Anthropic Releases Claude Fable 5 and Mythos 5: Two New Models

Anthropic just shipped two new AI models: Claude Fable 5 and Claude Mythos 5. It's a significant release for anyone tracking AI, and we're taking the time to explain it plainly, without assuming you follow AI news every day.

For context: Anthropic is the US company that builds Claude, an AI assistant directly competing with ChatGPT. Claude comes in several "families" (Opus, Sonnet, Haiku), a bit like a car comes in different trims. This week, Anthropic adds two new families: Fable and Mythos. Both jump to version 5.

---

Fable and Mythos: two models for two uses

The official announcement presents the two models as a pair. Mythos 5 targets the most demanding tasks: complex reasoning, science, law, biology. Fable 5 is built to be an excellent programming and automation companion at a more accessible cost.

Anthropic clarifies in its methodology note: "reported scores are within a 1-to-3 percentage point difference for both models". In plain terms, they are very close in performance. The differences play out mostly in price and intended use.

To frame this: both models beat the current best competitors (OpenAI's GPT 5.5 and Google's Gemini 3.1 Pro) on most published benchmarks. That's the headline takeaway.

What it changes in practice

The chart below illustrates the gap on one of the most-watched evaluations: agent-driven coding. This is where Fable 5 shines the brightest:

Without diving into jargon, look at the orange bars (Fable) on the left: on the first benchmark, Fable 5 succeeds at eight programming tasks out of ten where GPT 5.5 succeeds at fewer than six. On the second, which is harder, the gap widens further.

Beyond code, Mythos 5 and Fable 5 make notable progress in scientific reasoning, legal analysis, and cybersecurity. On the "with tools" evaluations, Mythos 5 solves 64.5% of Humanity's Last Exam questions, more than twelve points ahead of GPT 5.5.

💡 Generative AI is entering a phase where the measure is no longer speed, but reliability. A model that succeeds at eight tasks out of ten changes the nature of the work you can hand to it.

Two demos that beat any explanation

Anthropic released two short videos that show better than any text what Fable 5 can do. The first: the model plays Pokémon FireRed using only vision, no text guidance.

The second: the model simulates the solar system and predicts an eclipse, starting from general knowledge only. That's the kind of task that requires chaining math, code, and physics. It shows where Mythos 5 has pushed the scientific reasoning bar.

---

Should you care if you're not a developer?

Yes, and here's why. When AI gains ten to twenty points on answer quality, this isn't only an engineers' concern. It changes what your AI assistant can do in an ordinary workday: read a legal document, prepare a market analysis, build a presentation from scattered notes.

Mythos 5 also scores better on "aligned behavior" evaluations, where a lower score is better. Mythos 5 reaches 2.06 versus 2.81 for the previous generation (Sonnet 4.6). The model strays less, which matters as much as raw speed.

On the practical side, both models are available right now through Anthropic's API, on Claude.ai, in Claude Code, on Amazon Bedrock and Google Cloud Vertex AI. If you already use these platforms, you get access just by changing the model name in your settings.

The AB-Arts take

At AB-Arts, we track these releases because they shape what we can build with and for our clients. Our Claude masterclass is built specifically to help non-technical teams hold the tool with precision. Fable 5 and Mythos 5 don't change the teaching method, but they significantly widen what teams can accomplish coming out of the training.

To dig deeper on the meaning of this acceleration, read our piece on Dario Amodei's prediction that AI would write 90% of code. With Fable 5, what looked bold fifteen months ago enters the measurable concrete.

→ To train your team on driving Claude through your real pipelines, browse our Claude masterclasses. For tailored guidance in your studio or organization's context, reach out via our contact page.