Rates verified July 27, 2026
Claude Fable 5 cost calculator
Most token calculators multiply your token count by the list price and stop there. Three things move the real number, and this one models all of them.
What this adds over a plain token calculator
First, the effort parameter moves output volume by several times between low and max — the expensive model at low effort is often the cheap option. Second, a share of Fable 5 requests hit a safety classifier and are answered by Opus 4.8, billed differently and invisible to your error handling. Third, if your token counts were measured on Opus 4.6 or earlier, they are low by roughly 30%: the tokenizer changed at Opus 4.7 and every current model inherits it. Skip any of the three and your estimate is wrong in a direction you did not choose.
Monthly cost estimator
My token counts came from a pre-Opus-4.7 model
The tokenizer introduced with Opus 4.7 splits the same text into roughly 30% more tokens. Fable 5, Opus 4.8 and Opus 5 all share it — turn this on only if you measured on Opus 4.6 or earlier.
Route through the Batch API
50% off both directions, but responses are not real-time.
Claude Fable 5
$238.19
per month · $5.81/MTok effective
Claude Opus 5
$122.15
per month · $2.98/MTok effective
Claude Opus 5 is cheaper by $116.04 per month on this workload.
- Fresh input
- $111.15
- Cache reads
- $25.93
- Output
- $95.00
- Safeguard fallback
- $6.11
Estimates, not a quote. Two inputs are modelled rather than published (1.3x tokenizer, effort estimates) · Rates verified 2026-09-02
DeepSeek, Qwen, GLM and Kimi rates are vendor-published as of August 2026 and not independently verified. Effort-tier multipliers below are Anthropic-only estimates, applied to competitors as a rough proxy.
The three layers nobody else models
Each of these is documented behaviour, not speculation — but none of them appear in the list price.
The tokenizer changed at Opus 4.7 — but not between today's models
Anthropic's token-counting documentation states that Fable 5 and Mythos 5 use the tokenizer introduced with Opus 4.7, which produces roughly 30% more tokens than models released before it for the same text. The migration guide gives the range as 1.0x–1.35x depending on content, with code, JSON, XML and YAML at the top of it. The practical consequence is narrower than it first looks: moving from Opus 4.8 or Opus 5 to Fable 5 involves no tokenizer change at all, because they already share it. The toggle above matters only if your baseline came from Opus 4.6 or earlier — in which case your budget is short by about a third.
Effort moves output volume several times over
Effort is an API parameter (low, medium, high, xhigh, max — default high), not a prompt trick, and it changes how many tokens the model spends. On one published measurement, an identical prompt produced about 1,900 output tokens at low effort and about 14,400 at max. Anthropic has not published a cost curve, so the multipliers here are interpolated from that single anchor and labelled as estimates.
Safeguard fallbacks are billed, and they are not free
When a request trips Fable 5's cybersecurity or biology classifiers it is answered by Opus 4.8 instead, returned as HTTP 200 with stop_reason: refusal. You pay for the model that actually served the request: if the block lands before any output — the common case — the declined Fable 5 attempt is not billed at all and the turn bills at Opus 4.8's rates. A mid-stream block is the expensive shape, because the prefix that already streamed bills at Fable 5 rates. Anthropic reports fallbacks in under 5% of conversations on average; Artificial Analysis measured 9% on one hard benchmark, and prompts that ask the model to narrate its reasoning raise the rate sharply.
Where this calculator is honest about not knowing
Token counts are estimated from your inputs, not measured — for an exact figure, call Anthropic's count_tokens endpoint with your real prompt. The effort multipliers are interpolated from a single public measurement and will not match every workload. Cache-write costs are excluded because they depend on how often your prefix changes. Treat the output as a planning range, not an invoice.
Cut the number before you pay it
Caching a stable system prompt, dropping effort one notch on routine work, and removing reasoning-narration instructions each move the bill more than switching models does.