Anthropic shipped Claude Fable 5.1 on 1 September 2026 as a point release on top of Claude Fable 5. The headline price didn't move, the context window didn't move, and the model ID barely moved — a combination that makes it easy to file the release under "minor" and move on.
It isn't minor. The most consequential change is buried in a pricing footnote, and there's a genuine breaking change that returns a 400 error on existing agent code. This guide covers both, with every number traced back to Anthropic's own documentation.
What is Claude Fable 5.1?
Claude Fable 5.1 is the current top of Anthropic's Claude lineup — the model the docs point you to "for demanding reasoning and long-horizon agentic work." Anthropic is explicit that it is not the default recommendation: the models overview tells you to start with Claude Opus 5 and reach for Fable 5.1 "when your evals on Claude Opus 5 at higher effort still fall short." A sibling model, Claude Mythos 5.1, shares its exact specifications and pricing but is restricted to Anthropic's invitation-only Project Glasswing.
| Specification | Claude Fable 5.1 |
|---|---|
| Claude API model ID | claude-fable-5-1 |
| Released | 1 September 2026 |
| Context window | 1,000,000 tokens (default and maximum) |
| Max output | 128,000 tokens |
| Input / output | Text and images in, text out |
| Thinking | Adaptive, always on |
| Default effort | high |
| Reliable knowledge cutoff | June 2026 |
| Training data cutoff | June 2026 |
| Comparative latency | Slower (slowest in the current lineup) |
| Retirement commitment | Not sooner than 1 September 2027 |
| Platforms | Claude API, Amazon Bedrock (anthropic.claude-fable-5-1), Google Cloud, Microsoft Foundry, Claude Platform on AWS |
Two details in that table matter more than they look. The 1M-token context window is billed at standard per-token rates across the whole window — a 900k-token request costs the same per token as a 9k-token one, with no long-context premium. And Fable 5.1 uses the tokenizer introduced with Claude Opus 4.7, which produces roughly 30% more tokens for the same text than the pre-4.7 tokenizer. If you are comparing costs against Claude Sonnet 4.6 or anything older, the token counts are not like-for-like.
What changed from Claude Fable 5?
Anthropic groups the changes into three breaking and five additive. The breaking ones are the reason you should not treat this as a drop-in string swap. For where Fable 5 started, see our Claude Fable 5 launch guide.
Forced tool use now returns an error
This is the one most likely to break production code. On Fable 5.1, tool_choice set to {"type": "any"} or {"type": "tool", "name": "..."} returns a 400 invalid_request_error:
tool_choice: type "tool" and "any" are not supported for this model.tool_choice: {"type": "auto"} (the default) and {"type": "none"} still work. The same validation applies to the token-counting endpoint, so you can't even price a request that uses forced tool calls.
The reasoning is coherent: thinking is always on for this model, and a forced tool call would skip it — the model would write its working-out into the tool arguments instead, which degrades argument quality. Anthropic's recommended replacements are to keep tool_choice: {"type": "auto"} and set strict: true with strict tool use, move the schema to structured outputs, or simply instruct the model in the prompt ("Use the get_weather tool to answer"). The docs claim Fable 5.1 follows explicit tool instructions reliably.
Thinking blocks are bound to the model and to the conversation prefix
Every thinking block now records which model produced it, and preservation runs one way only. Fable 5.1 can read earlier models' thinking blocks; no earlier model can read Fable 5.1's. A router or fallback that switches models mid-conversation silently drops those blocks — they aren't billed and don't count toward input_tokens, but the reasoning is gone for turns that run elsewhere.
Separately, modifying anything before a Fable 5.1 thinking block — the system prompt, the tools array, or an earlier message — invalidates every thinking block after it and errors on the next request, enforced for accounts created on or after 31 August 2026. Treat the conversation as append-only: injecting a per-request reminder into history and deleting it next turn is exactly the pattern that now breaks. Claude Code, claude.ai, Claude Managed Agents and the Claude Agent SDK handle this for you; if your code builds the messages array by hand, this is the migration item to budget time for.
Cheaper cache reads, and three new beta features
The additive changes are the cache-read price cut (covered in detail below), plus three betas worth knowing about:
- Per-message effort (
mid-conversation-output-config-2026-07-01) — change the effort level partway through a conversation without invalidating the prompt cache. Raise it for a hard step, drop it for routine ones. - Turn-scoped system messages (
mid-conversation-system-clear-at-2026-08-21) — setclear_at: "next_user_message"on a system message so it carries system-prompt authority for one turn, then stops rendering. It stays inmessages, so history remains append-only and the cache keeps matching. A cleared message costs no input tokens. - Readable progress updates (
thinking-display-updates-2026-08-18) — setthinking.displayto"updates"to receive between-tool-call status lines as text while raw reasoning stays hidden.
Anthropic also documents a set of behaviour changes that need no code change but will show up in your evals. Fable 5.1 batches parallel tool calls less consistently than Fable 5 (sometimes one call per turn where Fable 5 issued several), writes fewer progress updates, answers from memory more often at low effort instead of searching, writes denser prose with less formatting, is more likely to reproduce source passages in summaries without marking them as quotations, and is more likely to rewrite a whole file when a targeted edit would do. None of those reduce answer quality, but several cost tokens and wall-clock time.
What does Claude Fable 5.1 actually cost?
Here is the full price sheet, from Anthropic's pricing page.
| Token type | Claude Fable 5.1 | Claude Fable 5 | Claude Opus 5 |
|---|---|---|---|
| Base input | $10 / MTok | $10 / MTok | $5 / MTok |
| Output | $50 / MTok | $50 / MTok | $25 / MTok |
| 5-minute cache write | $12.50 / MTok | $12.50 / MTok | $6.25 / MTok |
| 1-hour cache write | $20 / MTok | $20 / MTok | $10 / MTok |
| Cache read (hit) | $0.25 / MTok | $1.00 / MTok | $0.50 / MTok |
| Batch API | $5 in / $25 out | $5 in / $25 out | $2.50 in / $12.50 out |
The cache-read row is the whole story. Every other Claude model prices a cache hit at 0.1x the base input price. Anthropic's pricing table carries a footnote saying cache hits and refreshes on Fable 5.1 and Mythos 5.1 are priced at 0.025x instead. Ten percent of $10 is the $1.00 Fable 5 charges; 2.5% of $10 is $0.25. That is a 75% reduction on cache reads with no change to the sticker price — and because it lives in a multiplier footnote rather than a headline, it is easy to miss entirely.
A worked example on a real agent workload
Take a coding agent with a 200,000-token cached prefix (system prompt, repo map, project conventions, a few reference files), running a 40-turn session. Each turn re-reads the cached prefix, adds ~2,000 tokens of new input, and produces ~1,500 output tokens. Use the 1-hour cache so the prefix survives the session.
| Line item | Tokens | Fable 5 | Fable 5.1 |
|---|---|---|---|
| 1h cache write (once) | 200,000 | $4.00 | $4.00 |
| Cache reads (40 x 200,000) | 8,000,000 | $8.00 | $2.00 |
| Fresh input (40 x 2,000) | 80,000 | $0.80 | $0.80 |
| Output (40 x 1,500) | 60,000 | $3.00 | $3.00 |
| Session total | $15.80 | $9.80 |
That is a 38% cut on identical work, purely from the cache-read line. It also matches what Anthropic claims in the announcement: roughly 25% less than Fable 5 for typical workloads, with savings "up to approximately 45%" on agentic work. The more cache-heavy your workload, the closer you sit to the top of that range.
An honest caveat: at $10 in and $50 out, this is still a premium-tier model, and the discount only applies to cache reads. Writes are unchanged, and the minimum cacheable prompt is 512 tokens. If your workload is short one-shot calls with no reusable prefix, Fable 5.1 costs you exactly what Fable 5 did — twice Opus 5's rate for both input and output. Artificial Analysis puts Fable 5.1's blended price at $7.17 per million tokens and notes it is "particularly expensive when comparing to other models of similar price."
How good is Claude Fable 5.1?
Worth separating what Anthropic reports from what third parties measure.
Vendor benchmarks
From Anthropic's own launch announcement:
| 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% |
| AutomationBench | 31.4% | 17.1% | 26.9% | 19.6% |
| 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% | — |
| CursorBench 3.2.0 | 73.4% | 70.5% | 70.0% | 67.2% |
| GDPval-AA v2 (knowledge work) | 1853 | 1723 | 1824 | 1711 |
Terminal-Bench-Science more than doubles, which is the single biggest jump on the board. The general-purpose numbers are more incremental: CursorBench moves 70.5% to 73.4%, and Humanity's Last Exam without tools moves 57.8% to 60.9%. Notably, Anthropic did not publish SWE-bench Verified, GPQA Diamond, ARC-AGI or tau-bench figures for this release, so anyone quoting those for Fable 5.1 is not quoting Anthropic.
Independent measurement
Artificial Analysis scores Fable 5.1 at 66 on its Intelligence Index — first out of 196 models evaluated — against 63.1 for Claude Opus 5 and 62.1 for Claude Fable 5 on the same index. Its coding index sits at 81.6 and its agentic index at 61.3, both up from Fable 5's 76.5 and 56.6.
The same evaluation surfaces the trade-off. Output runs at about 66 tokens per second, and time-to-first-token measures roughly 285 seconds against a 3.35-second median across models. That is what "adaptive thinking, always on, default effort high" costs you in practice: this model thinks before it says anything, and on hard prompts it thinks for minutes. Anthropic's own comparison table simply lists Fable 5.1's latency as "Slower" — the slowest in the current lineup. For a batch job or a background agent that is irrelevant. For anything a human is waiting on, it is disqualifying.
How does Claude Fable 5.1 compare with Claude Opus 5?
Short version: Fable 5.1 wins on capability, Opus 5 wins on price and speed, and Anthropic itself recommends Opus 5 as the starting point.
Both have a 1M-token context window and 128K max output. Fable 5.1 leads on every published benchmark and on the Artificial Analysis index. But Opus 5 is exactly half the price on input, output and cache writes, has "Moderate" rather than "Slower" latency, supports forced tool use and a non-default temperature, and can hit 300K output tokens on the Batch API with the output-300k-2026-03-24 beta header — a capability Fable 5.1 does not have. Opus 5 also supports Fast mode, priced at $10/$50, which is a different way of spending the same money.
The decision rule Anthropic gives is a good one: run your evals on Opus 5 at a higher effort level first, and only move to Fable 5.1 if they still fall short. For the full head-to-head, see Claude Fable 5.1 vs Opus 5, and our Claude Opus 5 launch guide for Opus 5's own specs and benchmarks.
Is Claude Fable 5.1 available on Claude Pro and Max, or API only?
It is available on the consumer plans, not just the API. Anthropic's pricing page lists Fable alongside Opus, Sonnet and Haiku on Free, Pro ($17/month annual, $20 monthly) and both Max tiers (from $100/month), as well as Team and Enterprise.
The catch is in the fine print. The pricing page attaches a "50% of weekly limits" note to Fable usage on the consumer tiers — Fable burns your weekly allowance roughly twice as fast as the other models. And in Claude Code, Fable is not the default on any plan; it must be explicitly selected, and depending on your plan and seat tier, Fable usage may bill to usage credits rather than your plan's included limits. If you switch a Pro or Max subscription to Fable 5.1 and leave it there, expect to hit limits sooner or start drawing down credits. Our guide to Claude Fable usage credits and plan limits covers how that accounting works.
One more constraint that matters for regulated buyers: Fable 5.1 and Mythos 5.1 carry 30-day data retention and are not available under zero data retention unless Anthropic expressly authorises it. Both are Covered Models, as Fable 5 and Mythos 5 were.
How do you use Claude Fable 5.1?
The model string is claude-fable-5-1 on the Claude API, Google Cloud, Microsoft Foundry and Claude Platform on AWS, and anthropic.claude-fable-5-1 on Amazon Bedrock. Migrating from Fable 5 is a one-line change plus the tool_choice audit described above.
curl https://api.anthropic.com/v1/messages \
-H "x-api-key: $ANTHROPIC_API_KEY" \
-H "anthropic-version: 2023-06-01" \
-H "content-type: application/json" \
-d '{
"model": "claude-fable-5-1",
"max_tokens": 4096,
"messages": [
{"role": "user", "content": "Summarise this repo migration plan in three steps."}
]
}'In Python, with prompt caching switched on — which is the point of this release:
import anthropic
client = anthropic.Anthropic()
response = client.messages.create(
model="claude-fable-5-1",
max_tokens=4096,
system=[
{
"type": "text",
"text": LARGE_SYSTEM_PROMPT, # must be >= 512 tokens to be cacheable
"cache_control": {"type": "ephemeral", "ttl": "1h"},
}
],
messages=[{"role": "user", "content": "What changed in the auth module this week?"}],
)
print(response.usage.cache_read_input_tokens) # these bill at $0.25 / MTok
print(response.usage.cache_creation_input_tokens)Three things to get right if you want the saving:
- Meet the 512-token minimum. Anything shorter is not cacheable, and you pay full input price.
- Pick the right TTL. The 5-minute write costs 1.25x base input, the 1-hour write costs 2x. At 0.025x reads, a 5-minute cache pays for itself after a single hit and a 1-hour cache after two. Both break even far faster than they did on Fable 5.
- Keep history append-only. Editing an earlier turn invalidates thinking blocks and breaks the cache prefix. Use turn-scoped system messages for per-turn reminders and server-side context editing or compaction for trimming, not client-side surgery on
messages.
In Claude Code, Fable 5.1 is the model behind the fable alias. Select it with /model fable, claude --model fable, claude --model claude-fable-5-1, or export ANTHROPIC_MODEL=fable. Claude Code v2.1.255 or later is required for Fable 5.1. If you're weighing this against other agent harnesses, our complete guide to AI coding agents covers the wider landscape.
Should you upgrade from Claude Fable 5?
If you already run Fable 5 with prompt caching, upgrade — the same money buys the same tokens at higher benchmark scores, and cache-heavy sessions get materially cheaper. Budget an afternoon for the migration: strip any tool_choice of type any or tool, verify your code never edits earlier turns, then re-run your evals and re-tune effort from the high default.
If you run Fable 5 without caching, the upgrade is free but the cost story doesn't apply to you. Fix the caching first; that's where the money is, on either model.
And if you are not on Fable at all, the honest answer is that Anthropic doesn't think you should be — its documentation points you at Opus 5 at half the price and better latency. Fable 5.1 is the right tool for long-horizon agentic work where a minutes-long first token is acceptable and correctness is worth double the rate. It is the wrong tool for anything interactive. Prove you need it before you pay for it.
FAQ
What is Claude Fable 5.1?
Claude Fable 5.1 is Anthropic's most capable Claude model, released on 1 September 2026 as a successor to Claude Fable 5. It has a 1M-token context window, 128K max output, always-on adaptive thinking, and a June 2026 knowledge cutoff. Anthropic positions it for demanding reasoning and long-horizon agentic work rather than as a general default.
How much does Claude Fable 5.1 cost?
$10 per million input tokens and $50 per million output tokens on the Claude API — identical to Fable 5. Cache writes are $12.50/MTok for 5-minute and $20/MTok for 1-hour. Cache reads are $0.25/MTok. The Batch API halves input and output to $5 and $25 per million tokens.
What's new in Fable 5.1 versus Fable 5?
Cache reads dropped 75% to $0.25 per million tokens, and benchmarks improved across agentic coding, research, vision and computer use. Three changes are breaking: forced tool use now errors, thinking blocks are bound to the producing model, and editing earlier turns invalidates them. Per-message effort, turn-scoped system messages and readable progress updates are new betas.
Is Fable 5.1 better than Opus 5?
On capability, yes — it leads Opus 5 on every benchmark Anthropic published and scores 66 versus 63.1 on the Artificial Analysis Intelligence Index. But it costs twice as much, is slower, and drops forced tool use. Anthropic recommends starting with Opus 5 and moving up only if your evals at higher effort still fall short.
Is Claude Fable 5.1 available on the Pro plan?
Yes. Anthropic's pricing page lists Fable on Free, Pro, both Max tiers, Team and Enterprise. However, Fable usage counts against roughly 50% of your weekly limits, it is never the default model on any plan, and in Claude Code it may bill to usage credits rather than included limits depending on your plan and seat tier.
What is Fable 5.1's context window?
1,000,000 tokens, which is both the default and the maximum. The entire window is billed at standard per-token rates with no long-context premium. Maximum output is 128,000 tokens; unlike Opus 5 and Sonnet 5, Fable 5.1 does not support the 300K-output batch beta. Roughly 555k English words fit in 1M tokens on the current tokenizer.
Should I upgrade from Fable 5?
Yes if you use prompt caching, since identical workloads get 25-45% cheaper at the same headline price. Before switching, remove any tool_choice of type any or tool (it now returns a 400 error), make sure your code never edits earlier conversation turns, and re-run your evals — several default behaviours changed.
Is Claude Fable 5 deprecated?
Not deprecated, but reclassified. Anthropic's models overview lists Claude Fable 5 under "Legacy models (still available)" alongside Opus 4.8 and Sonnet 4.6. It continues to serve requests at $10/$50 with $1/MTok cache reads. Fable 5.1 carries a retirement commitment of not sooner than 1 September 2027.