Anthropic's Claude Fable 5.1 — same sticker price, cache reads cut 75%
Anthropic released Claude Fable 5.1 and Mythos 5.1 on September 1, 2026, and the striking thing about the price list is that only one line moved. Input remains 10 dollars per million tokens and output 50 dollars, identical to Fable 5. Cache reads dropped 75 percent, from 1.00 dollar to 0.25 dollars per million tokens, while cache writes hold at 12.50 dollars for five-minute retention and 20 dollars for one hour. Anthropic says the change cuts bills by 25 to 45 percent, but outside tests disagree on whether the cost of completing a given task actually falls. Performance rose sharply on agentic work: Terminal-Bench-Science 0.1 went from 24.7 to 52.6 percent, Terminal-Bench 4.0 from 42.0 to 55.8, and AutomationBench from 17.1 to 31.4. Mythos 5.1 is the same weights with looser safeguards, available only to vetted cybersecurity and life-sciences organisations
The three lines
- Price — input $10 and output $50 unchanged; cache reads $1.00 → $0.25 (−75%)
- Performance — Terminal-Bench-Science 24.7% → 52.6%, AutomationBench 17.1% → 31.4%
- Two builds — Mythos 5.1 is the same model with looser safeguards, vetted access only
Key questions
- How much does Claude Fable 5.1 cost?
- **Ten dollars per million input tokens and fifty per million output tokens — identical to Fable 5.** One line changed: **cache reads fell from 1.00 dollar to 0.25 dollars per million tokens, a 75 percent cut.** Cache writes are unchanged at **12.50 dollars** for five-minute retention and **20 dollars** for one hour, and batch processing still takes 50 percent off input and output. The context window is **1 million tokens** with a maximum output of **128,000 tokens**. For comparison, OpenAI's GPT-6 Astra, released September 3, also prices at 10 dollars input and 50 dollars output with a 1.05 million token context (see "OpenAI launches GPT-6 Astra, September 3, 2026"). **Frontier sticker prices have converged**, which is why competition has moved down the list into caching, batching and context length.
- How much does a cheaper cache read actually save?
- **Anthropic says bills fall 25 to 45 percent; outside tests are less certain.** Both claims can be true, and the reason is in what caching saves. Prompt caching stores a repeated prefix — a system instruction, a reference document, a codebase — so it does not have to be processed again on every call. You pay a premium once to write it and a small amount each time you read it. **The saving therefore depends on how many times the same prefix is reused.** Load one codebase and ask fifty questions against it, and the read-price cut passes straight to the bill. Send a different document every time, and write costs dominate while the read cut barely lands (see "What prompt caching is"). There is a second variable. **In agentic work, the more steps a model takes on its own, the more total tokens it consumes.** A unit price cut to a quarter is cancelled if usage quadruples, and stronger models tend to try longer and harder. That is precisely where the outside tests diverge: **the unit price definitely fell; the total depends on the workload.**
- What is the difference between Mythos 5.1 and Fable 5.1?
- **The model is the same; the safeguards are not.** Anthropic describes Mythos 5.1 as **the same underlying model** as Fable 5.1 with **more permissive safeguards**, available only to **vetted cybersecurity and life-sciences organisations**. Why ship two builds. These are the two domains where **the dangerous knowledge and the defensive knowledge are the same knowledge.** You cannot block a vulnerability you do not know how to find, or counter a pathogen whose mechanism you do not understand. Leave the general-release safeguards in place and legitimate defensive research is blocked too; remove them and anyone can use the result. **Splitting by access eligibility** is the compromise Anthropic chose, and the pattern is spreading — Google DeepMind shipped a defenders-only Cyber variant alongside Gemini 3.8 Flash on September 2 (see "Google releases Gemini 3.8 Flash"). On the Fable 5.1 side, Anthropic says unnecessary safeguard interventions in production fell by roughly 60 percent — a claim about false positives, not about permitting harmful requests, and one that has not been independently verified.
Anthropic released Claude Fable 5.1 and Claude Mythos 5.1 on September 1.
The first thing anyone reads on a model launch is the price list, and this time there is almost nothing on it. Input 10 dollars per million tokens, output 50 dollars. Identical to Fable 5.
Exactly one line moved.
1. Cache reads: $1.00 → $0.25
| Item | Fable 5 | Fable 5.1 |
|---|---|---|
| Input (per 1M tokens) | $10 | $10 |
| Output (per 1M tokens) | $50 | $50 |
| Cache write (5-minute) | $12.50 | $12.50 |
| Cache write (1-hour) | $20 | $20 |
| Cache read | $1.00 | $0.25 (−75%) |
| Batch processing | 50% off input/output | Same |
| Context window | — | 1M tokens |
| Max output | — | 128,000 tokens |
The sticker price held and only an ancillary line fell. That is a fair picture of where the frontier market now sits.
OpenAI's GPT-6 Astra, released September 3, also prices at 10 dollars input and 50 dollars output, with a 1.05 million token context (see "OpenAI launches GPT-6 Astra, September 3, 2026"). The two leading models carry the same headline numbers to the decimal. When the marquee figures converge, competition drops to the line beneath — caching, batching, context length, and the total cost of actually finishing a task.
2. Is it 25 to 45 percent, or isn't it?
Anthropic says the change cuts bills by 25 to 45 percent. Outside tests were reported as disagreeing on whether the cost of completing a task actually falls.
Both can be true. It depends on what caching saves.
Prompt caching stores a repeated prefix — a system instruction, a reference document, a codebase — so it is not reprocessed on every call. You pay a premium once to write it ($12.50–$20) and a small amount each time you read it ($0.25).
The saving is set by how many times the same prefix is reused.
- Load one codebase, ask fifty questions → the read-price cut passes straight through
- Send a different document each time → write costs dominate, the read cut barely lands
And there is a second variable. In agentic work, total tokens grow with the number of steps the model takes by itself. A unit price cut to a quarter is erased if usage quadruples, and more capable models tend to try longer. That offset is real (see "What prompt caching is").
This is exactly where the outside tests split. The unit price definitely fell. The total depends on the workload.
3. The gains are in agentic work
The benchmark movement Anthropic published:
| Benchmark | Fable 5 | Fable 5.1 |
|---|---|---|
| Terminal-Bench-Science 0.1 | 24.7% | 52.6% |
| Terminal-Bench 4.0 | 42.0% | 55.8% |
| AutomationBench | 17.1% | 31.4% |
| CursorBench 3.2.0 | — | 73.4% |
Terminal-Bench-Science more than doubled. Benchmarks in this family test whether a model can drive tools in a terminal across many self-directed steps to finish a task. They measure completion, not the quality of a single answer.
The standard caveat applies. These are all vendor-supplied numbers, and measurement conditions vary between companies — how many attempts are allowed and which tools are available change what the same benchmark name means (see "What an AI risk rating is"). What is solid here is that this is a like-for-like comparison of two models under one company's own harness.
4. Mythos 5.1 — same model, different eligibility
Mythos 5.1 uses the same weights as Fable 5.1. What differs is the strength of the safeguards, and it is open only to vetted cybersecurity and life-sciences organisations.
Why two builds. These are the two fields where the dangerous knowledge and the protective knowledge are the same body of knowledge. You cannot close a vulnerability you cannot find, or counter a pathogen whose mechanism you do not understand. Leave general-release safeguards in place and legitimate defensive work is blocked; remove them and anyone can use the result. Splitting by who is allowed in is the compromise.
The pattern is spreading. Google DeepMind shipped a defenders-only Cyber variant with Gemini 3.8 Flash on September 2 (see "Google releases Gemini 3.8 Flash"), and OpenAI classified Astra as the first model to reach its own critical cyber threshold in early September (see "OpenAI's Astra hits a critical cyber rating").
On the Fable 5.1 side, Anthropic says unnecessary safeguard interventions in production fell by about 60 percent — a claim about wrongly blocked benign requests, not about permitting harmful ones. It has not been independently verified.
5. What is still open
- The 25-45 percent figure is the vendor's. The workload mix behind it was not published. Check your own cache reuse ratio before assuming the number applies to you.
- The Mythos vetting process is undisclosed. Who decides eligibility, on what criteria, and whether approved organisations are listed — none of it is public. When dangerous capability is controlled by eligibility, the eligibility decision is the safeguard.
- The axis of price competition has moved. Fable 5.1 (September 1), Gemini 3.8 Flash (September 2) and GPT-6 Astra (September 3) landed in three days. The headline prices have converged; the differences now live in caching, batching and context.
- Read it next to the Nvidia deal. On September 3, Nvidia acquired Hugging Face for 12.9 billion dollars (see "Nvidia buys Hugging Face for 12.9 billion dollars"). While model prices converge, the layer that distributes and runs those models is consolidating.
Sources
- VentureBeat — Anthropic's Claude Fable 5.1 and Mythos 5.1 arrive with a 75% cost reduction for Fable cache reads
- MarkTechPost — Anthropic Releases Claude Fable 5.1 and Claude Mythos 5.1: 52.6% on Terminal-Bench-Science and 75% Cheaper Cache Reads
- implicator.ai — Anthropic Fable 5.1 Keeps $10/$50 Price, Cuts Cache Reads 75%
- LLM Stats — Claude Fable 5.1 API pricing, context window and benchmarks
- gHacks — Anthropic Releases Claude Fable 5.1 and Cuts Cached Token Pricing by 75%
- Anthropic — Newsroom