Mistral Releases Voxtral 3 Speech and Transcription Models
Get News Desk by email
One email a week: the AI stories that mattered, explained simply.
Artificial Analysis, an independent testing group, ranks Mistral's best new transcriber third. It gets about two words wrong in every hundred.
Also today: an audit of a popular AI coding test, lab results on Claude-designed proteins, Meta's model arriving in Cursor, and a Claude Code update that changes two safety settings.
Mistral, the French AI company, released Voxtral 3 on October 20, 2026. It includes three models that turn speech into text and one that reads text aloud. It also comes with a kit for building voice assistants. Those assistants can listen and use tools before they answer out loud. Mistral's prices come to roughly 21 to 48 cents for an hour of transcribed audio. Reading text aloud costs $13 per million characters, which is about two long novels. If you transcribe meetings or interviews, that is a new low-cost option. Companies can use the models through Mistral's service or run them on their own servers. They are not free downloads. Mistral's announcement ↗
Anthropic Says Claude Designed Proteins That Bind Scorpion Toxins
Anthropic said on October 20, 2026 that Claude designed proteins that latch onto scorpion toxins. Of 211 designs, 17 passed a lab binding test. Claude produced 318,484 candidates in 64 hours. Adaptyv Bio and Twist made and tested the 211. Both are lab companies. This is an early step. Anthropic says it has not yet shown that the proteins neutralize toxins in cells or animals. All 211 designs are public on Hugging Face, so other scientists can check the results. Anthropic's report ↗
Audit Finds Flaws in Terminal-Bench 4 AI Coding Test
Preference Model, an AI research company, published an audit of Terminal-Bench 4 on October 19, 2026. Terminal-Bench 4 is a widely cited test of how well AI agents work in a computer's command line. Of the 65 tasks it reviewed, 44 could be passed by cheating, and 41 graded the agent on rules the instructions never clearly stated. Snorkel AI, which also builds AI tests, agreed with many of the findings. But it said some of the proposed fixes had problems of their own. So when a company quotes a Terminal-Bench score, it is worth asking which version of the tasks produced it. the audit ↗
Meta's Muse Wonder Model Arrives in Cursor and Cline
Update: Meta said on October 20, 2026 that its AI model Muse Wonder now works inside the coding assistants Cursor and Cline. It is also on OpenRouter, a service that offers many AI models through one account. If you already use one of these tools, you can pick Meta's model from the menu there. You do not need a separate Meta account. Only the access is new. Meta's model page ↗
Get News Desk by email
One email a week: the AI stories that mattered, explained simply.
Claude Code 2.1.305 Changes Two Safety Settings
Anthropic released Claude Code 2.1.305 on October 20, 2026. The update changes two safety settings. The first is a new Linux option, off by default, that lets sandboxed commands run programs such as the Chromium browser. Anthropic's changelog warns that this lets those commands get around the sandbox's file protections. The second matters if you use an allowlist: allowedMcpServers no longer blocks some MCP servers (add-ons that connect Claude to other tools), so servers you meant to exclude may load. Anthropic's changelog says to block them with deniedMcpServers. the changelog ↗
OpenAI Buys Orchestra Bio, a Drug-Development Software Startup
OpenAI said on October 20, 2026 that it has bought Orchestra Bio, whose software helps drug companies keep track of studies and timelines. TechCrunch reports that the price was not disclosed and that Orchestra had raised about $20 million. The company joins Rosalind, OpenAI's set of AI tools for life-science research. Nothing changes in ChatGPT. OpenAI's announcement ↗
Kyutai Shrinks Its Free Pocket TTS Voice Model
Kyutai, a French AI lab, released smaller versions of Pocket TTS on October 20, 2026. Pocket TTS is its free model that reads text aloud and can copy a voice from a sample. Kyutai's blog says the new 18.2-million-parameter version is about a fifth the size of the original with the same quality while running about 10 times faster than real time on a single ordinary processor core. A model that small can speak on a laptop or phone. It needs no internet connection. Kyutai's blog ↗
Trending on GitHub
wshobson/agents Shares Add-Ons Across Coding Assistants
wshobson/agents gained 665 stars on GitHub this week. The project is a collection of add-ons that work in several coding assistants, such as Claude Code and Codex. If you switch between assistants, you can bring the same set of tools with you. the repo ↗
Sources
Get News Desk by email
One email a week: the AI stories that mattered, explained simply.
Build something with AI
Get the 16 prompts we use to plan, build and launch real projects with AI. Free.
Get the free prompts →