Claude Fable 5.2: a lower price, four changes that can break code, and 30-day data retention

Anthropic released Claude Fable 5.2 and Claude Mythos 5.2 on 9 November 2026, and says the two are the same underlying model. Fable carries safeguards for cybersecurity and biology, so Anthropic offers it to all users. Mythos is for organisations verified through Anthropic's Cyber Verification Program or Life Sciences Verification Program.

Both cost $8 per million input tokens and $40 per million output tokens. A token is the small unit of text that API use is billed in. Fable 5.1 and Mythos 5.1 cost $10 and $50, so the new price is 20% lower on both, by our arithmetic.

The prompt cache stores a repeated part of a prompt so it is not paid for in full each time. Reading from it costs $0.20 per million tokens, and writing to it costs $10 for five minutes or $16 for an hour. The Batch API, for work that can wait, costs $4 and $20.

Fast mode costs $16 and $80 per million tokens, twice the standard price by our arithmetic. Anthropic says it generates output up to 2.5 times faster. It does not work with the Batch API, and Amazon Bedrock is not listed among its platforms.

Anthropic says the 5.2 models write output up to 30% faster than 5.1. Each reads up to 1 million tokens in one request and writes up to 128,000, with no extra charge for long prompts.

On the same pricing page, Claude Opus 5.5 costs $4 and $20, and Claude Sonnet 5.5 costs $2 and $10. Fable 5.2 is twice the Opus 5.5 rate per token, by our arithmetic. Anthropic recommends starting most coding and knowledge work on Opus 5.5, which it calls "our strongest daily driver". It suggests stepping up only when tests on Opus at higher effort fall short.

Effort is a setting for how much the model thinks before it answers. On Anthropic's own FrontierCode 2.0 coding test, Fable 5.2 at high effort scored 45.3%, against 45.5% for Opus 5.5 at maximum effort. Anthropic says Fable did that for roughly half the cost per task, and at its second-highest effort it scored 49.1%.

Anthropic says Fable 5.2's safeguards block most cybersecurity tasks, such as penetration testing and exploit writing, and most research biology. A blocked request gets its answer from an older model instead: Opus 4.8 for cybersecurity and Opus 5 for biology. Anthropic says early testing saw about 55% fewer blocks than on Fable 5.1.

On subscription plans, Pro plans and standard Team seats get Fable models within the plan from 9 to 16 November 2026. That covers up to half their weekly limits, plus one limit reset. After that, Fable use is billed to usage credits if they are turned on. Anthropic's Help Center says Fable models use up limits faster than other Claude models.

Teams already on Fable 5.1 are likely to gain most from the cut. Each token costs 20% less from the day they switch, once the code changes in the next section are made.

For a team on Opus 5.5, Fable 5.2 costs $8 and $40 against $4 and $20. Anthropic's half-the-cost claim rests on its own test, where Fable at high effort matched Opus at maximum effort. Run both models on your own tasks and compare the bill per task to see whether that holds for your work.

Anthropic describes two ways to use Fable beside cheaper models. As an advisor, it answers questions from smaller models, and as an orchestrator, it hands simpler tasks to them. Either set-up can keep the $40 output rate to the few steps that need Fable.

For security teams, Anthropic's account means most security testing requests on Fable 5.2 are answered by an older Opus model. Mythos 5.2 needs verification through Anthropic's programme first. Claude Enterprise customers outside that programme can use Claude Security, which scans code for weaknesses and now uses Mythos 5.2 among its models.

For Team admins, the launch week can show who reaches for Fable before any usage credits are spent. Max, Ultra and premium seats get one to three limit resets, depending on how long the subscription has run. Those resets expire on 8 March 2027. Standard seats on legacy seat-based Enterprise plans, self-serve Enterprise and usage-based Enterprise plans get no launch offer.

Anthropic's documentation lists four changes from Fable 5.1 that can stop working code. The model's reasoning from earlier turns is sent back to it in later requests, and two of the four changes concern how that works.

First, Fable 5.2 checks on every account that the start of a conversation has not been edited. If code changes the instructions, the tool list or an earlier message, then sends back reasoning from a later turn, the request fails. On Fable 5.1 this check applied by default only to accounts created on or after 31 August 2026. Code on an older account can run cleanly on 5.1 and fail after the switch.

Nothing changes where Claude Code, claude.ai, Claude Managed Agents or the Claude Agent SDK builds the requests. Anthropic offers a stopgap that drops the rejected reasoning instead of failing, and warns it can raise output tokens.

Second, only Fable 5.2 and Mythos 5.2 can read Fable 5.2's reasoning. When a conversation moves from 5.2 to another model, including Opus 5.5 and Fable 5.1, that reasoning is dropped. The request still succeeds, and the dropped reasoning is not billed.

Third, computer use, where the model works a screen, keyboard and mouse, runs only through Anthropic's newer computer toolset. The earlier tool returns an error on every platform except Amazon Bedrock.

Fourth, in the advisor tool (in beta), a Fable 5.2 or Mythos 5.2 model takes only Fable 5.2 or Mythos 5.2 as its advisor. Every model accepts Fable 5.2 as an advisor.

Some limits carry over from Fable 5.1. Thinking is always on and cannot be given a fixed budget, and forcing the model to call a set tool returns an error. Organisations that manage their own encryption keys cannot use structured outputs, which hold answers to a fixed format, on Fable 5.2 or Mythos 5.2.

Fable 5.1 is not deprecated, and Anthropic will not retire it before 1 September 2027. Fable 5.2 stays until at least 9 November 2027.

The first change is likely to carry the most risk. Agents that trim or rewrite earlier turns to save tokens can fail on Fable 5.2, even where they ran cleanly on 5.1.

The stopgap buys time, at the cost of more output tokens. A lasting fix is likely to mean code that leaves earlier turns alone once reasoning has been sent back.

The second change affects any system that moves a conversation between models, for example falling back to Opus 5.5 when Fable is busy. The request succeeds, so nothing raises an alarm, but the second model starts without Fable's reasoning.

Fable 5.1 stays until at least 1 September 2027. Teams can move one workload at a time, starting with the ones that gain most from the lower price.

Anthropic treats Fable 5.2 and Mythos 5.2 as Covered Models, as it did Fable 5.1, Fable 5 and their Mythos versions. Covered Models need 30-day data retention so Anthropic can run safety monitoring. Zero data retention (ZDR), under which prompts and responses are not stored, is not available for them unless Anthropic expressly authorises it.

Where the data sits depends on the route. On the Claude API, including Claude Platform on AWS, Anthropic holds the retained data. On Amazon Bedrock and Google Cloud's Agent Platform, it stays inside the cloud provider's environment.

An organisation with a ZDR arrangement can turn on 30-day retention for one workspace and keep ZDR in the others. The Default Workspace always follows the organisation's setting.

Anthropic is building Enterprise Frontier Safeguards, safety monitoring that runs on prompts and responses stored in the customer's own systems, so it works with ZDR. It is in private beta, and Enterprise customers ask for access through their account team or a form. Until it is ready, Anthropic says eligible customers can keep using both models with ZDR. AWS says the same for Fable 5.2.

In GitHub Copilot, Business and Enterprise admins must turn on a Claude Fable 5.2 policy, which is off by default. GitHub says turning it on accepts the retention requirement, and that Anthropic does not train its models on retained data. Copilot Pro+, Max, Business and Enterprise get the model in a gradual rollout, billed at the provider's list price under usage-based billing.

GitHub says eligible customers can use Fable 5.2, 5.1 and 5 with ZDR until the end of the calendar year, under a time-bound exemption. After that, they will need Enterprise Frontier Safeguards. GitHub sends customers to their account team to find out whether they qualify.

Our reading is that retention is the main question for legal and security teams. Where a contract or AI policy promises zero retention, Fable 5.2 falls outside that promise unless the customer qualifies for the exemption.

The single-workspace option lets an organisation trial Fable 5.2 in one place, with known data, while the rest of the account keeps ZDR.

GitHub treats turning on the Copilot policy as accepting the retention requirement, so one admin's click accepts it for the whole organisation. The person who signs off data terms should approve that first.

GitHub's exemption runs to the end of the calendar year. A buyer relying on it can plan now for Enterprise Frontier Safeguards, or for 30-day retention, in January.