Claude Fable 5.1 Launch: Pricing, Benchmarks, and What’s New

Claude Fable 5.1 launched September 1, 2026 at Fable 5’s prices ($10/$50 per million tokens), with cache reads cut 75% and clear benchmark leads over Opus 5. This guide covers the full Claude Fable 5.1 launch: pricing, benchmarks, and what’s new, including three breaking API changes, five beta features, and the migration path from Fable 5 or Opus 5. If you run Claude in production, the details below tell you whether to upgrade, and what breaks if you do.

The launch numbers that matter most:

• $0.25 per million tokens for cache reads in 2026, down from $1 on Fable 5, per Anthropic’s Claude pricing documentation
• Up to 45% cheaper agentic and coding workloads versus Fable 5, by Anthropic’s 2026 estimates
• 52.6% on Terminal-Bench-Science 0.1 in September 2026, more than double Fable 5’s 24.7%, per Anthropic’s launch announcement
• Three breaking changes for existing Fable 5 code in 2026: forced tool use, model-bound thinking blocks, and history editing, per Anthropic’s what’s-new documentation
• June 2026 knowledge cutoff, the most recent in Anthropic’s current lineup

What’s New in Claude Fable 5.1 at Launch

Claude Fable 5.1 shipped on September 1, 2026 as the successor to Claude Fable 5, keeping the same $10/$50 per-million-token prices while cutting cache reads to a quarter of the old rate and layering in eight documented changes: three breaking, five additive. Anthropic positions it for demanding reasoning and long-horizon agentic work, and is explicit that most teams should start with Claude Opus 5 and reach for Fable 5.1 only when Opus 5 at higher effort still misses on their evals.

The core specs carry over. A 1M-token context window at standard per-token pricing across the whole window, 128K max output tokens, and adaptive thinking that’s always on, with depth steered through the effort parameter (default: high). The reliable knowledge cutoff moves to June 2026.

For teams already calling Fable 5, Anthropic’s what’s-new documentation for Fable 5.1 splits the changes cleanly:

Breaking (three):

• Forced tool use (tool_choice set to any or a named tool) now returns a 400 error.
• Earlier Claude models can’t read Fable 5.1’s thinking blocks, though Fable 5.1 reads theirs.
• Editing earlier conversation turns invalidates thinking blocks, enforced for API accounts created on or after August 31, 2026.

Additive (five):

• Per-message effort changes mid-conversation without breaking the prompt cache (beta).
• Turn-scoped system messages that expire after the next user turn (beta).
• Readable progress updates between tool calls via thinking.display: "updates" (beta).
• Cache reads at $0.25 per million tokens, down from $1.
• Content provenance: a statistical text watermark on all output, plus C2PA credentials on generated media.

One naming note worth clearing up early. Claude Mythos 5.1 is the same model as Fable 5.1, per Anthropic’s own docs, with different safeguard configurations. It’s invite-only, restricted to participants in Project Glasswing, Anthropic’s critical-infrastructure security program. Everything below about specs, pricing, and behavior applies to both. Access is the entire difference.

Each of these changes gets a full section later in this guide, with the migration mechanics spelled out for anyone running Fable 5 or Opus 5 in production today.

What Does Claude Fable 5.1 Actually Cost?

Claude Fable 5.1 costs $10 per million input tokens and $50 per million output tokens, unchanged from Fable 5, and the real story is cache reads at $0.25 per million: a quarter of Fable 5’s rate. Here’s the full 2026 rate card:

That cache-read number deserves a closer look. On Fable 5.1, a cache hit costs 2.5% of the base input price. Every other current Claude model, Opus 5 and Sonnet 5 included, charges the standard 10% multiplier. Anthropic priced this deliberately: long agentic sessions re-read the same cached prefix over and over, so the read rate dominates real-world spend far more than the sticker price suggests. By Anthropic’s own 2026 estimates, the cheaper cache reads alone cut typical workload costs by roughly 25%, and heavily agentic or coding workloads by up to about 45% versus Fable 5.

Anyone who’s watched an agent harness burn budget knows the pattern: the model writes 2,000 tokens and re-reads 400,000. That asymmetry is exactly what this pricing targets.

Two more billing details matter in practice. On AWS, Fable 5.1 bills through AWS Marketplace in Claude Consumption Units, converted at $0.01 per CCU. And the 512-token minimum cacheable prompt length carries over unchanged, so very short prompts still can’t cache at all.

On claude.ai, plan availability for Fable-class models splits by tier: Pro, Max, Team, and Enterprise get access, Free doesn’t. Max-tier and premium seats can spend up to 50% of weekly usage limits on Fable models at no extra cost. Pro-tier and standard Team or Enterprise seats pay through usage credits billed at standard API rates.

How Claude Fable 5.1 Compares With Claude Opus 5, Sonnet 5, and Haiku 4.5

Claude Fable 5.1 sits at the top of Anthropic’s 2026 lineup on capability and price, and at the bottom on speed. The spec sheet makes the trade explicit:

Opus 5 matches Fable 5.1 on context and output ceiling at half the token price. That’s why Anthropic’s own guidance says to default to Opus 5 and step up only when your evals demand it. Refreshing candor from a vendor, honestly.

Does the premium buy anything? The launch benchmarks say yes. Per Anthropic’s September 2026 launch announcement, Fable 5.1 beats both its predecessor and Opus 5 across most measured evaluations:

The Terminal-Bench-Science jump is the eye-opener: 52.6% against Fable 5’s 24.7%, better than double. AutomationBench nearly doubles too. Coding gains are real but tighter, and Opus 5 stays within about three points on Terminal-Bench 4.0, which reinforces the start-with-Opus advice for everyday engineering work.

One efficiency claim stands out for budget planning: Anthropic reports that Fable 5.1 at low or medium effort matches or beats Fable 5 at full effort, at meaningfully lower cost. If that holds on your workloads, the upgrade can pay for itself before you ever touch the higher effort levels.

Where Claude Fable 5.1 Improves: Agentic Coding, Research, and Document Work

Claude Fable 5.1’s gains over Fable 5 concentrate in six documented areas, and the gap between the two models is widest at higher effort levels. Per Anthropic’s Fable 5.1 model overview, the improvements land where long-horizon work actually happens:

• Agentic coding over long sessions: multi-file features, large refactors and migrations, debugging, and code review across sessions that run for hours.
• Knowledge work with documents, spreadsheets, and slides: taking an analysis from a first question to a finished document, a live-formula spreadsheet, or a slide deck built from a blank page.
• Research and search: higher accuracy on multistep web research and deep-research tasks that follow up on what they find.
• Vision: reading dense charts, filings, and tables nested inside PDFs, including with crop-and-zoom tools on charts.
• Long-context reasoning: connecting details across the full 1M-token window instead of just retrieving from it.
• Computer use: operating a browser and desktop apps more reliably, and recovering from failed steps.

That last clause matters more than it sounds. Recovering from a failed step is the difference between an agent that finishes a task and one that stalls at the first mispositioned click. Anyone who’s shipped a production agent knows the failure mode: everything works until step 14, then the run dies quietly.

Two caveats keep expectations honest. The gains scale with effort, so a team running everything at low effort to save money will see less of the improvement than the launch benchmarks suggest. And multilingual performance is on par with Fable 5: if your workload is heavily non-English, the 5.1 upgrade buys you the six areas above and nothing extra on language coverage.

For teams building agents in these exact areas, the model upgrade is only half the work. The harness around it, the tool design, the context strategy, decides whether those capability gains reach production. That’s the kind of gap an AI integration audit is built to find before the migration, instead of three weeks after.

The Three Breaking Changes for Existing Claude Fable 5 Integrations

Three changes in Claude Fable 5.1 will break existing Fable 5 code: forced tool use now returns a 400 error, thinking blocks are bound one-way to the model that wrote them, and editing earlier conversation turns invalidates every thinking block that follows. Each has a documented replacement.

1. Forced tool use returns a 400 error. Setting tool_choice to {"type": "any"} or {"type": "tool", "name": "..."} fails with an invalid_request_error on the Messages API, the Batches API, and even the token counting endpoint. The reasoning is sound: thinking is always on, and a forced call would skip it, pushing the model’s working-out into the tool arguments and degrading their quality. The replacements:

• For schema-valid JSON: keep tool_choice: auto and set strict: true on the tool, or move the schema to structured outputs.
• To make the model call a tool: name it in the prompt (“Use the get_weather tool to answer”). Fable 5.1 follows explicit tool instructions reliably.
• When your application requires a call on a specific turn: append a mid-conversation system message naming the tool, which keeps earlier turns byte-identical and cache-warm.

tool_choice: none still works for turns that must skip tools.

2. Thinking blocks are model-bound, one way. Fable 5.1 reads thinking blocks from Opus 5, Fable 5, and every earlier Claude model, so upgrading a live conversation keeps its reasoning. The reverse fails: no earlier model reads Fable 5.1’s blocks. When a router, retry, or refusal fallback moves a conversation to an older model, the API drops the unreadable blocks before the model sees them, unbilled, and the target re-plans without that reasoning. Without the thinking-binding-controls-2026-08-01 beta header, the drop is silent. With it, an input_transformations array names every dropped block.

3. Editing history invalidates thinking blocks. Each Fable 5.1 thinking block is valid only against the exact system prompt, tools, and messages that preceded it. Rebuilding the system prompt between requests, injecting a per-turn reminder you delete later, reordering turns, even a document URL that serves different bytes on replay: each one invalidates every later block. Append-only histories, server-side compaction, cache_control moves, and effort changes all stay safe.

Enforcement is the trap here. Accounts created on or after August 31, 2026 get a hard 400. Older accounts sail through unless they opt in, which means a tool vendor testing on a 2024 account ships code their newer-account users watch fail on day one.

New Beta Features: Per-Message Effort, Turn-Scoped System Messages, and Progress Updates

Claude Fable 5.1’s additive features are three beta APIs aimed at long-running agent sessions, plus content provenance that ships on by default. None requires code changes to adopt the model; each removes a real cost or visibility problem once you do.

Per-message effort (beta header mid-conversation-output-config-2026-07-01) lets you change the effort level mid-conversation without invalidating the prompt cache. An effort-only role: "system" message carrying output_config takes effect from the next user turn:

{"role": "system", "content": [], "output_config": {"effort": "low"}}

On Fable 5, changing effort between requests dropped every cached prefix. Now you can run routine turns at low and spike to xhigh for the hard step, and the cache keeps matching. All five levels are accepted: low, medium, high, xhigh, max.

Turn-scoped system messages (beta header mid-conversation-system-clear-at-2026-08-21) solve the inject-and-delete pattern that now breaks thinking blocks. Set clear_at: "next_user_message" on a system message and it carries system-prompt authority for one turn, then stops rendering once a later user message exists. You keep sending it back verbatim, so history stays append-only, the cache stays warm, and a cleared message costs zero input tokens. Built for per-turn agent-loop reminders like “check your inbox before running more code.”

Progress updates as readable text (beta header thinking-display-updates-2026-08-18) fix the silent-agent problem. Fable 5.1 writes short status updates between tool calls, but under the default thinking.display of "omitted" those blocks come back empty, so a long agentic turn looks dead to your users. Set display: "updates" and any thinking block with non-empty text becomes a status line you can render, while raw reasoning stays hidden.

Content provenance rounds out the set, and it’s live everywhere without a header. Every piece of Fable 5.1 and Mythos 5.1 text carries Anthropic’s statistical watermark, adding no tokens and no hidden characters, and generated images and video carry signed C2PA Content Credentials when retrieved through the Files API.

Behavior Differences That Show Up Without Any Code Change

Claude Fable 5.1 behaves differently from Fable 5 in seven documented ways even when your integration compiles and runs untouched. None of these throws an error. They surface as changed output, changed cost, or changed pacing, which makes them easy to miss in a quick smoke test and expensive to discover in production.

The drifts Anthropic documents, each with a prompting fix:

• More variable parallel tool calling: the model may issue one tool call per turn where Fable 5 batched several, adding round trips and wall-clock time in long agent loops without hurting answer quality. A one-line batching instruction restores the old behavior.
• Fewer progress updates during long tool runs: less narration between calls, especially at higher effort. Ask explicitly for an opening line, periodic updates, and a recap if your UI depends on it.
• Memory-first answers at low effort: the model calls search and retrieval tools less often at the lowest effort level. Raise effort for turns that need fresh data, or add a verification nudge.
• Denser prose: longer sentences, fewer paragraph breaks in places.
• Less chat formatting: fewer bold phrases, headers, and lists, so anti-formatting rules written for older Claude models can now suppress structure the content genuinely needs.
• Unmarked quotations in summaries: the model reproduces source passages more often without flagging them as quotes.
• Whole-file rewrites for small edits: same end result, more output tokens and time.

That last one stings at $50 per million output tokens. A ten-line change that rewrites a 2,000-line file is real money at scale.

What carries over from Fable 5 unchanged is just as useful to know. Prefilling the assistant response still returns a 400. So do non-default temperature, top_p, or top_k values. The 512-token minimum cacheable prompt length holds, thinking.display still defaults to "omitted", and interleaved thinking stays automatic with no beta header.

How to Migrate to Claude Fable 5.1 From Claude Fable 5, Claude Opus 5, or Earlier

Migrating to Claude Fable 5.1 is a model-ID swap plus a checklist whose length depends on where you’re coming from: shortest from Fable 5, longer from Opus 5, longest from Opus 4.8 or earlier. Anthropic ships an automated path too. In Claude Code, /claude-api migrate this project to claude-fable-5-1 applies the ID swap, breaking parameter changes, prefill replacement, and effort calibration across your codebase, then hands you a manual verification list. It detects Bedrock and Claude Platform on AWS clients and adjusts model-ID formats for those platforms.

From Fable 5, the move is mostly drop-in. Per Anthropic’s Fable 5.1 migration guide, the API surface, pricing, tokenizer, and refusal categories all match. The work concentrates in four checks:

1. Replace forced tool_choice with auto plus an explicit instruction and strict: true tools, or JSON outputs.
2. Run the history-editing audit: a multi-turn session with the thinking-binding beta header and prefix_mismatch_behavior: "drop_block", logging input_transformations on every response. Empty arrays mean your history is intact. Any prefix_binding_mismatch entry means something edited the prefix.
3. Pick a production mismatch setting ("error" or "drop_block") and monitor it.
4. Re-sweep effort on your own evals instead of carrying over a Fable 5 setting. Gains are largest at xhigh and max, but those levels add thinking time.

From Opus 5, add five deltas: thinking: {type: "disabled"} now returns a 400 at any effort level, narration between tool calls arrives in thinking blocks instead of text blocks, refusal categories broaden beyond Opus 5’s cybersecurity-only classifiers (expect "bio" and "reasoning_extraction"), token prices double to $10/$50, and zero data retention eligibility goes away unless Anthropic expressly authorizes it. Integrations written for Opus 4.8 or earlier often rebuild the system prompt or truncate turns on every request, habits those models never punished, so budget extra time for the history audit there.

From Mythos 5, everything above applies except the history-editing check, which Mythos 5.1 doesn’t run. This is exactly the kind of migration where a production agent harness either survives quietly or fails loudly, and the audit tells you which before your users do.

Availability, Data Retention, and the Claude Mythos 5.1 Access Program

Claude Fable 5.1 launched on every major platform simultaneously on September 1, 2026, while Claude Mythos 5.1 remains invite-only through Project Glasswing. No access approval is needed for Fable 5.1 anywhere it ships:

• Claude API: all customers, as claude-fable-5-1
• AWS: Amazon Bedrock (anthropic.claude-fable-5-1) and Claude Platform on AWS
• Google Cloud: Vertex AI, as claude-fable-5-1
• Microsoft Foundry: on Anthropic infrastructure
• End-user products: claude.ai, Claude Code (defaulting to high effort), and Claude Cowork (defaulting to medium)

Anthropic commits to no retirement before September 1, 2027, a one-year floor worth noting in any procurement conversation.

The compliance fine print bites harder than the platform list. Both models carry mandatory 30-day data retention and are designated Covered Models; a request from an organization without 30-day retention returns a 400 error, and zero data retention requires express authorization from Anthropic. Neither model supports Priority Tier, which Fable 5 does. Refused requests can fall back automatically, with Claude Opus 4.8 and Claude Opus 5 as the only permitted targets.

Mythos 5.1 is where the naming story pays off. Same model, same specs, same pricing, with access gated to vetted defenders hardening critical software. The program has numbers behind it: Anthropic’s initial Project Glasswing update reports that roughly 50 partners used Claude Mythos Preview to find more than 10,000 high- or critical-severity vulnerabilities, including 6,202 flaws across 1,000+ open-source projects with a 90.6% independently confirmed true-positive rate. Anthropic has since announced expansion to roughly 150 additional organizations across more than 15 countries, spanning power, water, healthcare, and communications infrastructure. Access runs through Anthropic, AWS, or Google Cloud account teams.

Frequently Asked Questions About Claude Fable 5.1

These are the questions engineers and buyers ask most often about the Claude Fable 5.1 launch, with the short answers up front.

How much does Claude Fable 5.1 cost?

Claude Fable 5.1 costs $10 per million input tokens and $50 per million output tokens as of September 2026, identical to Fable 5. Cache reads cost $0.25 per million tokens, a quarter of the Fable 5 rate, and the Batch API halves input and output prices to $5/$25. Cache writes run $12.50 per million for five-minute entries and $20 for one-hour entries.

Is Claude Fable 5.1 better than Claude Opus 5?

On Anthropic’s September 2026 launch benchmarks, yes: Fable 5.1 leads Opus 5 on every reported evaluation, including 55.8% versus 52.3% on Terminal-Bench 4.0 and 52.6% versus 29.0% on Terminal-Bench-Science. It also costs twice as much per token and runs slower. Anthropic’s own guidance is to start with Opus 5 and move up only when your evals at higher effort still fall short.

What is Claude Mythos 5.1?

Claude Mythos 5.1 is the same model as Claude Fable 5.1 with different safeguard configurations, offered by invitation only to participants in Project Glasswing, Anthropic’s critical-infrastructure security program. It shares Fable 5.1’s specs and pricing. Access runs through an Anthropic, AWS, or Google Cloud account team.

Can I disable thinking on Claude Fable 5.1?

No. Adaptive thinking is always on, and both thinking: {"type": "disabled"} and manual budgets via budget_tokens return a 400 error. You control depth through the effort parameter instead, across five levels from low to max, and in beta you can change it mid-conversation without losing the prompt cache.

Why does tool_choice return a 400 error on Claude Fable 5.1?

Forced tool use is removed because a forced call would skip the model’s always-on thinking, pushing its working-out into the tool arguments and degrading their quality. Setting tool_choice to any or a named tool fails on the Messages API, the Batches API, and the token counting endpoint. Use auto with strict: true tools and an explicit instruction, or structured outputs.

Is Claude Fable 5.1 available on the claude.ai Free plan?

No. As of September 2026, Fable-class models require a Pro, Max, Team, or Enterprise plan. Max-tier and premium seats can spend up to 50% of weekly usage limits on Fable models at no extra cost, while Pro-tier and standard seats pay through usage credits at standard API rates.

What is the knowledge cutoff for Claude Fable 5.1?

Both the reliable knowledge cutoff and the training data cutoff are June 2026, the most recent in Anthropic’s lineup. Opus 5 sits at May 2026, Sonnet 5 at January 2026, and Haiku 4.5 back at February 2025.

Will Claude Fable 5.1 be retired soon?

No sooner than September 1, 2027. Anthropic commits to at least a one-year availability window from the September 1, 2026 release date, which gives production teams a firm planning floor.

Where to Start With Claude Fable 5.1

Start with an eval run, and let the numbers decide whether the Claude Fable 5.1 launch pricing and benchmarks justify the premium for your workload. Run your suite on Opus 5 at higher effort first. If it clears the bar, you just saved half your token budget. If it doesn’t, swap the model ID, work through the migration checklist (forced tool_choice, append-only history, the input_transformations audit), and sweep effort levels fresh instead of carrying over old settings.

Then pilot before you commit. Route a slice of traffic, watch for prefix mismatches and one-call-per-turn agent loops, and re-baseline cost with the $0.25 cache reads factored in. The quarter-price cache is where most agentic workloads will feel this release.

If you’d rather have someone who’s shipped production agents run that evaluation with you, talk to AlphaCorp AI. The eval harness you build this week outlasts this model and the next one.