Claude Fable 5.1 API: Model ID, Rates, Effort and Cache TTL

Claude Fable 5.1 shipped on 1 September 2026. This post is the reference: the identifiers, the five published rates, and the cache-TTL choice that costs more than the price cut saves. For the cost argument against Opus 5, see Fable 5.1 vs Fable 5 vs Opus 5.

The identifiers

Adaptive thinking cannot be switched off on this model. That matters for budgeting, because thinking tokens are billed as output — the $50 line, not the $10 line.

All five rates

Anthropic publishes five rates per model. Most write-ups quote one of them. Dollars per million tokens:

Read the two Fable columns against each other. Four of the five numbers are identical. The version bump moved exactly one line.

Effort does not change the rate — it changes the token count

Effort is not a pricing tier. The per-token rates above hold at every effort level. What effort changes is how much the model writes before it answers.

Artificial Analysis, which took part in pre-release evaluation of the model, measured Fable 5.1 at max effort using roughly 1.7x the output tokens of Fable 5. Output is the $50 line. That is why their cost-per-task figure came out at $3.76 against Fable 5’s $3.14 — 20% higher — even though the cache read got 75% cheaper.

So the budgeting rule is: raising effort raises your bill through volume, not through rate. If you are cost-sensitive, cap effort before you optimize anything else.

What caching is worth

A cache read at $0.25 is one-fortieth the price of base input at $10. The saving compounds with reuse.

Take one 1M-token prefix — a codebase, a document set, a long system prompt. Write it once, read it back N times:

The table assumes a single 5-minute write with no expiry between reads. The “no caching” column is the same N+1 full passes at $10 per million.

At a single read the two versions are within 6% of each other. At a hundred reads Fable 5.1 costs a third of Fable 5. The model is identical across those rows; only the cache-hit count changes. That is the number to measure.

The TTL decision

Anthropic sells two cache-write durations: $12.50 per million for five minutes, $20 for an hour. The five-minute clock resets on reads. Anthropic’s documentation is explicit that the cache is refreshed for no additional cost each time the cached content is used, and that the lifetime is measured from each request that writes or reads the entry. An agent that keeps hitting the same prefix holds a five-minute entry open on reads alone and pays for exactly one write.

The money leaks in the gaps. Any pause over five minutes — a slow tool call, a human reading a diff — lets the entry expire, and the next call pays the write again. An hour of that costs at worst 12 × $12.50 = $150, against $20 for one 1-hour write.

That $130 gap per prefix per hour is decided by your traffic pattern rather than your model choice, which is why it survives the version bump entirely.

Migrating from Fable 5

The change is a string. Base rates, context window, output ceiling and default effort are all unchanged, so a swap of the model ID is the whole migration:

- model="claude-fable-5"
+ model="claude-fable-5-1"

Two things to re-check after the swap, because neither is visible in the diff:

• Your output budget. Same rate, but Artificial Analysis measured roughly 1.7x the output tokens at max effort. If you had alerts calibrated on Fable 5’s output volume, they will fire.
• Your knowledge-cutoff assumptions. A June 2026 cutoff means prompts that patched around stale library knowledge may now be redundant or, worse, actively wrong.

What Anthropic says changed besides price

The Claude Fable product page describes behavior rather than benchmarks. It says Fable 5.1 “avoids easy-seeming shortcuts, fixes the root causes of problems rather than the symptoms, and keeps you updated as it works”.

One correction worth making, since it has spread: several write-ups presented the drop in unnecessary refusals as a Fable 5.1 feature. That change is dated 7 August 2026 and applies to Fable 5’s biology safeguards, where Anthropic reports the update “reduced biology-related fallbacks by about 85% across our product surfaces”. Separate release, earlier date, different model.

Calling it through Ofox

As of 3 September 2026, Fable 5.1 is not in the Ofox catalog. The catalog lists 138 models and the only Fable entry is anthropic/claude-fable-5. There is no configuration that makes 5.1 work today.

What does work today, on one key: anthropic/claude-fable-5, anthropic/claude-opus-5 and anthropic/claude-sonnet-5, at Anthropic list prices with no markup. All three are reachable over either protocol:

The native path gives you the Messages API, including extended thinking and tool use. Caching parameters pass through untouched, so the TTL arithmetic above applies unchanged.

import anthropic

client = anthropic.Anthropic(
    base_url="https://api.ofox.ai/anthropic",
    api_key="<OFOX_API_KEY>",
)

message = client.messages.create(
    model="anthropic/claude-fable-5",
    max_tokens=1024,
    messages=[{"role": "user", "content": "Summarize this repository."}],
)

If and when 5.1 is added, the change should be the model string and nothing else. Until then, running the comparison on Fable 5 tells you what your cache-hit ratio actually is — which is the number that decides whether 5.1’s cheaper read is worth anything to you.

Rates and specifications verified against Anthropic’s pricing documentation, model overview, product page and launch announcement on 3 September 2026. Cost-per-task figures from Artificial Analysis, retrieved the same day. Ofox catalog and protocol base URLs checked against the live /v1/models endpoint and the Ofox documentation on 3 September 2026.