admin The company said Fable 5 would remain available on Pro, Max, Team and eligible Enterprise subscriptions for up to half of each user’s weekly allowance. Claude Code customers will also retain weekly usage limits that are 50 per cent higher until the new deadline, set for 11.59pm Pacific Time.
The one-week extension is the third change to the access timetable since Fable 5 returned to Anthropic’s services at the start of July. The initial inclusive period was due to expire on July 7 before being moved to July 12 and then July 19, giving subscribers additional time to test the company’s most capable generally available model without buying separate usage credits.
Anthropic has not indicated whether access will be extended again. Under the planned arrangement, subscribers who want to continue using Fable 5 after exhausting any inclusive allowance will have to pay through usage credits, bringing the consumer experience closer to the token-based billing already applied to developers using the Claude application programming interface.
Fable 5 costs $10 per million input tokens and $50 per million output tokens through the API, twice the corresponding rates for Claude Opus 4.8. It supports a one-million-token context window, output of up to 128,000 tokens and adaptive reasoning designed for long-running coding, research and enterprise tasks.
The higher cost reflects the computing resources needed to operate frontier models over large codebases and extended autonomous sessions. It also illustrates the financial challenge facing artificial intelligence companies as subscription customers increasingly use models for workloads that can consume millions of tokens during a single project.
The extension may help Anthropic measure demand and ease subscriber concerns before separating access to Fable 5 from fixed-price plans. Users accustomed to unlimited-looking interfaces have criticised additional metering across the AI industry, particularly when agentic coding tools can generate charges without the predictable interaction pattern of conventional chatbots.
Fable 5 was launched on June 9 alongside Claude Mythos 5, a closely related model with fewer restrictions in selected cybersecurity and scientific areas. Mythos access remains limited to approved organisations, while Fable includes safeguards that may redirect flagged biology or cybersecurity requests to Opus 4.8.
Anthropic temporarily suspended both models in June after US export controls restricted their availability to foreign nationals. Fable 5 was restored globally on July 1 after those restrictions were lifted, beginning the sequence of promotional access windows now extended through July 19.
Attention has also shifted to an unidentified model labelled “Claude Honeycomb EAP” that briefly appeared as an option for some Cursor users. EAP commonly refers to an early-access programme, but neither Anthropic nor Cursor has publicly identified Honeycomb as a forthcoming commercial model.
Online claims linking Honeycomb to a supposed Claude Opus 5 release remain unverified. Anthropic has announced no Opus 5 model, release date or pricing, and the company’s official product catalogue continues to list Opus 4.8 as the current model in the Opus family.
Cursor forum users reported seeing Honeycomb with a large context capacity and an extra-high reasoning setting before the listing disappeared. Temporary model names are often used for internal testing, staged deployments and evaluation builds, meaning the appearance alone does not establish that a public launch is imminent.
The episode nevertheless highlights the role of coding platforms in revealing changes within the competitive AI market. Cursor integrates models from several providers and regularly tests unreleased configurations, creating opportunities for model identifiers or experimental settings to become visible before formal announcements.
Anthropic is competing with OpenAI, Google, xAI and specialist coding providers for developers building software through autonomous agents. Fable 5 has been positioned for complex migrations, multi-stage engineering projects and professional work that requires prolonged reasoning with limited supervision.
The company says the model can operate across large repositories, generate and run tests, inspect visual outputs and revise its own work. Its safeguards, mandatory 30-day data retention and higher token charges may limit adoption among organisations with strict privacy requirements or workloads where lower-cost models provide adequate performance.
