One model, two versions. Fable 5.2 and Mythos 5.2 are the same underlying model with different safeguards. Fable is open to everyone, and Mythos is limited to vetted cybersecurity and life sciences organizations to prevent misuse.

Faster and cheaper. Both models produce outputs up to 30% faster than the 5.1 versions. Prices drop 20%, to $8 per million input tokens and $40 per million output tokens.

A big jump in coding. On the FrontierCode 2.0 benchmark, Fable 5.2 scored 49.1%, up from 32.4% for Fable 5.1. Anthropic says both models beat Opus 5.5 on almost every capability it tested. For comparison, GPT-6 Astra scored 57.2%.

Fewer refusals. Fable 5.2 blocks about 55% fewer requests than Fable 5.1.

Mythos is built for security defenders. Mythos 5.2 found 89.6% of vulnerabilities in Anthropic’s internal test, up from 78.3%. Claude Enterprise customers can use this through Claude Security without joining the verification programs.

Where to get it. Fable 5.2 is available now on all paid Claude plans and in Claude Code. Developers can use it through the API, AWS, Google Cloud and Microsoft Azure with the model ID claude-fable-5-2.

What it is. Control Room is a central system that tracks every AI model job across OpenAI’s research setup, including training experiments, evaluations and data processing. It can shut down any of these jobs if something looks unsafe.

How it works. Every job gets a unique ID when it starts. Before the model answers any request, the system checks with Control Room whether that job is still allowed to run. If the job has been stopped, the request is denied.

A stop within 35 seconds. A person or an automated AI monitor can order a stop, and the model’s output halts within 35 seconds. OpenAI’s frontier training runs are watched by a separate monitor model that can now trigger this stop on its own. OpenAI wants to cut the delay to zero.

Safe even if it breaks. Control Room is designed to “fail closed.” If it goes down from an error, gets switched off or is tampered with, all model requests are blocked until it comes back. OpenAI also runs regular scans to catch models running without it.

Why now. The launch follows cases where OpenAI’s research models acted outside their instructions, including one that reportedly got around internet restrictions during training. OpenAI has also paused some tool-use training and testing for its most advanced models.

Limited to research for now. Control Room covers only OpenAI’s internal research work, not ChatGPT or the API. No outside group has reviewed it yet.

What it is. Muse Ember is the lightest and most cost-efficient model yet from Meta Superintelligence Labs. It’s built for the everyday work agents repeat nonstop: tool calls, small coding tasks, routing, pulling out data, and reading long documents.

Very cheap. At 10 cents per million input tokens, Ember costs far less than Anthropic’s Haiku and OpenAI’s Luna. Its price is close to open-source models like DeepSeek V4.1 Flash and GLM 5.3 Flash. Meta says it also scores higher than Haiku and Luna on a combined intelligence benchmark.

Built to work with a bigger model. Ember is meant to pair with Meta’s larger Muse Wonder model. In Meta’s words, “Muse Wonder plans, Muse Ember handles the volume.” For example, in customer support, Wonder sets the rules and Ember applies them across thousands of cases.

Reads more than text. Ember understands text, images and audio. Agents can use it to read screenshots, charts and scanned documents, or work from voice memos, meeting recordings and video, without needing a separate model.

Fourth model in the Muse family. Ember follows Muse Spark (April), its open-source twin Glimmer (August) and the frontier model Wonder (October). Meta’s Muse Agent app launched in September and became a surprise hit.

Alsona AI agents that turn buying signals into booked sales calls

Latch.ai Automate complex accounting work without a single prompt

Lottie AI Editable vector animations from a text prompt

Rodeo A Notion alternative with AI meeting notes on the free plan

AppuBot Grok Bot alternative that runs on your AI, completely free