Anthropic releases Fable 5.2 and Mythos 5.2 with lower token prices Anthropic releases Fable 5.2 and Mythos 5.2 with lower token prices

The unit price falls; the work budget has more parts

The previous Fable 5.1 and Mythos 5.1 rates were $10 for a million input tokens and $50 for a million output tokens. At identical standard-rate usage, the new prices turn a $60 bill for one million tokens in each direction into $48. That calculation excludes caching, tools and other charges. Optional fast mode costs $16 for input and $80 for output per million tokens. The standard model and the speed option therefore have different economics. Anthropic still recommends starting most workloads with its latest Opus model, then moving to Fable when evaluations at higher Opus effort fall short.

The same underlying model has different access paths

Fable and Mythos share an underlying model, according to Anthropic. Fable includes cybersecurity and biology safeguards that support broader availability. Mythos is reserved for organizations admitted through verification programs for those areas. That distinction matters when a team chooses a model for security or scientific work: a general release does not provide every customer with the same permitted capabilities. Anthropic lists Fable availability across its own platform, Amazon Web Services, Google Cloud and Microsoft Azure. It describes both releases as intended for ambitious technical work, including coding and research.

Applications must preserve the conversation they send back

The migration guide describes a change that reaches beyond replacing the model name. Fable 5.2 checks whether earlier messages, system instructions or tool definitions changed before a returned thinking block. That check now applies to every account. Sending an invalid block back normally produces a request error. An append-only conversation retains earlier turns and adds new ones; a client that silently rewrites old context has different behavior. Model switching has a separate constraint. On the Claude API and Google-managed Claude service, other models cannot read Fable 5.2 thinking blocks, except Mythos 5.2. Those blocks are dropped when the conversation moves to an incompatible model.

Retention and effort remain deployment decisions

The model overview specifies a one-million-token context window and up to 128,000 output tokens. Adaptive thinking is always enabled, with high effort as the API default. Effort controls how much reasoning the model performs, so the token rate is only one component of a completed job. The overview also requires 30-day data retention unless Anthropic expressly authorizes a zero-retention arrangement. The launch describes Enterprise Frontier Safeguards, currently in private beta, as a route that runs automated monitoring on prompts and responses stored in customer infrastructure. That makes retention approval a separate deployment dependency from gaining access to the model itself.