Anthropic shipped Claude Fable 5.1 and Claude Mythos 5.1 on September 1, 2026, and the coverage is leading with "up to 45% cheaper." That number is real but it is easy to misread. The base price did not move at all.
Fable 5.1 still costs $10 per million input tokens and $50 per million output tokens, exactly what Fable 5 cost. What changed is one line on the pricing table: cache reads dropped from $1 to $0.25 per million tokens, a 4x cut. Anthropic's own framing is a 25% saving on typical workloads and up to 45% on heavily agentic ones, and both figures come entirely from that single line.
So whether you save anything depends on how much of your input is cached. If you send fresh prompts every time, your bill is identical.
Table of contents
- What actually changed on price
- Whether you see the saving depends on your cache hit rate
- The benchmark numbers
- Terminal-Bench-Science more than doubled
- Three breaking changes if you already call Fable 5
- Effort levels are the real cost lever
- What Mythos 5.1 is
- Should you switch?
- FAQ
- Sources
What actually changed on price
Here is the full comparison from Anthropic's pricing table:
| Fable 5 | Fable 5.1 | |
|---|---|---|
| Base input | $10 / MTok | $10 / MTok |
| 5m cache write | $12.50 / MTok | $12.50 / MTok |
| 1h cache write | $20 / MTok | $20 / MTok |
| Cache read | $1 / MTok | $0.25 / MTok |
| Output | $50 / MTok | $50 / MTok |
One row moved. Cache reads on Fable 5.1 are priced at 0.025x the base input price, where every other Claude model uses the standard 0.1x. That is the whole pricing story.
Context window stays at 1M tokens, max output at 128K.
Whether you see the saving depends on your cache hit rate
This is the part worth doing arithmetic on before you plan around a 45% cut.
Take a workload with 1M input tokens and 200K output tokens per run. Output cost is fixed at $10 either way. On the input side:
- No caching at all: 1M fresh tokens = $10 on both models. Saving: zero.
- Half cached: 500K fresh ($5) + 500K cached. On Fable 5 that is $0.50, on 5.1 it is $0.125. Total run: $15.50 versus $15.13, a saving of 38 cents or 2.4%.
- 90% cached: 100K fresh ($1) + 900K cached. Fable 5: $0.90 on the cached part. Fable 5.1: $0.23. Total run: $11.90 versus $11.23, a saving of 5.7%.
Small numbers, because output dominates at $50/MTok. The picture inverts when output is small and cached input is enormous, which is exactly what long-running agentic loops look like: the same large system prompt and codebase context replayed across dozens of turns, each turn emitting a short tool call.
Push it to that extreme and the numbers flip hard. Two million cached input tokens against a 5K output costs $2.25 on Fable 5 and $0.75 on Fable 5.1, a 67% cut. Anthropic's 45% sits between that and the modest single-digit savings above, which is why it reads as a claim about a specific shape of workload rather than a general discount. Our API cost calculator now carries both models, so you can put your own token mix in rather than trusting a headline.
The benchmark numbers
Anthropic published these with production safeguards enabled, and notes that on tasks where safeguards intervened the model scored zero, which likely drags the figures down.
| 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 (strict) | 41.7% | 36.1% | 39.6% | — |
| Humanity's Last Exam (no tools) | 60.9% | 57.8% | 56.6% | — |
| AutomationBench | 31.4% | 17.1% | 26.9% | 19.6% |
| CursorBench 3.2.0 | 73.4% | 70.5% | 70.0% | 67.2% |
Mythos 5.1 scores 60.9% on Terminal-Bench 4.0 against Fable 5.1's 55.8%. Anthropic explains that gap as its older, less precise cyber safeguards intervening on some tasks, and expects the difference to shrink now that those safeguards have been improved.
Note what the table does not say. On several benchmarks the jump over Opus 5 is a few points, and Opus 5 costs half as much ($5/$25). Anthropic's own documentation says to start with Opus 5 for most workloads and reach for Fable 5.1 only for demanding reasoning and long-horizon agentic work, or when your evals on Opus 5 at higher effort still fall short. That is unusually direct advice against their most expensive model.
Terminal-Bench-Science more than doubled
The one result that is not incremental: 24.7% to 52.6% on Terminal-Bench-Science 0.1, agentic scientific research.
Read the footnote before getting excited. Anthropic states the standard error is ±3.5–4.5 points per model, and that the public leaderboard reports Opus 5 at 30.0% and Fable 5 at 21.4% while their own setup reproduces 29.0% and 24.7%. Both are within noise, but it tells you these numbers move by several points depending on harness.
Even discounting generously, doubling is outside that error band. This is the claim most worth independent verification as people run it.
Three breaking changes if you already call Fable 5
Not a drop-in swap. Three things break:
- Forced tool use returns an error. If your code requires a specific tool call, that request now fails.
- Earlier models cannot read its thinking blocks. Pipelines that hand thinking output to an older Claude will break.
- Editing earlier turns invalidates thinking blocks. Conversation-rewriting patterns need rework.
Five additions, all optional: per-message effort (beta), turn-scoped system messages (beta), readable progress updates between tool calls via display: "updates" (beta), the lower cache read price, and content provenance.
Model IDs are claude-fable-5-1 and claude-mythos-5-1.
Effort levels are the real cost lever
Buried in the announcement and more useful than the price cut: at Low or Medium effort, Fable 5.1 matches or beats Fable 5's results at much lower cost.
Defaults differ by surface, which is worth knowing because it decides your bill:
- Claude Code: High
- Claude Cowork and Claude.ai: Medium
- API: per-message effort control, in beta
If you are on Claude Code and have not thought about effort, you are paying for High on every request. Dropping to Medium for routine work is a larger lever than the cache read change for most people.
What Mythos 5.1 is
Same model, different access. Mythos 5.1 has identical capabilities and identical pricing, available only to Project Glasswing participants by invitation. If you are not in that program, Fable 5.1 is your version.
Anthropic also mentions two things around it: Enterprise Frontier Safeguards, which store data in customer-controlled cloud infrastructure rather than Anthropic's, arriving in phases from later this fall, with zero data retention available to eligible customers until then. And a biology access program built with the US government, enrollment opening soon.
On safeguards they claim 60% fewer false positives in cybersecurity, and note Fable 5.1 can now be used to find software vulnerabilities but not to develop exploits for them.
Should you switch?
Depends on which of these you are:
Running long agentic loops with heavy prompt caching. Yes. The cache read cut is aimed at you, the Terminal-Bench 4.0 and AutomationBench gains are the largest, and the effort controls give you a second lever.
On Opus 5 and happy. Anthropic says stay. Opus 5 is half the price and within a few points on most of this table. Switch only if your own evals on Opus 5 at higher effort fall short.
Sending mostly fresh prompts. The price change does you nothing. Judge purely on capability, and 2x output cost against Opus 5 is a real gap to justify.
Doing agentic scientific research. The Terminal-Bench-Science jump is worth testing, with the error bars in mind.
FAQ
Is Claude Fable 5.1 cheaper than Fable 5? Base input and output prices are identical at $10 and $50 per million tokens. Only cache reads changed, from $1 to $0.25 per million. Anthropic estimates 25% saving on typical workloads and up to 45% on heavily agentic ones, and both come from that one line.
When was Fable 5.1 released? September 1, 2026, alongside Claude Mythos 5.1.
What are the model IDs?
claude-fable-5-1 and claude-mythos-5-1.
Is it a drop-in replacement for Fable 5? No. Forced tool use now errors, earlier models cannot read its thinking blocks, and editing earlier turns invalidates thinking blocks.
How does it compare to Opus 5? Better on every published benchmark, but by a few points on most, and it costs twice as much ($10/$50 versus $5/$25). Anthropic's own docs recommend starting with Opus 5.
What is the context window? 1M tokens input, 128K max output, unchanged from Fable 5.
Can I use Mythos 5.1? Only as a Project Glasswing participant. It is invitation-only with the same capabilities and price as Fable 5.1.
Sources
- Introducing Claude Fable 5.1 and Claude Mythos 5.1 — Anthropic, September 1, 2026
- Claude Fable 5.1 overview — Claude Platform Docs
- What's new in Claude Fable 5.1 — Claude Platform Docs
- Claude model pricing table — Claude Platform Docs
