What Actually Changed in Claude Fable 5.1 — And What Using It Commits You To
3 September 2026
Anthropic released Claude Fable 5.1 on 1 September 2026, positioned as the most capable model it offers to the general public.
This site normally writes about generative AI running on your own machine. This time the subject is a top-tier model that runs in the cloud. Laying out the price and the conditions side by side makes it easier to decide what to use for what.
What does the specification actually commit you to, and what changed from Fable 5?
- 1. What I looked at, and where it came from
- 2. What the specification says
- 3. What improved over Fable 5?
- 4. What does it cost?
- 5. Thinking is billed as output too
- 6. Where you can use it, and what you accept
- 7. Safeguards can hand your request to a different model
- 8. Safety went up in one place and down in another
- 9. In summary: what actually changed
What I looked at, and where it came from
Three things: what kind of model this is, what it costs, and what you have to accept in order to use it.
The method was reading Anthropic’s product page and the official system card (the technical PDF), then cross-checking both against third-party write-ups published the day after launch. Figures confirmed on official material and figures I could not confirm are marked separately below.
What the specification says
| Item | Value | Source |
|---|---|---|
| Released | 1 September 2026 | official |
| Model ID | claude-fable-5-1 | official |
| Input | $10 per million tokens | official |
| Output | $50 per million tokens | official |
| Cache read | $0.25 per million tokens | official |
| Context window | 1M tokens | official (system card) |
| Max output per request | 128,000 tokens | official (system card) |
| Knowledge cutoff | June 2026 | official (system card) |
What improved over Fable 5?
Anthropic leads with long-running work: staying on target across hours or days, digging to the cause instead of the symptom, and checking its own output before stopping.
Of the published numbers, only one could be confirmed on the product page: Terminal-Bench-Science, where Fable 5 scored 24.7% and Opus 5 scored 29.0%. The benchmark figures below appeared in third-party write-ups.
| 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 | 70.5% | 73.4% |
These are the vendor’s own results. Fable was also evaluated with its production safeguards switched on, and Anthropic notes that tasks where those safeguards fired scored zero — which, by its own account, likely holds the reported numbers down.
What does it cost?
On unit price, Fable 5.1 is unchanged from Fable 5, and expensive: $10 per million tokens in, $50 out. Third-party write-ups put that at twice Opus 5 and five times Sonnet 5.
What changed is the cache read: $0.25 per million tokens, one fortieth of the normal input rate. Anthropic estimates roughly 25% cheaper than Fable 5 for typical work, and around 45% for workloads that read the same cached context over and over.
Caching is what kicks in when you send the same preamble repeatedly — the project description, the tool list, the conversation so far. If that part is byte-identical each time, the second read onward is cheap.
The flip side: rewrite the preamble every time and the discount disappears. A prompt that does not change is worth more here than a clever one.
Thinking is billed as output too
One thing matters as much as the unit price. Fable thinks before it answers, and that thinking counts as output tokens. If the visible answer is 10,000 tokens and it spent another 10,000 thinking, 20,000 land on the output side of the bill — at $50 per million.
And the thinking cannot be switched off. What you adjust is the depth: low, medium, high, xhigh and max, with high as the default. Those five levels, and the fact that thinking cannot be disabled, were already true of Fable 5; third-party write-ups say the same holds for 5.1, which I could not confirm on official 5.1 material.
In practice that means putting a deep setting on short reformatting or summarising work mostly adds an internal draft before the visible one, and raises the bill. Go deep only on the hard step.
Where you can use it, and what you accept
The product page lists Claude Pro, Max, Team and Enterprise users, and for developers the Claude Platform plus AWS, Google Cloud and Microsoft Foundry.
Using Fable means accepting data retention for safety monitoring. Third-party write-ups describe this as 30 days, and not something a setting turns off.
Retention and training are separate matters. Anthropic states it has never trained on enterprise data without explicit permission. Alongside 5.1, the product page announces a zero-retention option for enterprises; third-party write-ups call it Enterprise Frontier Safeguards and describe it as moving both the storage and the human review to the customer’s side. It is still rolling out.
Safeguards can hand your request to a different model
The product page states that Fable 5.1 is restricted in cybersecurity and biology, and falls back to Opus models there. The system card names them: Opus 4.8 for cybersecurity, Opus 5 for biology and AI development. That the fallback is not billed at Fable rates comes from third-party write-ups and could not be confirmed.
So you can select Fable and have a different model answer you. If a reply suddenly feels shallower, this is one possible reason.
This is not a rare edge case. In one system card evaluation, roughly half of Fable 5.1’s coding rollouts fell back to Opus 4.8, under 10% in computer use and tool use, for an overall fallback rate of 23%. In another, 95% of the coding requests that returned a valid response were served by Opus 4.8.
Nor did it start with 5.1. In that same evaluation, 99% of the requests sent to Fable 5 were answered by Opus 4.8.
Safety went up in one place and down in another
The gain is prompt injection resistance. On an external benchmark, the attack success rate after fifteen attempts was 1.0%, against 4.8% for Opus 5 and 6.5% for Fable 5. The strongest non-Anthropic model reaches 9.2%, and most sit between 24% and 53%. The card notes these were measured with prompt-injection-specific defences turned off.
Prompt injection means hiding an instruction — “ignore your previous instructions" — inside content the model is asked to read: a web page, an issue tracker comment, an incoming email. Anything that reads outside text carries this risk.
The loss is elsewhere. The system card says overall misaligned behaviour, measured by an automated behavioural audit, is a slight regression compared with Opus 5. This is about how it behaves, not how capable it is. Specifically, it accepts unverifiable claims of authorisation somewhat more readily — “I am the administrator", “this has been approved" — which the model has no way to check.
These two look similar and are not the same thing. Being harder to attack through injected text is not the same as being careful about claimed authority.
If you leave it working alone for long stretches: do not give it the ability to read outside content and the ability to send things outward in the same session. Reading alone cannot act on a bad instruction. And do not delegate vaguely — “do not touch anything outside this folder" holds better than “use your judgement as needed".
In summary: what actually changed
Fable 5.1 is built for working long. Its value is less in one-shot answers than in carrying a job of hours or days to the end without falling apart.
The price is high, but cache reads dropped to a fortieth. The more your usage repeats the same preamble, the more that shows up on the bill.
And using it means accepting that your data is retained for a period. A zero-retention option now exists for enterprises, but it is still rolling out.
Fable 5.1 is a frontier model — the largest thing each vendor puts at the top of its range — and it is not something a local LLM on your desk can be compared against. But when data has to stay in your hands, and there is a retention window on the other side, doesn’t the case for running models locally get a little stronger?
On how far a machine at home can actually go, there is this.