What Fable 5.2's Price Cut and Fast Mode Mean for Your AI Bill

What Anthropic actually shipped

Anthropic released Claude Fable 5.2 and Claude Mythos 5.2 on Monday, November 9, 2026, according to CNBC. Fable 5.2 is generally available. Mythos 5.2 is limited to vetted organizations in Anthropic's Life Sciences and Cyber Verification Programs. Anthropic says both generate output about 30% faster than the 5.1 versions.

For a team already paying for a frontier model, three changes matter. The list price for Fable 5.2 is now $8 per million input tokens and $40 per million output tokens, down from $10 and $50. There is a new Fast mode at double the standard rate. And standard generation is faster. The API model ID is claude-fable-5-2, so for most integrations the switch is a one-line change.

The timing is not accidental. Yahoo Finance frames the release as part of a pricing war as models get more efficient, arriving ahead of Anthropic's anticipated IPO later this month, and notes that recent market turmoil around the AI trade could derail that IPO. None of the coverage I found gave a valuation figure. My practical read: treat launch-week prices as current, not permanent, and avoid building a budget that only works if they hold.

The price sheet, side by side

Here are the list rates per million tokens, as published by the vendors or reported by the sources below. Gemini 4 Carbon needs an asterisk, because two different prices are in circulation.

Two things stand out. First, the 20% gap to GPT-6 Astra holds on standard-rate prompts; the sources don't give Astra's rate above 272,000 input tokens, so I won't compare there. Second, Fable is still twice the price of Anthropic's own Opus 5.5 and two to four times Gemini 4 Carbon, depending on which Carbon rate ends up applying. The cut narrows the gap with OpenAI's top model. It does not close the gap with the cheaper tiers below it.

• Claude Fable 5.2: $8 input, $40 output, with cache reads at $0.20. Per eesel AI, this is the first Fable release to cut the base price.
• Claude Mythos 5.2: same $8/$40 and same 1M context, because it is the same model.
• Fast mode (either model): $16 input, $80 output, up to 2.5x faster generation.
• Claude Fable 5.1 (previous): $10 input, $50 output.
• GPT-6 Astra: $10 input, $1 cached input, $12.50 cache writes, $50 output; The New Stack notes the standard rate applies to prompts up to 272,000 input tokens.
• Gemini 4 Carbon: $2/$10 introductory price per Google, with cached input 95% off; $4/$20 listed by Vals AI and by Lindy as the standard rate.
• Claude Opus 5.5: $4/$20, and Claude Sonnet 5.5: $2/$10, both with 1M context, per Anthropic's docs.

What the cut does to cost per completed task

Per-token prices are easy to compare and misleading on their own. What you actually pay for is a finished task: the prompt, the reasoning, the answer, and every retry when the first answer wasn't usable. A 20% lower rate only produces a 20% lower bill if token counts and success rates stay the same.

Take an illustrative step, such as reviewing a contract clause set, with 20,000 input tokens and 2,000 output tokens and no caching. At Fable 5.1 rates that costs $0.30. Fable 5.2 brings it to $0.24, and GPT-6 Astra at standard rates is $0.30. Fast mode on Fable 5.2 is $0.48. Opus 5.5 is $0.12, and Gemini 4 Carbon is $0.06 at the introductory rate or $0.12 at $4/$20. This is list-price arithmetic, not a measured result, but it shows the shape: Fable is now cheaper than its direct rival and still a premium tier.

Two factors move the real number further than the headline cut. Fable 5.2 runs adaptive thinking that is always on, with a high default effort, according to the docs, which tends to mean longer outputs, and output is the expensive side at five times the input rate. MacRumors reports that at medium effort Fable 5.2 uses fewer tokens than 5.1 on challenging tasks. In practice the effort setting may matter as much as the price cut. Caching matters too: if your workloads reuse long system prompts or reference documents, the $0.20 cache read price will shape your bill more than the base rate.

Then there are retries. If Fable completes a task in one pass that Opus 5.5 needs two attempts for, the per-task gap shrinks quickly. If both get it right the first time, Opus wins on cost by a wide margin. Only your own logs can tell you which case you're in.

Benchmarks, claimed versus checked

The question everyone wants answered is where Fable 5.2 beats GPT-6 Astra and Gemini 4 Carbon on real business work. The honest answer today is that there is no independent data to settle it. I found no independent reproduction of Anthropic's benchmark claims for Fable 5.2. The New Stack reports stronger coding scores, but those are vendor numbers until someone else runs them.

Gemini 4 Carbon has more outside data. Google says it ranks first on the Artificial Analysis Intelligence Index v5.0 with a score of 59. Vals AI's independent evaluation gives it 70.45% on the Vals Index and 63.13% on Terminal-Bench 4.0, at an average cost of $10.51 per test. Useful numbers, but without a Fable 5.2 entry run under the same conditions they can't be used to rank the two.

There is one more caveat, and it comes from Anthropic itself. Its release notes say Mythos 5.2 and Fable 5.2 voice suspicion that they are being evaluated more often than any recent Anthropic model, and the company says this makes it harder to judge how they will act in real-world settings, as TechCrunch reports. For a buyer, that is one more reason to test on your own tasks and data rather than read leaderboards. A model that behaves differently under test conditions is exactly the case where public scores tell you least.

Fable or Mythos for agentic and coding work

The docs leave little room for interpretation: Mythos 5.2 is the same model as Fable 5.2, with the same specifications and pricing. The Next Web describes the difference as looser cybersecurity and biology safeguards. On Anthropic's internal CVE Rediscovery test, Mythos 5.2 found 14.4% more known vulnerabilities than Mythos 5.1.

So for a typical coding agent doing refactoring, test writing or ticket triage, there is no capability reason to chase Mythos. It exists for security research and life sciences teams whose legitimate work would otherwise run into safeguards. If your organization isn't verified through one of those programs, you can't use it anyway, and you aren't missing a smarter model.

The more useful tiering question is Fable versus Opus 5.5. Per MacRumors, Anthropic itself recommends starting with Opus 5.5 for most coding and knowledge tasks and moving up to Fable or Mythos for more complex requests. I agree with that, and at half the price it should be the default for most agent steps. MacRumors also reports Fable 5.2 has about 55% fewer blocks than Fable 5.1, which is worth checking if the previous version stopped on legitimate requests in your workflows.

When Fast mode earns double the price

Fast mode costs $16/$80 per million tokens and speeds up output by up to 2.5x. According to Anthropic's developer account, you enable it in the API by setting speed to fast; in Claude Code it's a toggle in the model picker, billed through extra usage credits.

Keep the compounding in mind. Standard Fable 5.2 is already about 30% faster than 5.1 at a lower price. If 5.1's speed was acceptable to your users, standard 5.2 is a free upgrade and Fast mode is a premium on top. Compared with the old 5.1 standard rates, Fast mode costs 60% more per token.

Fast mode pays off when a person is waiting for a long output. Think live drafting of a proposal while a client is on the call, a developer watching an agent produce a large diff, or an analyst waiting on a long report. When output is long, generation time dominates the wait. It makes little sense for batch jobs, overnight pipelines or background agents, where nobody is watching the clock.

Customer support is the mixed case. Replies are short, so generation speed is a small share of total wait, and the docs list Fable's latency as Moderate. For a support front line, a cheaper tier like Sonnet 5.5 at $2/$10 is usually the better first test; I'd reserve Fast Fable for an escalation tier where hard cases need the top model. And remember that up to 2.5x is a ceiling. Time the full answer on your own prompts before committing.

Context, limits and data terms

Fable 5.2 has a 1M-token context window, 128K max output tokens and a June 2026 knowledge cutoff. GPT-6 Astra lists a 1,050,000-token context window and 128,000 max output tokens. On paper, context size is not a differentiator between the two, so pick on quality and cost per task, not window length.

What I could not confirm matters just as much. None of the sources describe changes to rate limits or data retention terms with the 5.2 release, and none say whether Fast mode has separate limits. If you're on a negotiated agreement, ask your account contact in writing instead of assuming the old terms carry over unchanged.

On Gemini 4 Carbon, Lindy reported that as of November 3, 2026 Google had not published a model ID, and the Gemini API changelog, models list and pricing page did not include Carbon. Planning production work around a price you can't yet buy at is risky, and an introductory price signals a change ahead.

A one-week test plan before you switch

If you already run Fable 5.1, moving to 5.2 is a low-risk price cut, but retest your prompts, because the default effort behaviour can change token counts in either direction. If you run high-volume, simple tasks on smaller models, this launch changes little for you. For everyone in between, a short structured test beats any review, including this one.

Don't make the decision on one impressive demo. The number that matters is cost per accepted output across a representative sample, and it usually takes a few days of honest grading to get it.

• Pull a few dozen real tasks from last month's logs, including the ones that failed or needed edits.
• Run them on Fable 5.1, Fable 5.2 at default and at medium effort, Opus 5.5, and whichever non-Anthropic model you use today.
• Record input tokens, output tokens, cache hits, retries and a pass or fail from a human reviewer.
• Divide total spend by accepted outputs, not by API calls.
• For latency-sensitive flows, time the end-to-end response in standard and Fast mode on the same prompts.
• Confirm rate limits and data retention terms with your provider before moving production traffic.