What this means for makers and artists
If you build tools for creation, Claude Fable 5 represents a shift towards raw computational power rather than conversational polish. The model is notably slow and costly, yet it demonstrates an uncanny ability to churn through complex tasks without immediate failure. For developers and artists, the challenge is no longer finding what the model can do, but rather identifying the specific edge cases where it falters.
The safety trade-off
Anthropic has released two distinct versions today. Claude Fable 5 offers the same raw capabilities as the previous Mythos 5 iteration but operates under significantly stricter safety guardrails. These filters are active enough to trigger new API mechanisms that notify users when a request is blocked. Uniquely, the API now allows for an automatic fallback to a different model if a response is rejected.
Conversely, the newly launched Claude Mythos 5 shares the full capabilities of Fable 5 but explicitly lacks these safety classifiers.
Technical specifications and pricing
Both models share a context window of one million tokens and can generate a maximum of 128,000 output tokens. Their knowledge is current only up to January 2026.
The pricing structure has doubled compared to the Opus 4.5/4.6/4.7/4.8 series. Users pay $10 per million input tokens and $50 per million output tokens. Notably, the cost does not increase based on the length of the context window.
Documentation for the upgrade is notably sparse compared to the detailed guides provided for Opus 4.8.
A shift in model scale
The most distinct characteristic of Fable 5 is its sheer scale. It feels “big” not just in terms of processing speed and expense, but in the depth of its internal knowledge.
To illustrate this, a prompt was used to list all of Simon Willison’s open-source projects in reverse chronological order. When asked to perform this task without search tools, Opus 4.8 admitted it could not provide a reliable, date-verified list of every project. It listed a few well-known items like LLM, Datasette, sqlite-utils, and Django before stopping.
In contrast, Fable 5 generated a comprehensive list of fourteen projects, including specific release dates for tools like files-to-prompt (April 2024), symbex (June 2023), and shot-scraper (March 2022). It correctly identified that a truly complete list was impossible due to its knowledge cutoff, yet it provided far more detail than the previous model.
While the author previously argued that baking knowledge into weights is less useful than active search, the sheer volume of details retained is a strong proxy for model size. The speed, pricing, and demonstrated knowledge suggest Fable 5 is likely the largest model yet released by any vendor.
Integration across the ecosystem
Fable 5 is currently available across all Anthropic interfaces, including the Claude.ai chat interface, Claude Code for web, the CLI, and Claude Cowork. Access is included in subscription plans until June 22nd, after which it will incur extra charges.
Since September 2025, the Claude.ai interface has provided a full container environment. This allows for the installation of packages and cloning of repositories directly from GitHub.
In a recent test, the model was tasked with upgrading a project called micropython-wasm to run full Python instead of MicroPython. Fable identified that Brett Cannon’s cpython-wasi-build could be used for this purpose. Although the model could not download the necessary files due to environment restrictions, the user uploaded the zip files manually.
Following this, the model successfully compiled and tested a build. It produced a 13.9MB wheel file (cpython_wasm-0.1.0-py3-none-any.whl) containing the system, wrappers, and standard library. This wheel can be used with uv to run Python code in a sandbox environment.
Complex software engineering tasks
Before the release date, a goal was set to add a feature to Datasette Agent: the ability for tool calls to pause execution and request user approval.
Fable 5 not only solved this problem but also identified and fixed six underlying issues in the LLM library to support this pause-resume mechanism. The model initially relied on complex workarounds but quickly refactored them into supported features once the scope was clarified.
Key takeaways
- Claude Fable 5 is a massive, slow, and expensive model that prioritises raw capability and deep knowledge over conversational speed.
- The release introduces a dual-track strategy: the safety-restricted Fable 5 and the unrestricted Mythos 5, both with identical underlying performance.
- Practical testing shows the model can successfully refactor complex software projects and integrate new features into existing toolchains without search assistance.
Stay ahead of AI. Get the most important stories delivered to your inbox — no spam, no noise.
