Claude Fable 5.2: $8 per Million Tokens, Access Only With 30 Days of Retention

Anthropic released Claude Fable 5.2 and Mythos 5.2 on 9 November: 30 hours of work on the Fermat proof, two solved problems in black-hole physics, unconfirmed from outside at first. Two things were measurable on launch day: the price sits 20 percent under its predecessor, and access only works with a duty to keep data for 30 days. Both hit our gateway runs, worked through on our own token counts.

What Anthropic ships and who confirms it

Fable 5.2 goes to the Claude apps, the Claude Platform, Bedrock, Google Cloud and Foundry, Mythos 5.2 only to approved partners. Anthropic writes, verbatim: "In early testing on the FrontierCode 2.0 evaluation, Fable 5.2 at High effort scored about level with Opus 5.5 at Max effort, for roughly half the cost per task." Early testing, High against Max effort, measured by Anthropic itself.

The post names the UK AI Security Institute, Frontier Design and METR as reviewers, footnote 4 withdraws the confirmation, verbatim: "None of the external organizations reviewed the contents of this blog post or the contents of the System Card. Those organizations do not necessarily agree with any of the claims made in either publication." Externally tested here means measured before release, nobody endorses the post. None of the three had a statement on 5.2 that evening. METR has been struggling since 5 November with traceability: Anthropic redacted most of the evidence METR would find most informative. The math results point to hexagonmath.org, nobody has reproduced them.

The price: $8 instead of $10, with four footnotes

List price per million tokens: $8 input, $40 output. The predecessor 5.1 stands at $10 and $50, both sides 20 percent lower. Cache reads cost $0.20, writes $10 with a five-minute hold and $16 with an hour, batch runs at half price.

First, Fast Mode: $16/$80 with no batch discount. The documentation promises "up to 2.5x higher output tokens per second", measured on output speed rather than on time to first token. Second, the counter: Fable 5.2 shares the tokenizer with every model since Opus 4.7, so a prompt counts roughly 30 percent higher. Third, the API reference: its ID list still named fable-5-1 on the launch evening, the example uses claude-fable-5-2. Fourth, the competitors: two of the three price fields carry the availability catch.

Meta announced Muse Ember the same day at $0.10 input, 80 times cheaper than Fable 5.2. The Astra catch: above 272,001 input tokens the input price of the whole request doubles, and that is where agentic load sits. No long-context surcharge is measured for the 5.2 models.

The retention duty behind model access

Anthropic keeps a list called Covered Models, both 5.2 models on it since launch day. The policy dates to 9 June 2026, five months before the model. Core, verbatim: "Data retention: 30-day minimum retention by default." On usage-policy violations data stays for up to two years, even under otherwise zero retention. And as an access condition: "In order to access this model, your organization or workspace must have data retention enabled." Without retention the API answers HTTP 400, before the first line of response.

Zero data retention is not bookable there, it is grantable: the documentation excludes workspaces, Claude Enterprise organizations and third-party platforms, "unless explicitly authorized". The three levers: the retention setting of the organization (at Enterprise only by the Primary Owner), a ZDR grant from Anthropic, Enterprise Frontier Safeguards from autumn 2026. On Bedrock and Google Cloud the prompts and outputs stay in the customer's cloud account, on Foundry the boundary runs per subscription. On staff: "By default, no Anthropic personnel can read your retained conversations."

Context: OpenAI does it the same way, its policies name up to 30 days of abuse logs by default, zero retention needs an approval. Retention is the industry's mode, not an Anthropic detour.

What the launch would cost across our gateway

Our own AI stack runs through a LiteLLM gateway in-house, open-weight models on our own hardware. The 30 complete days from 11 October to 9 November carry 176,380 requests across 6,333,231,339 input tokens and 75,801,053 output tokens; zero cache hits, zero dollars to third parties. At Fable 5.2 prices that is $50,666 for input and $3,032 for output, so $53,698 in 30 days, about $653,000 a year. At 5.1 prices it would be $67,122, the savings from the cut around $13,400 a month.

The $0.20 cache price only bites on standardised prompt prefixes. The embedding lane (1.15 million requests, 2.99 billion tokens) is netted out, no chat model runs there.

What stays open on day one

Artificial Analysis lists Fable 5.2 at index 65, first of 74, and at $26.54 per task it is also the priciest model on that board. Its label reads "Max, Default Fallback", Anthropic's own figures name other effort levels, so the comparison runs across different settings. LiveBench 2.0 reports 84.9 points at $9.85 per solved task, measured by the operator of the board who enters models himself. Vals AI names 71.31 percent at 43 minutes 42 seconds, the only latency figure from outside Anthropic. LMArena has no 5.2 row, the Aider board has written nothing new since 3 October 2025, both of those are empty measurements rather than verdicts. An output speed measured from outside is missing, and so is a reproduced vendor number.

On the data location the announcement stays quiet: data residency does not appear in the launch post. That matters since the GPAI duties of the AI Act have been running since 2 August 2026 and demand external evaluation. The Commissioner's speech of 14 October names external evaluation as an RFI topic for GPAI providers; the fine ceiling is 3 percent of turnover, no company named. OpenAI prices EU residency at a 10 percent surcharge and does not switch Fast Mode on for GPT-6 Astra there at all.

Our decision: Fable 5.2 does not enter our stack, if only because the retention question is open in our own tenant. We watch three things: independent measurements of output speed from 100k context, whether the EFS grant solves the ZDR question in practice, whether Anthropic says anything about an EU data location. The price cut moves the question anyway, from whether to when: at what monthly volume the cloud pays off.

Further Reading

• Anthropic: Claude Fable and Mythos 5.2 – launch post with the FrontierCode quote and footnote 4, retrieved 9 November 2026
• Anthropic pricing docs – list price, cache, fast mode and batch for both 5.2 models, as of 9 November 2026
• Fast mode docs – same intelligence, 2.5x output tokens per second, no batch discount
• Covered Models policy of 9 June 2026 – designation list, 30-day retention, access lock
• Data retention practices – ZDR exceptions, access protection, staff access
• METR assessment of 5 November 2026 – Anthropic's redaction of the evidence
• Artificial Analysis: Claude Fable 5.2 – index 65, $26.54 per task, as of 9 November 2026
• LiveBench – board state on 9 November, board publication on 6 November
• Commission speech SPEECH/26/2150 – external evaluation as an RFI topic, 14 October 2026