Anthropic released Claude Fable 5.2 and Claude Mythos 5.2 on November 9, introducing two models for software development, analysis and agents carrying out complex technical work. The update brings lower prices than the 5.1 generation and compatibility changes for applications connected through the API. Access and data handling rules differ from those applying to the regular models in the portfolio.
In the technical catalogue, Fable 5.2 has a context window of one million tokens and a generation limit of 128,000 tokens. It accepts text and images and produces text. It is available through the Claude API, Amazon Bedrock, Claude services on Google Cloud and Microsoft Foundry. The direct developer identifier is claude-fable-5-2; the Bedrock integration uses its own identifiers, published separately for that platform.
The same underlying model, with different access conditions
According to Anthropic, Fable and Mythos use the same underlying model. Fable includes additional safeguards for cybersecurity activities and biological research, while Mythos is offered through verification programmes. The Fable 5.2 documentation confirms that Mythos 5.2 has the same specifications and pricing, but its availability is limited to organisations approved for the relevant area of use.
The portfolio positioning is explicit: Anthropic recommends Opus 5.5 as the starting point for most work and Fable 5.2 for demanding reasoning and agents operating over long periods. Sonnet 5.5 is presented as an option balancing speed and capability, while Haiku 5.5 targets classification, extraction and routing of large volumes of requests. The current models support tool use and multilingual inputs.
Base pricing and fast mode have separate calculations
For Fable 5.2 and Mythos 5.2, published base API pricing is $8 per million input tokens and $40 per million output tokens. The corresponding figures for the 5.1 versions are $10 and $50. Batch processing for the new models costs $4 for input and $20 for output, measured on the same per-million-token basis.
Fast mode runs the same model with a different inference configuration. According to the documentation, the increase of up to 2.5 times concerns the number of output tokens generated per second, rather than time to the first token. Direct pricing is $16 per million input tokens and $80 for output. Fast mode is unavailable in the Batch API; cloud platforms have their own delivery and pricing conditions.
Another control is effort level, which influences reasoning, token consumption and response time. Fable 5.2 supports five levels, from low to max, with high as the API default. Adaptive thinking remains active. The output limit includes both thinking and response text, so a small allowance can be exhausted before any text intended for the user appears.
Three distinct roles in an agent application
- Execution
The main model works on the task - Consultation
Advisor returns guidance based on the conversation - Delegation, in beta
Managed Agents distributes work and tracks usage
Compatibility also concerns the conversation history
For applications building their own API requests, changing the model involves checking how the history is preserved. The new versions validate the content preceding thinking blocks. Changes to system instructions, tools or earlier messages can invalidate those blocks and produce an HTTP 400 error. The documentation specifies that integrations must preserve the conversation by appending new messages without rewriting those already used.
There are also differences in computer operation tools. The older computer_20251124 tool is rejected by Fable 5.2 on the platforms listed by Anthropic, except Amazon Bedrock; the alternative is the new toolset. Fable 5.1 also cannot read thinking blocks produced by Fable 5.2. When switching back to the previous version, the API removes incompatible blocks and the conversation continues without that reasoning.
The advisory role is separate from actual execution. In Claude Code, the experimental Advisor tool lets the main model request an assessment from another model at important moments. The advisor receives the complete conversation, including tool calls and their results, and then returns guidance. Each consultation consumes additional tokens, billed at the advisor model’s rates. The feature has its own platform and model compatibility requirements.
In Managed Agents, multiagent orchestration is in beta. A session can delegate work to subagents, run workflows combining results from several agents, or consult an Advisor. The configuration determines which models and tools are available, while delegated agents work with their own contexts. Each agent’s tokens count towards session consumption; the session budget can cap total spending, including workflows executed in the background.
Access verification and data retention
For security work, the Cyber Verification Program separates defensive access, authorised offensive testing and specialised activities. Approval for internal use covers the organisation’s own code, products and infrastructure. Providing services to customers requires separate approval. The programme accepts applications for managed services in which staff review deliverables, while the customer neither interacts directly with the model nor receives its raw outputs.
Fable 5.2 and Mythos 5.2 are designated Covered Models. The standard policy requires requests and responses to be retained for at least 30 days, with automated safety analysis. Deletion follows that period, subject to the stated exceptions for investigations or legal requirements. These conditions follow the model onto third-party platforms; a zero data retention configuration requires the applicable express authorisation.
Separately, Enterprise Frontier Safeguards allows monitoring data to be retained in cloud infrastructure controlled by the customer, under its own encryption keys and access policies. Automated systems identify patterns requiring attention, and the signals are sent to the customer’s team for review. When announcing the new models, Anthropic stated that this solution was in private beta, with access requested through the account team or a dedicated form.
Storage, encryption and automated review controls are optional within the Enterprise Frontier Safeguards architecture. Anthropic states that they do not change model behaviour, API pricing or usage limits, and that it does not charge a fee for the solution. When data is retained in the customer’s cloud account, the infrastructure provider bills for storage, reads, writes and data transfers according to its own rates.
