Home → Guides → Claude Fable 5.1 (claude-fable-5-1) API pricing, cache rate and the four parameters that return 400

Claude Fable 5.1 (claude-fable-5-1) API pricing, cache rate and the four parameters that return 400

One unusual price rule and four hard parameter limits. Both are easy to miss and both change what this model costs you.

The short version

• Cache reads bill at $0.25 per 1M — 0.025× the input rate, against 0.1× on every other Claude model. That is not a typo and not a promotion; the arithmetic below checks it against a real invoice.
• Its cache reads are cheaper than Opus 5's ($0.25 against $0.50 per 1M) even though its list input price is twice as high. In a loop where almost all input is cache reads, the expensive-looking model is the cheaper one.
• Four parameters return a 400 rather than being ignored. Forced tool use, temperature, top_p and top_k. Code written for other Claude models will break, loudly.
• 1,000,000 tokens of context at standard per-token pricing across the whole window, 128k max output. List price is $10.00 per 1M input, $50.00 per 1M output.

The cache rate, checked against a real bill

Documentation is worth what you paid for it. Here is the arithmetic across all 61 billed calls that have gone through this platform:

input             50,641 tokens  ×  $10.00 / 1M              =  $0.506410
output            20,943 tokens  ×  $50.00 / 1M              =  $1.047150
cache read   412,044 tokens  ×  $0.25 / 1M   (0.025×)  =  $0.103011
cache write     147,794 tokens  ×  $12.50 / 1M  (1.25×)   =  $1.847425
                                                     ─────────
computed                                             $3.503996
actually billed                                      $3.5040

Those agree, to the sixth decimal. And the cache-read line is where the unusual rate shows: those same 412,044 tokens cost $0.103011 at 0.025×. Priced at the 0.1× every other Claude model charges they would cost $0.412044 — 4 times as much. That gap is far too large to be a rounding artefact.

What that rate actually changes

List price says Fable 5.1 costs twice what Opus 5 does on input ($10.00 against $5.00 per 1M). Cache reads reverse it: $0.25 against $0.50. So which model is cheaper depends entirely on how much of your input arrives as a cache read.

Setting the two input-side costs equal and solving gives a single crossover:

95.2% cache reads

Below that share, Opus 5 is cheaper on input. Above it, Fable 5.1 is — despite the doubled headline price.

That threshold is not theoretical. Our heaviest Claude workload, claude-sonnet-5 over 503 calls, ran 7,289,753 fresh input tokens against 31,008,712 cache reads — 81.0%. At that share, the same input tokens cost $80.65 on claude-fable-5-1 against $51.95 on claude-opus-5 — still short of the crossover, but by 14.3 points of cache share, not by an order of magnitude.

The practical reading: a long-running agent with a stable prompt prefix is exactly the shape where Fable 5.1's pricing stops being expensive. A chat endpoint that rebuilds its prompt every call is exactly the shape where it stays expensive. Same model, same price list, opposite conclusions — and you cannot tell which one you have without reading the cache column in your own usage log.

Four parameters that return a 400

Most APIs quietly ignore a parameter they do not support. This one rejects the request — better behaviour, and worse for anyone porting working code across.

What you send	Result	What to do instead
tool_choice type any or tool	400	Keep type auto and say in the prompt when the tool applies
temperature off default	400	Steer with the prompt; there is no sampling knob here
top_p off default	400	Same
top_k off default	400	Same

Forced tool use is the one that catches people. Thinking is always on for this model, and a forced tool call would skip it — so rather than silently degrading the model into something you did not ask for, the API refuses. If your agent framework sets tool_choice to any to guarantee a structured response, that is the line to change.

Prefilling the assistant turn is also unsupported. All of these apply to claude-fable-5 as well.

Calling it

from openai import OpenAI

client = OpenAI(api_key="YOUR_KEY", base_url="https://apiclan.us/v1")

resp = client.chat.completions.create(
    model="claude-fable-5-1",          # hyphens, not "5.1"
    messages=[{"role": "user", "content": "Hello"}],
)

The model string is claude-fable-5-1. Writing it with a dot is the most common failure here, and it returns a model-not-found rather than falling back to anything.

What you pay here

Per 1M tokens	List price	APICLAN	You save
Input	$10.00	$4.50	55%
Output	$50.00	$22.50	55%

Cache reads and writes bill at the same fractions of their own rates, so the 0.025× read advantage carries through to what you actually pay.

When not to use it

• Anything short. Classification, routing, tagging, extraction. The long-horizon reasoning you are paying for contributes nothing, and always-on thinking means paying for tokens nobody reads.
• Anything that needs forced tool use. Not a preference — a 400.
• Anything that needs temperature control. If you tune sampling for output diversity, this model cannot do it.
• Fresh-prompt-every-call workloads. Below 95.2% cache reads the pricing works against you and Opus 5 is the cheaper reasoning model.

Where it earns the price: hours-long agent loops, multi-file refactors, multistep research over a large stable corpus — work where context is built once and read many times, which is exactly the shape the cache rate is designed for.

Questions people actually ask

Start using it

No subscription, no monthly minimum, no sales call. Top up with USDT and spend what you use — 1 USDT gives you 2 credits of API balance.

Read the 30-second quickstart

Prices quoted on this page are regenerated automatically from live billing data. Third-party terms are quoted from that party's own published documentation.