Claude Fable 5 API: Cost, Pricing and How to Get a Key (2026)
Fable 5 lists at $10/$50 per million tokens — roughly twice Opus 5. Here is what that works out to on a real workload, where the cheaper model is the right answer, and how to get a key when your local card will not go through.
1. The one number that should drive your model choice
Most Fable 5 write-ups compare capability. Capability is real, but it is not what breaks budgets — the output price is. Fable 5 lists at $50 per million output tokens against Opus 5's $25, and output is where agentic workloads spend. A coding agent that reads 30K tokens of context and writes 8K tokens of patch is dominated by that 8K. Work the arithmetic once for your own shape of traffic before you pin a default model: take a representative day, pull the input and output token counts from your usage dashboard, and multiply. Teams are routinely surprised to find the flagship premium lands almost entirely on generation, not on reading.
2. Fable 5 against Opus 5, in blended terms
At a 3:1 input-to-output mix, Fable 5 blends to roughly $19 per million tokens on cocodot and Opus 5 to about $9 — a little over 2x. That ratio, not the headline capability gap, is the thing to hold in your head. A useful way to decide: if you cannot describe the specific failure mode that Opus 5 exhibits on your task and Fable 5 does not, you are paying 2x for a feeling. Conversely, when a task genuinely runs long — a twenty-step agent chain, a cross-file refactor where one early mistake poisons everything after it — the cheaper model can cost more in wasted retries than the expensive one costs outright.
3. Getting a key when your local card is declined
This is the wall most people actually hit, and it has nothing to do with your code. Card-not-present acquiring screens the issuing country, read from the card's first 6–8 digits (the BIN). Cards issued in some regions are declined at a high rate on AI-platform checkouts regardless of balance, and switching from Visa to Mastercard at the same bank changes nothing, because the BIN still encodes the same issuing country. What has to change is the issuing country. So: either obtain a card whose issuing country is accepted and keep using Anthropic's console directly — which preserves your direct commercial relationship and vendor SLA — or route your calls through an OpenAI-compatible endpoint that accepts local payment. Neither is strictly better; they trade a vendor relationship against a payment method.
4. The setup, in three lines per SDK
Nothing about the call changes except the base URL and the model name. With any OpenAI SDK, point base_url at https://cocodot.co/api/ai/v1 and set model to claude-fable-5 (or the short code mcf-1); official names auto-route, so you do not have to memorise codes. If you are on the Anthropic SDK or Claude Code, set ANTHROPIC_BASE_URL=https://cocodot.co/api/ai instead and leave the rest of your configuration alone — an Anthropic-compatible endpoint is served on the same host (beta). Streaming, system prompts, tool use and multi-turn all behave as they do upstream.
from openai import OpenAI
client = OpenAI(
base_url="https://cocodot.co/api/ai/v1",
api_key="sk-YOUR_KEY",
)
resp = client.chat.completions.create(
model="claude-fable-5", # or the short code: mcf-1
messages=[{"role": "user", "content": "Summarise this changelog."}],
)
print(resp.choices[0].message.content)5. What the free trial credit actually covers
Verifying your email grants $0.5 of trial credit, and it is worth being precise about what that buys, because the honest answer is narrower than most signup pages imply. The trial credit is scoped to the budget models — it is enough to prove that your base URL, key, headers and streaming all work end to end, which is the part that usually goes wrong. It is not enough to exercise Fable 5, and it is not scoped to the flagships at all: at $47.50 per million output tokens, fifty cents is a rounding error. Use it the way it is actually useful — smoke-test the plumbing on a cheap model first, confirm a real response comes back, and only then top up and switch the model string.
6. Caching: where we are cheaper, and where Anthropic is cheaper
Prompt caching works here, and the billing rule is simple: a cache hit is charged at half our input rate, and a cache write is charged at the standard input price with no surcharge. For Fable 5 that puts a cache hit at about $4.75 per million tokens. Now the part a vendor page does not usually tell you: Anthropic's own cache-hit pricing is a tenth of its list input price, roughly $1 per million for Fable 5 — so on this specific line item, calling Anthropic directly is several times cheaper than calling us. If your application re-sends a large fixed prefix on nearly every request (a long system prompt, a retrieved corpus, a big tool schema) and your hit rate is high, that difference can dominate your bill, and the honest recommendation is to go direct if you can pay them. Where an intermediary earns its place is the other three quadrants: output tokens, uncached input, and being able to pay at all.
7. Route by task, and make the router the default
The single most effective cost control is not negotiating a rate — it is making sure the expensive model only sees the requests that need it. Set your default to a cheap tier and escalate explicitly, rather than defaulting to the flagship and hoping to remember to downgrade. In practice that means: variable renames, log statements, boilerplate and routine Q&A on the cheapest tier; ordinary feature work and tests on Sonnet 5; and Fable 5 or Opus 5 reserved for cross-file refactors, long-context analysis and multi-step agent runs. Most coding agents let you bind different models to different modes in settings. It is one afternoon of configuration against a recurring monthly bill.
8. Verify you are getting the model you are paying for
Any intermediary creates the same structural temptation: bill for a flagship, serve something cheaper, and count on easy prompts hiding the difference. The incentive scales with the price gap, which makes the most expensive model on the menu the one most worth checking. Do not settle this with promises — settle it with a probe. cocodot's downgrade checker is open source with a hosted version at probe.cocodot.co: paste in any OpenAI-compatible base URL and a key and it runs the probes and reports back. It is deliberately usable against us as readily as against anyone else, and a provider's willingness to be tested is better evidence than any no-downgrade pledge. Whatever you choose, fund a small amount, run your own hard prompts, and scale only once your numbers come back clean.
Claude flagship tiers: list price vs cocodot, USD per million tokens (live figures at /pricing)
| Model | Model id | List (in / out) | cocodot (in / out) | Blended at 3:1 |
|---|---|---|---|---|
| Fable 5 | claude-fable-5 / mcf-1 | $10 / $50 | $9.50 / $47.50 | ~$19 |
| Opus 5 | claude-opus-5 / mco-7 | $5 / $25 | $4.50 / $22.50 | ~$9 |
| Sonnet 5 | claude-sonnet-5 | $2 / $10 | $1.80 / $9 | ~$3.60 |
| Cache hit (cocodot) | any of the above | — | half our input rate | see section 6 |