What changed today: Anthropic removed Claude Fable 5 from the standard usage allowance on Pro, Max, Team, and seat-based Enterprise plans. Continued access requires usage credits drawn down at API pricing. Fable 5 remains available — it simply costs more than it did yesterday.
The pricing in plain English
Fable 5 via usage credits on any Claude subscription plan costs $10 per million input tokens and $50 per million output tokens. Those are the same rates as the public Claude API, applied inside your subscription account.
A 1,000-word document summary (roughly 1,300 tokens input, 400 tokens output): approximately $0.033 — around 2.6p. A full proposal draft with a detailed brief (roughly 4,000 tokens input, 2,000 tokens output): approximately $0.14 — around 11p. A month of daily proposal drafts (20 working days): approximately £2.20. Those numbers are small individually. They accumulate if Fable 5 is your default for everything rather than your premium option for high-stakes tasks.
For comparison, Sonnet 4.6 remains included in your plan limits at no per-token charge. The practical question is not whether Fable 5 costs money — it clearly does — but whether the quality difference for your specific task justifies the cost.
Fable 5 versus Sonnet 4.6: where the gap actually matters
Fable 5 scored 80.3% on SWE-Bench, making it the strongest publicly available coding model at launch. It also outperforms Sonnet 4.6 on multi-step reasoning, nuanced document analysis, and complex instruction-following. However, Sonnet 4.6 is itself an excellent model — for most conversational, drafting, and summarisation tasks, the quality difference is marginal.
Complex code generation or debugging with multiple files and dependencies. Legal or contract document analysis where missing a nuance has a real cost. Multi-step reasoning tasks like financial modelling logic or structured business case development. Long-form proposals where the quality of argument structure and client-specific framing is the differentiator.
Customer email drafting and follow-up sequences. Social media captions and short marketing copy. Meeting notes summarisation. FAQ and knowledge base content. Internal process documentation. General research summaries. Most day-to-day admin automation.
What Anthropic has said about restoring Fable 5 access
Anthropic has confirmed that Fable 5 was moved to credits due to capacity constraints, not a permanent pricing decision. The company has said it intends to restore Fable 5 as a standard plan feature once capacity allows, but has not provided a date. The pattern from previous Anthropic model rollouts suggests this could be weeks to a few months.
This means that if you decide not to purchase credits now, you are not permanently locked out — you are simply waiting for capacity to expand. The question is whether the delay in using Fable 5 for your high-stakes tasks has a real cost to your business over the next few weeks.
What to do
The thirty-minute decision test
Take your hardest recurring AI task — the one where output quality has the most direct business impact. Run it on Fable 5 (using credits, so this test will cost a small amount). Run the same task on Sonnet 4.6 (plan-included, no additional cost). Compare the outputs honestly on the criteria that matter for that task: accuracy, structure, tone, completeness. If the Fable 5 output is materially better in a way that affects a real business outcome — a better proposal, a more accurate analysis, a cleaner code function — calculate the monthly credit cost for that specific task and decide whether it is justified. If the difference is marginal or the output would need the same amount of editing regardless, Sonnet 4.6 is your daily driver and credits are not worth it at this stage. Do not drift into habitual Fable 5 use without this test. Drifting is how small per-token costs become surprising monthly bills.
The broader principle here is relevant beyond just Fable 5: as AI tools move from flat-rate to usage-based pricing, the operators who will manage costs well are those who are deliberate about which model they use for which task, rather than defaulting to the most powerful option for everything. That deliberateness is a skill worth building now, while the cost differences are still small.