RSS
News

Fable 5.2 pricing drops to $8/$40 per million tokens

Fable 5.2 costs $8 per million input tokens and $40 per million output, down from $10 and $50. The full price sheet, why cost per task matters more, how effort levels and fast mode change the bill, and a usage-credits gotcha in Claude Code.

silver round coins on white surface
Photo by Chris Briggs on Unsplash

Updated November 20, 2026: Expanded with the full price sheet, caching, effort and fast mode costs, and subscription caveats.

With the Fable 5.2 release on November 9, Anthropic cut the price of its top public model to $8 per million input tokens and $40 per million output, down 20% from Fable 5.1's $10 and $50. Mythos 5.2, the same underlying model, costs the same but is only offered to vetted organizations through Anthropic's Life Sciences and Cyber Verification Programs. The model is live on the Claude API; Claude Code needs v2.1.321 or later (run claude update). TechCrunch wrote that the cut makes Fable 5.2 cheaper than GPT-6 Astra; our sources don't give Astra's price, so we can't check that.

The price sheet

Claude API list prices in USD, from Anthropic's pricing page, checked November 20, 2026.
ModelInput / MTokOutput / MTokCache hit / MTok
Fable 5.2, Mythos 5.2$8$40$0.20
Fable 5.1, Mythos 5.1$10$50$0.25
Opus 5.5$4$20$0.20
Sonnet 5.5$2$10$0.10
Haiku 5.5 (prompts up to 100K)$0.10$0.50$0.01

Cache reads on Fable 5.2 cost 0.025 times the base input price, against 0.1 times on most models, so a long session re-sending the same large prefix pays $0.20 per million tokens for the repeated part. Cache writes cost 1.25 times base input ($10) for a 5-minute cache and 2 times ($16) for a 1-hour cache; they pay for themselves after one and two cache reads respectively. The full 1M-token context is billed at the standard rate, except on Haiku 5.5, where prompts over 100,000 tokens cost five times more ($0.50 in, $2.50 out). Two more things change the real bill:

  • Tokenizer: Claude 4.7 and later models use a newer tokenizer that produces roughly 30% more tokens for the same text, depending on the content. A 20% cut per token is not a 20% cut per document if you are coming from an older model.
  • Data residency: asking for US-only inference with the inference_geo parameter multiplies every token category by 1.1 on the Claude API.

Cost per task, not cost per token

Anthropic's launch post puts effort and efficiency up front. On FrontierCode 2.0, its benchmark of whether maintainers would merge a model's code change, Fable 5.2 at High effort scored 45.3%. That is essentially level with Opus 5.5's best score of 45.5%, at roughly half the cost per task, even though Opus 5.5 costs half as much per token. At Xhigh effort Fable 5.2 reached 49.1%, the highest score in Anthropic's table.

The other cost charts point the same way. On Perplexity's WANDR benchmark, Fable 5.2 at Medium matched Opus 5.5's top score for about 40% of the cost per task; on Humanity's Last Exam with tools, Fable at Low matched it for under a tenth of the cost; on Chartography, Fable at Medium scored 66.4% against 64.4% for Opus 5.5 at Max.

All of these are Anthropic's own measurements, and each compares Fable at a lower effort with Opus 5.5's best result; only the Chartography chart names Max as Opus 5.5's setting. Fable does not win everywhere in the launch table: on AutomationBench, Gemini 4 Carbon scored 53.9% against Fable 5.2's 51.4%.

Effort is the main cost dial

Fable 5.2 has five effort levels: low, medium, high, xhigh and max. Adaptive thinking is always on, so effort is the main control over thinking, latency and cost. The API default is high; the Claude apps and Claude Code default to medium. In Claude Code, change it with /effort or the /model picker.

Anthropic's effort docs advise starting at high, moving up to xhigh only for your hardest long-running coding and agent work where you have measured a gain, and dropping to medium or low wherever your evals show quality holds. Max is for measuring the ceiling: slower and pricier than xhigh, often for a small gain. Effort barely changes reply length, so use the prompt for that. At high and above, set a large max_tokens, which caps thinking and reply together; the 64000 below is our example, not an official figure. Re-run an effort sweep rather than reusing Fable 5.1 settings.

{
  "model": "claude-fable-5-2",
  "max_tokens": 64000,
  "output_config": {"effort": "medium"},
  "messages": [{"role": "user", "content": "..."}]
}

On supported platforms you can change effort per message without losing the prompt cache, which in principle lets one long session run routine turns cheaply and spend more only on hard ones.

Fast mode doubles the rate

Fast mode runs the same model on a faster inference setup: up to 2.5 times more output tokens per second (time to first token does not improve) for $16 per million input and $80 per million output. On the Claude API, set speed to "fast"; the fast mode docs list which models need a beta header, and Microsoft Foundry uses separate fast model IDs instead. Fast mode has its own rate limit (a 429 when exceeded), does not work with the Batch API, and doesn't share cached prefixes with standard speed, so falling back means a cache miss.

In Claude Code, toggle it with /fast. On Pro, Max, Ultra, Team and Enterprise plans it is paid only from usage credits, even if your plan has usage left, and Team and Enterprise need an Owner to enable it. It is not available in Claude Code on Amazon Bedrock, Microsoft Foundry or Claude Platform on AWS.

A gotcha for subscribers

Depending on your plan and seat tier, Fable usage in Claude Code can bill to usage credits instead of your plan's limits; the /model picker then shows "Requires usage credits" on the Fable row, and an interactive session can show a consent prompt. In background, Remote Control and agent-team sessions, Claude Code holds that prompt for five minutes by default, then ends the turn without sending the request. Runs with -p never ask: if a Fable request would bill to credits, Claude Code bills it. And in Claude apps gateway sessions the fable alias resolves to Fable 5, not 5.2; use claude-fable-5-2 there. Check both before pointing scripts or CI at Fable.

Our take

The 20% cut is real, but Anthropic's charts suggest the bigger saving comes from effort, and the newer tokenizer eats into the headline figure. Anthropic itself recommends starting most coding and knowledge work on Opus 5.5 and moving up to Fable only for the most ambitious, open-ended jobs. If you run Opus 5.5 at Max to get acceptable results, Fable 5.2 at High or Medium may do the same work for less. If Opus 5.5 at Medium already does the job, Fable's doubled per-token price is hard to justify; in our reading the launch charts don't directly compare that case. Run a small effort sweep on your own tasks, compare cost per finished task, and keep fast mode for moments when waiting costs more than tokens.

Source: Claude Fable 5.2 and Mythos 5.2 (Anthropic), Pricing (Claude Platform Docs), Effort (Claude Platform Docs), Fast mode (Claude Platform Docs), Speed up responses with fast mode (Claude Code Docs), Model configuration (Claude Code Docs), Anthropic releases new Mythos and Fable models amid RSI concerns (TechCrunch), Anthropic releases first frontier models since Amodei's call for slowdown as IPO buzz builds (CNBC)

This story first appeared in the claudefox daily briefing for November 10, 2026.