ByClaude Fable 5.1 is Anthropic’s successor to Claude Fable 5, built for long-running agentic coding, multistep research, and document and spreadsheet work, and it launched on September 1, 2026 alongside a restricted sibling model called Claude Mythos 5.1. The headline change is not a bigger model, it is a smarter and cheaper one: cache reads now cost a quarter of the previous rate, and capability gains concentrate at the higher effort levels developers actually use for hard problems.
This guide covers what changed under the hood, what it costs, which breaking changes could affect a production integration built on the previous version, and how it stacks up against the other major model releases that landed the same week.
What Is Claude Fable 5.1?
Claude Fable 5.1 is Anthropic’s latest large language model, positioned as the model to reach for when a task needs demanding reasoning or long-horizon agentic work that Claude Opus 5 does not fully cover. It ships with a 1 million token context window at standard per-token pricing across the whole window, and up to 128,000 output tokens per response, according to Anthropic’s official model documentation.
Anthropic describes Claude Fable 5.1 as achieving similar or better results than Claude Fable 5 at low or medium effort, with the gap widening substantially at higher effort tiers. Adaptive thinking is always on for this model, and the effort parameter controls how deeply it reasons before responding.
The model’s tokenizer is unchanged from Claude Fable 5, which itself was introduced with Claude Opus 4.7. Compared with Claude models older than Opus 4.7, the same input text produces roughly 30% more tokens under this tokenizer, which matters when estimating cost from older benchmarks or logs.
How Is Claude Fable 5.1 Different From Claude Fable 5?
The biggest practical difference is where the model gets better and where it behaves differently by default. Anthropic’s documentation groups the capability gains into six areas, each aimed at sessions that run far longer than a typical chat exchange.
- Agentic coding over long sessions — multi-file features, large refactors, migrations, and debugging across sessions that run for hours rather than minutes
- Knowledge work with documents, spreadsheets, and slides — taking an analysis from a first question to a finished document or a live-formula spreadsheet
- Research and search — higher accuracy on multistep web research that follows up on what it finds rather than stopping at the first result
- Vision — reading dense charts, filings, and tables nested inside PDFs, including crop-and-zoom on individual charts
- Long-context work — reasoning over and connecting details across the full 1 million token context window
- Computer use — operating a browser and desktop applications more reliably, and recovering from failed steps instead of getting stuck
Multilingual performance stays on par with Claude Fable 5 rather than improving, which Anthropic calls out directly rather than leaving implied. If your workload is multilingual chat rather than long-horizon agentic work, Fable 5.1 may not be a meaningful upgrade over Fable 5 for that specific use case.
Behavior changes matter just as much as capability gains for anyone building on the API. Claude Fable 5.1 tends to issue one tool call per turn where Claude Fable 5 often batched several, it writes fewer progress updates during long tool runs unless you opt into a display setting, and it is more likely to rewrite a whole file for a small text edit rather than making a targeted change.
None of these are regressions in output quality, according to Anthropic, but they do change token usage and response pacing, which is worth testing against your own evaluation suite before switching production traffic over.
What Is Claude Mythos 5.1 and How Does It Differ From Fable 5.1?
Claude Mythos 5.1 shares the exact same underlying capabilities as Claude Fable 5.1, but with a different configuration of safety safeguards, and it is available only to approved participants in Anthropic’s Project Glasswing program rather than to the general public.
Anthropic positions Mythos specifically for work in cybersecurity and life sciences, where Fable’s default safeguards would otherwise block legitimate research. Both models share the same context window, output limits, and pricing, so the distinction is entirely about who can access the model and under what safety configuration, not about raw capability.
How Much Does Claude Fable 5.1 Cost?
Base input and output pricing for Claude Fable 5.1 is unchanged from Claude Fable 5, but cache read pricing drops to a quarter of the previous rate, which is the single biggest cost lever for anyone running long agentic sessions that repeatedly reread a cached prompt prefix.
| Pricing Component | Rate (USD per million tokens) |
|---|---|
| Base input | $10 |
| 5-minute cache writes | $12.50 |
| 1-hour cache writes | $20 |
| Cache reads | $0.25 |
| Output | $50 |
Cache reads on Claude Fable 5.1 cost 0.025 times the base input price, compared with 0.1 times on most other Claude models, according to Anthropic’s pricing documentation. For a long-running coding agent that rereads the same cached context dozens of times per session, that difference compounds quickly across a full day of usage.
What New Features Does Claude Fable 5.1 Add?
Five additive features ship alongside Claude Fable 5.1, and four of them are currently in beta. Per-message effort lets you raise or lower reasoning depth mid-conversation without invalidating the prompt cache, which is useful when only one step in a workflow actually needs deep reasoning.
- Per-message effort (beta) — change the effort level between turns in the same conversation
- Turn-scoped system messages (beta) — a system message that applies to one turn only, then stops rendering
- Progress updates between tool calls (beta) — opt in to see the model’s short status updates during long agent runs
- Lower cache read pricing — a permanent 75% reduction versus Claude Fable 5
- Content provenance — generated text carries Anthropic’s statistical watermark by default
Content provenance applies automatically and needs no configuration. Anthropic states the watermark does not affect output quality, adds no extra tokens, and carries no identifying information about the account that generated the text.
What Do Turn-Scoped System Messages Look Like in Practice?
A turn-scoped system message carries system-prompt authority for exactly one turn, then stops rendering once a later user message arrives, without you needing to delete it from the conversation history afterward. This is useful for one-off reminders inside a long tool-calling loop, such as telling the model to check a specific inbox before running more code.
{
"role": "system",
"clear_at": "next_user_message",
"content": "Results have landed in your inbox. Check it before running more code."
}Because the message stays in the conversation and is simply skipped after its turn, the prompt cache keeps matching and later thinking blocks stay valid. A cleared turn-scoped message also costs no input tokens on later requests, which makes it cheaper than the old pattern of injecting and deleting text by hand.
What Breaking Changes Should Developers Know About?
Three changes can break existing integrations built for Claude Fable 5, and all three are documented directly by Anthropic rather than discovered after the fact. Forced tool use is the first: setting tool_choice to "any" or a named tool now returns a 400 error, because thinking is always on and a forced tool call would skip it.
tool_choice: type "tool" and "any" are not supported for this model.That is the exact error text the API returns. The fix is to keep tool_choice set to "auto", the default, and use strict tool use or structured outputs instead when you need guaranteed schema-valid arguments back from the model.
The second change involves thinking blocks. Every thinking block records which model produced it, and Claude Fable 5.1 can read earlier models’ thinking blocks, but no earlier Claude model can read blocks that Claude Fable 5.1 produced. A conversation that switches away from Claude Fable 5.1 loses that reasoning for the turns handled elsewhere.
The third change is stricter conversation handling: editing anything before a Claude Fable 5.1 thinking block, such as the system prompt, the tools array, or an earlier message, triggers an error on the next request unless you explicitly opt into dropping the affected block. Treating the message history as append-only avoids this entirely.
Where Can You Access Claude Fable 5.1?
Claude Fable 5.1 is available to all Claude API customers under the model ID claude-fable-5-1, and it is also available through Claude in Amazon Bedrock, Claude on Google Cloud, and Claude in Microsoft Foundry, according to Anthropic’s availability table.
Claude Mythos 5.1 is restricted to approved Project Glasswing participants and is not available through general self-serve signup on any platform. Anyone who needs Mythos access has to go through their Anthropic, AWS, or Google Cloud account team directly.
How Do You Migrate From Claude Fable 5 to Fable 5.1?
Start by swapping the model string from claude-fable-5 to claude-fable-5-1 in your API calls, then work through the four items that actually change behavior rather than assuming a drop-in replacement will perform identically.
- Remove any forced tool_choice of type “any” or “tool” and move schema enforcement to strict tool use or structured outputs instead
- Keep your message history append-only so thinking blocks stay valid across the conversation
- Re-tune your effort level from the default rather than assuming Claude Fable 5 settings still fit
- Watch agent loops for single tool calls where Claude Fable 5 used to batch several, and add an explicit batching instruction if that matters for your latency budget
Anthropic’s own guidance stresses re-running your evaluation suite after migrating, since refusal handling, fallback behavior, and token counting all carry over unchanged, but default response style and tool-calling patterns do not.
Should You Switch to Claude Fable 5.1 Right Now?
Switch now if you run long agentic coding sessions, multistep research pipelines, or document and spreadsheet generation, since these are exactly the workloads Anthropic optimized this release around, and the lower cache read price reduces cost for any of them immediately.
Hold off briefly if your integration depends on forced tool_choice, hand-built message history that edits earlier turns, or a UI that relies on the exact volume of progress-update text Claude Fable 5 used to produce. Test against a staging environment first in those cases, since all three are documented breaking or behavioral changes rather than bugs you can wait out.
For everyday chat, short-form content generation, or simple single-turn tasks, Anthropic’s own guidance still points to Claude Opus 5 first. Fable 5.1 earns its higher cost profile specifically on long, effort-heavy work, not on quick one-off requests.
How Does Claude Fable 5.1 Compare to Other September 2026 Model Releases?
Claude Fable 5.1 launched in a crowded week for model releases. The table below compares it against three other models that shipped within days of it, based on positioning each vendor has publicly stated.
| Model | Vendor | Release Date | Primary Focus |
|---|---|---|---|
| Claude Fable 5.1 | Anthropic | Sept 1, 2026 | Long-running agentic coding and knowledge work |
| GPT-6 Astra | OpenAI | Sept 3, 2026 | General-purpose flagship reasoning |
| Gemini 3.8 Flash | Sept 2026 | Fast, low-cost everyday tasks | |
| Qwen3.8-Max-0902 | Alibaba / Qwen | Sept 2026 | Open-weight large-scale reasoning |
For a closer look at OpenAI’s release from the same week, see our GPT-6 Astra guide. What separates Claude Fable 5.1 from the rest of this list is less about raw benchmark position and more about its focus on long, multi-hour agentic sessions rather than single-turn chat quality.
None of these four releases directly overlap in positioning the way earlier model generations did. Gemini 3.8 Flash targets cost-sensitive, high-volume tasks, Qwen3.8-Max-0902 targets teams that want open weights they can self-host, GPT-6 Astra targets a general-purpose flagship experience, and Claude Fable 5.1 targets teams already running production agents that need to go longer and deeper on a single hard problem. Picking between them comes down to matching the workload rather than chasing a single leaderboard number.
Frequently Asked Questions
Is Claude Fable 5.1 the same as Claude Opus 5?
No. Anthropic recommends starting with Claude Opus 5 for most workloads and reaching for Claude Fable 5.1 specifically for demanding reasoning or long-horizon agentic work where Opus 5 falls short at higher effort. They are separate models with separate model IDs and pricing.
Does Claude Fable 5.1 cost more than Claude Fable 5?
No, base input and output pricing is identical to Claude Fable 5. The only pricing change is cache reads, which drop to a quarter of the previous cost, making Claude Fable 5.1 cheaper overall for any workload that relies on prompt caching.
Can I access Claude Mythos 5.1 without joining Project Glasswing?
No. Claude Mythos 5.1 is restricted to approved Project Glasswing participants working in cybersecurity or life sciences. General API customers can use Claude Fable 5.1, which shares the same underlying capabilities under a different safeguard configuration.
Will my existing Claude Fable 5 prompts work without changes on Fable 5.1?
Most prompts work, but forced tool_choice settings will fail with a 400 error, and code that edits earlier conversation turns can invalidate thinking blocks. Review Anthropic’s migration guide and re-run your evaluation suite before moving production traffic over.
Why does my agent make fewer parallel tool calls on Claude Fable 5.1?
This is a documented behavior change, not a bug. Claude Fable 5.1 may issue one tool call per turn where Claude Fable 5 batched several independent calls together. Add an explicit instruction to batch independent tool calls if your workflow depends on that pattern.
What happens to my existing Claude Fable 5 API integration if I do nothing?
Nothing changes automatically. Claude Fable 5 remains available under its own model ID, so an integration that never updates its model string keeps calling Claude Fable 5 indefinitely. You only see Claude Fable 5.1’s pricing, features, or breaking changes once you explicitly switch the model ID in your code, which means you can test the migration on a branch or staging environment at your own pace rather than being forced onto the new version by a deprecation deadline.
Final Thoughts on Claude Fable 5.1
Claude Fable 5.1 is an incremental but meaningful upgrade: same base pricing, a much cheaper cache read rate, and capability gains concentrated exactly where long agentic sessions need them most. The behavior changes around tool batching and progress updates are worth testing before a production migration, but nothing here requires waiting.
The clearest signal in this release is Anthropic’s own framing: it is not pitched as a replacement for daily-driver chat, it is pitched at the specific moment when a team’s evals on Claude Opus 5 at higher effort still fall short. That is a narrower, more honest positioning than most model launches offer, and it makes the upgrade decision easier rather than harder. Read the migration notes once, run your own evals, and let the results decide rather than the marketing copy.
If you are deciding between Anthropic’s current lineup, our Claude Opus 5 guide and Claude Sonnet 5 pricing guide cover the two models most people should start with before reaching for Fable 5.1’s higher effort tiers, and our Claude AI updates 2026 roundup tracks every release across the year.
Subscribe for Newsletter
