 5, Opus 5, and GPT-5.6 Sol. These are Anthropic-run launch results, not an independent certification that Fable 5.1 wins every workload. [1]
Methodology details matter. Fable 5.1 was evaluated with production safeguards enabled. Some safeguard interventions affected scoring, and Anthropic used Opus 4.8 or Opus 5 as fallbacks for selected cyber or biology interventions. OSWorld 2.0 uses the August 2026 task release, so these scores are not directly comparable with previously published OSWorld 2.0 numbers. [1]
Terminal-Bench-Science 0.1: 52.6% and the largest jump in the table
Anthropic reports 52.6% for Fable 5.1 on Terminal-Bench-Science 0.1, compared with 24.7% for Fable 5, 29.0% for Opus 5, and 22.4% for GPT-5.6 Sol. In this specific setup, Fable 5.1 scores more than twice as high as Fable 5. [1]
Anthropic also reports a standard error of roughly ±3.5–4.5 points per model. It compared its reproduced Opus 5 and Fable 5 results with the public leaderboard and says the differences are within noise. That is a useful reminder not to over-interpret one- or two-point gaps. [1]
Terminal-Bench 4.0 and the Fable–Mythos safeguard gap
Fable 5.1 scores 55.8% on Anthropic’s Terminal-Bench 4.0 table. Mythos 5.1, the same underlying model with different safeguards, scores 60.9%. Fable 5 is at 42.0%, Opus 5 at 52.3%, and GPT-5.6 Sol at 37.3%. [1]
Anthropic explicitly attributes the Fable–Mythos gap to tasks where earlier, less precise cyber safeguards intervened. Mythos 5.1 therefore should not be presented as the unrestricted score of the generally available Fable configuration. [1]
Knowledge work, HLE, and AutomationBench
On GDPval-AA v2, Fable 5.1 scores 1853 versus 1723 for Fable 5, 1824 for Opus 5, and 1711 for GPT-5.6 Sol. On Humanity’s Last Exam, Anthropic reports 60.9% without tools and 65.0% with tools. [1]
AutomationBench shows 31.4% for Fable 5.1, compared with 17.1% for Fable 5 and 26.9% for Opus 5. These numbers support the launch narrative around agent workflows and professional knowledge work, but they should still be validated on your own processes. [1]
CursorBench 3.2.0 and OSWorld 2.0
On CursorBench 3.2.0, Fable 5.1 scores 73.4%, Fable 5 70.5%, Opus 5 70.0%, and GPT-5.6 Sol 67.2%. The improvement here is much smaller than on Terminal-Bench-Science, so a claim that Fable 5.1 is “twice as good” across the whole benchmark suite would be false. [1]
On OSWorld 2.0, Fable 5.1 records 77.9% partial and 41.7% strict. Anthropic notes that these scores use the August 2026 task release with files that differ from older versions, making direct comparison with earlier public OSWorld results invalid. [1]
Effort matters more than before
Fable 5.1 exposes effort controls, and Anthropic says the performance gap over Fable 5 is widest at higher effort. At the same time, Low or Medium effort can deliver similar or better results than Fable 5 at lower cost. [1][3]
The API also adds beta support for changing effort mid-conversation without invalidating the prompt cache. That enables practical routing: lower effort for routine steps, higher effort for difficult planning, debugging, or research. [3]
Core capability gains: long agent runs, research, and finished artifacts
Anthropic groups Fable 5.1 improvements into six areas: long-session agentic coding, knowledge work with documents/spreadsheets/slides, research and search, vision, long-context reasoning, and computer use. [3]
That positioning matters more than a chatbot-style score. Fable 5.1 is intended to orchestrate and execute multi-hour processes across tools, repositories, documents, and desktop applications. Enterprise evals should measure task completion, retries, tool calls, token use, wall-clock time, and cost per successful task. [3][6]
Scientific research: important examples, but distinguish Fable from Mythos
Anthropic devotes a major part of the launch to scientific research. Its most advanced examples often use Mythos 5.1, the restricted version, including protein-design experiments and autonomous research workflows. [1]
Anthropic reports that Mythos 5.1 designed binders with 10× higher affinity than the best designs in specified Adaptyv Bio competitions on three targets and reached a hit rate close to 50% across 12 targets. These are Anthropic-reported research results and should not be automatically attributed to the public Fable 5.1 safeguard configuration. [1]
Safeguards: fewer false positives, not fewer boundaries
Anthropic says its newest cybersecurity safeguards produce about 60% fewer false positives than before. Fable 5.1 can now be used for source-code vulnerability discovery, while exploit development, penetration testing, and binary-based vulnerability scanning remain restricted depending on category. [1][3]
For biology, Anthropic reports about 85% fewer interventions on benign elementary biology and medical requests compared with the safeguards that launched with Fable 5. Dual-use life-science questions can still be routed to other models or restricted. [1]
Data retention and Enterprise Frontier Safeguards
Platform documentation treats Fable 5.1 and Mythos 5.1 as Covered Models with 30-day data retention and no standard zero-data-retention access unless Anthropic expressly authorizes it. [3]
At the same time, Anthropic announced Enterprise Frontier Safeguards (EFS), designed to keep safety-monitoring data in customer-controlled cloud infrastructure. Anthropic says rollout will begin in phases later this fall, while eligible customers can receive authorized ZDR before EFS is fully available. AWS documents additional provider-specific retention modes. [1][8]
Content provenance: text watermarking and C2PA for supported files
Text generated by Fable 5.1 and Mythos 5.1 carries Anthropic’s statistical text watermark on every platform where the model is available. Anthropic says it adds no tokens or hidden characters and does not change meaning or readability. [3]
Supported image and video files generated through Claude, such as through code execution, can carry signed C2PA Content Credentials when retrieved through the Files API. This does not mean every asset in every Claude product automatically receives C2PA metadata. [3]
Breaking change #1: forced tool use is unsupported
{
"tool_choice": {"type": "auto"}
}Fable 5.1 does not support `tool_choice` values of `any` or forcing a specific `tool`. Those requests return HTTP 400 `invalid_request_error`. `tool_choice: auto` and `none` continue to work. [3][5]
For schema-valid JSON, Anthropic recommends strict tool use with `auto` or structured outputs. If the model must call a tool, state when the tool should be used in the prompt instead of forcing the call through the parameter. [3]
Breaking change #2: thinking blocks require append-only history
Fable 5.1 can read thinking blocks from earlier Claude models, but earlier models cannot read Fable 5.1 thinking blocks. When routing down to an older model, the API can therefore drop incompatible blocks. [3]
More importantly, editing an earlier system prompt, tools definition, or message can invalidate later thinking blocks. Anthropic recommends treating conversation history as append-only and using mid-conversation system/tool changes instead of rewriting the conversation prefix. [3][5]
New API features: per-message effort, turn-scoped system messages, and progress updates
Fable 5.1 adds several beta features for long-running agents. Developers can change effort per message, use turn-scoped system messages, and receive readable progress status between tool calls through `thinking.display: "updates"`. Each feature requires the beta header documented by Anthropic. [3]
This matters because Fable 5.1 tends to emit fewer user-facing progress updates between tools than Fable 5. Applications whose UI depends on narration should configure and prompt for that behavior explicitly. [3][6]
Behavior differences: parallel tools, low-effort search, and writing style
Anthropic warns that parallel tool calling is more variable in Fable 5.1: the model may issue one tool call per turn where Fable 5 batched several. That can add round trips and wall-clock time without reducing answer quality. [3][6]
At `low` effort the model is also more likely to answer from memory and less likely to trigger search or retrieval. The docs call out denser prose, less formatting in chat, and a greater tendency to reproduce passages without quotation marking during document summaries. Integrations should retest prompting and attribution behavior. [3][6]
Migrating from Fable 5: a practical checklist
The obvious change is `claude-fable-5` → `claude-fable-5-1`, but the migration guide requires more: remove forced tool choice, preserve thinking blocks unchanged, keep history append-only, retune effort, and inspect agent loops for reduced tool batching. [5]
Anthropic also ships a Claude Code migration skill: `/claude-api migrate this project to claude-fable-5-1`. It can handle model-ID and some API updates, but production rollout still requires your own regression and task-level evaluations. [5]
- Update the model ID.
- Remove `tool_choice: any` and forced specific tools.
- Do not edit earlier turns when preserving thinking blocks.
- Test routers and fallbacks to older models.
- Recalibrate effort.
- Measure tool-call count and latency.
- Recalculate costs with the new cache-read rate.
- Run your own evals before production rollout.
Fable 5.1 vs Opus 5 vs Sonnet 5: which should you use?
| Model | Context | Max output | Input | Output | Latency | Best starting point |
|---|---|---|---|---|---|---|
| Fable 5.1 | 1M | 128K | $10 | $50 | Slower | Hardest long-horizon tasks |
| Opus 5 | 1M | 128K | $5 | $25 | Moderate | Default for difficult workloads |
| Sonnet 5 | 1M | 128K | $2 | $10 | Fast | Scale, speed, cost |
Anthropic itself recommends starting most workloads with Claude Opus 5. Fable 5.1 is positioned for demanding reasoning and long-horizon agentic work, or when your own Opus 5 evals at higher effort still fall short. [2][3]
Opus 5 costs half as much as Fable 5.1 for base input/output tokens: $5/$25 versus $10/$50. Sonnet 5 is cheaper again at $2/$10 and has a faster latency profile. All three current models provide 1M context and 128K max output. [2][4]
Availability, plans, and lifecycle
Fable 5.1 is available through Claude Platform and major partner clouds. Anthropic also lists it for Pro, Max, Team, and Enterprise users in Claude products. [2][7]
Its lifecycle status is Active (latest), released September 1, 2026, with retirement not sooner than September 1, 2027. Regional availability, pricing, and retention can differ on partner clouds, so production deployments should verify provider-specific documentation. [2][8][10]
What actually changed from Fable 5?
Fable 5.1 is not simply “Fable 5 with a lower price.” Base input/output pricing is identical, but cached context is much cheaper, agentic and research benchmarks improve, the API has breaking changes, and long conversations are more tightly coupled to thinking-block integrity. [1][3][4]
There is still no evidence from the launch table that Fable 5.1 will be the best model for every task. Even Anthropic recommends Opus 5 as the starting point for most workloads. A production decision should be based on your own evals, cost per successful task, latency, tool reliability, long-context quality, and safety/data-retention constraints. [2][3]
