Anthropic released Claude Fable 5.1 in September 2026. The headline price did not move: $10 per million input tokens and $50 per million output tokens, the same as Fable 5 in June. The change that matters is the cache price.
If you run agents, coding assistants or anything that resends a large system prompt and tool definitions, Fable 5.1 can lower your bill without changing product code. If you send single-turn requests with little repeated context, it changes almost nothing.
This is the upgrade note. For the three-vendor comparison, see the AI model comparison for 2026.
What changed
| Claude Fable 5 (June 2026) | Claude Fable 5.1 (September 2026) | |
|---|---|---|
| Input | $10 / MTok | $10 / MTok |
| Output | $50 / MTok | $50 / MTok |
| Cache read | About $1.00 / MTok | $0.25 / MTok |
| Position | Flagship | Flagship |
| Focus | Coding, writing, long-horizon agents | Same, with cheaper repeat context |
The input and output rates are unchanged. Cache reads fell to a quarter of a dollar per million tokens.
Published benchmarks
Anthropic published detailed scores alongside the release. These are vendor-run, not independent, but they are the most specific numbers available for this generation.
| Benchmark | Fable 5.1 | Fable 5 | Opus 5 | GPT-5.6 Sol |
|---|---|---|---|---|
| Terminal-Bench 4.0 | 55.8% | 42.0% | 52.3% | 37.3% |
| Terminal-Bench-Science 0.1 | 52.6% | 24.7% | 29.0% | 22.4% |
| CursorBench 3.2.0 | 73.4% | 70.5% | 70.0% | 67.2% |
| AutomationBench | 31.4% | 17.1% | 26.9% | 19.6% |
| Humanity’s Last Exam (with tools) | 65.0% | 63.8% | 63.6% | — |
The gain over Fable 5 is largest on agentic and scientific terminal work. That fits Anthropic’s positioning: this release is aimed at long-horizon work, not chat.
Two caveats. The benchmarks are vendor-reported, and OpenAI has not published GPT-6.1 scores on the same tests. Read the benchmark guide before treating any of these as a ranking.
Why cache reads dominate agent bills
An agent does not send a fresh question each turn. It resends the system prompt, tool schemas, retrieved documents and conversation history. On a long run, most input tokens are repeats.
That means the cache-read rate, not the headline input rate, sets your bill. A model that looks expensive per token can be cheap per task if most of its input is cached.
An illustrative monthly bill
Assume 100 million input tokens a month, 10 million output tokens, and 90% of input served from cache:
| Line item | Fable 5 | Fable 5.1 |
|---|---|---|
| Cached input (90M) | $90 | $22.50 |
| Fresh input (10M) | $100 | $100 |
| Output (10M) | $500 | $500 |
| Total | $690 | $622.50 |
That is about a 10% saving on an already cache-heavy workload. Anthropic estimates around 25% for typical workloads and up to around 45% for highly agentic work where cache reads dominate the bill. The difference depends on your cache-hit ratio and how much output you generate.
A quick way to estimate your own saving
The saving is simple to size. Every million cached tokens saves about $0.75 compared with Fable 5:
monthly saving ≈ cached tokens in millions × $0.75
At 200 million cached tokens a month, that is about $150. At 1 billion, about $750. Teams that resend the same large context all day see the largest effect.
What did not change
Anthropic has not repositioned the model. Fable 5.1 remains the flagship for coding, writing and long-horizon agent work. The $10 / $50 public rate is the same, and the model is still the most expensive of the three major vendors.
That matters for expectations. An upgrade does not turn Fable into a budget model. It makes the same model cheaper to run when your workload repeats context.
Data controls that changed
Fable 5.1 also arrived with a new data model. Enterprise Frontier Safeguards (EFS) keep customer data in cloud infrastructure controlled by the customer rather than Anthropic, while still allowing misuse detection. It is rolling out in phases, and until it is available eligible customers can use Fable 5.1 with zero data retention.
Anthropic also added an invisible watermark to outputs from models released after 2 August 2026, to comply with the EU AI Act code of practice, with a detection API in private preview for regulators, media and researchers.
For Indian businesses, the data-residency detail is the one to watch. See AI and the DPDP Act.
Who should upgrade
- Agent platforms with long system prompts and tool schemas resent on every turn.
- Coding assistants that keep repository context in the prompt.
- RAG systems that hold the same documents in context across a session.
- Teams already on Fable 5 with a high cache-hit ratio and meaningful monthly spend.
Who can wait
- Single-turn chat or short prompts with little repeated context.
- Low monthly volume, where the absolute saving is small.
- Teams still deciding whether they need Fable at all. Opus 4.8 at $5 / $25 is half the price for routine work, as covered in Claude Fable 5 vs Opus 4.8.
What to test before switching
- Cache-hit ratio. Pull it from your provider’s billing data, not from assumptions.
- Turns to completion. Fable often finishes a task in fewer turns than cheaper models, which reduces total tokens.
- Output quality on your hardest tasks. Test the 20 cases that matter, not the easy ones.
- Cost per completed task. Include retries and human fallback, not just token price.
- Refusal and fallback behaviour. Check how the model handles edge cases your product will hit.
Fable 5.1 against the alternatives
| Claude Fable 5.1 | GPT-6.1 Sol | Gemini 3.6 Flash | |
|---|---|---|---|
| Input | $10 | $1 | $0.75 |
| Output | $50 | $5 | $3.75 |
| Strongest at | Coding, writing, long agent runs | Agentic coding, computer use, professional work | Volume, multimodal, batch |
| Weakness | Price per token | Slightly behind Fable on the hardest coding | Less depth on hard reasoning |
Fable 5.1 is the quality choice, not the value choice. If your work is genuinely hard, the price is justified by fewer turns and better first-pass output. If it is not, the cheaper models are close enough that the difference rarely reaches the customer.
A practical upgrade checklist
- Confirm the model name and version with your provider before switching.
- Run your evaluation set on both versions and compare blind.
- Check that prompt caching is actually enabled in your integration.
- Measure cost per completed task for two weeks.
- Keep the old model available for rollback.
- Update your internal documentation and cost forecasts.
FAQ
Is Claude Fable 5.1 more expensive than Fable 5?
No. The public list price is unchanged at $10 input and $50 output per million tokens. The cache-read price fell to $0.25 per million tokens.
Did Fable 5.1 get better at coding?
Yes, on Anthropic’s published benchmarks. It scored 55.8% on Terminal-Bench 4.0 against 42.0% for Fable 5, and 52.6% on Terminal-Bench-Science 0.1 against 24.7%. Those are vendor-run results, so test on your own repository before committing.
Should I upgrade if I do not use caching?
It is not urgent. Without repeated context, the main change does not apply to you, and the input and output rates are the same.
What is a cache read?
A cache read is an input token the provider serves from a previously stored prompt prefix instead of processing it fresh. It is billed at a much lower rate, which is why long, repeated prompts benefit most.
Next step
If you are choosing a model for a product you are building, our AI app development cost guide covers the build side. For a short evaluation on your own data, contact 24Bit System.
Related reading: GPT-6.1 Sol vs Terra, GPT-6.1 vs Claude Fable 5.1 and Claude models 2026.
