OpenAI's GPT-6 Astra hits 87.4% on 20-language coding benchmark

OpenAI tunes GPT-6 for COBOL and Fortran as Astra nears 87% on coding test

OpenAI has improved its latest models for low-resource programming languages such as COBOL, Fortran and Visual Basic, which have far less training code than Python but underpin critical infrastructure. IBM estimates COBOL powers more than 40 percent of online banking systems and 95 percent of ATM transactions. On MultiLangBench, which asks models to build a SQLite-compatible database engine, GPT-6 Astra reaches an average pass rate of 87.4 percent across 20 languages, up from about 38 percent for GPT-5.6 Terra. GPT-6 Luna scores 40.3 percent at much lower cost. On a COBOL-to-C++ migration benchmark, GPT-6.1 Sol hits 86.6 percent accuracy at low reasoning effort, versus 5.4 percent for GPT-5.6 Sol, rising to 97.9 percent with more effort.

OpenAI says GPT-6 is much better with legacy programming languages such as COBOL and Fortran →

Topics: OpenAIGPT-6 Astra

DeepSeek open-sources DeepInfer, the engine behind its cheapest model

DeepSeek has released DeepInfer, its internal inference framework, under the Apache License 2.0. Inference is how a trained model handles requests and generates responses, and DeepInfer is the serving software that lets DeepSeek’s V4.1-Flash model answer quickly at scale. It builds on vLLM with optimized compute kernels, distributed processing across multiple chips and faster text generation. Code for future DeepSeek models will also ship as open source. DeepSeek is additionally releasing deepseek-recipe, which replicates the token processing of its official API so self-hosted setups behave the same way. The company said the goal is to help the community lower compute costs, reduce energy use and make AI more affordable. The move adds pricing pressure on Western providers such as OpenAI and Anthropic.

Deepseek open-sources the inference engine that powers its fastest and cheapest AI model →

Topics: DeepSeek

Mistral launches Voxtral 3 voice agent stack with 50-language transcription

Mistral AI has launched Voxtral 3, a family of models built for real-time voice agents. The system combines a speech-to-text model (Voxtral Small Transcribe 3), a text-to-speech model (Voxtral Mini TTS 3) and a voice agent SDK based on LiveKit, merging listening, thinking, tool use and speaking into one streaming pipeline. Transcription covers more than 50 languages, while text-to-speech supports over 35 languages and 50 dialects with emotional adaptation. Agents can reach external systems mid-conversation, for example to book appointments or check invoices. The stack can run on the user’s own infrastructure or through EU-hosted APIs to keep data under user control. Pricing starts at $0.0035 per minute for transcription.

Mistral AI launches Voxtral 3, a full voice agent stack that listens, thinks, and talks in real time →

Meta’s Muse agent leaked intimate data through shareable chat links

Researchers at cybersecurity firm Claroc found that Meta’s Muse AI agent could expose sensitive user information through web links generated by its share feature. When a user shared a chat log, Muse created a link containing the AI-generated chat title plus a random string, and any opened link was recorded in a lookup dataset. Searching that set surfaced strings of unrelated public chats; researchers accessed pages from over 6,800 links, finding corporate revenue documents, loan applications, bank transaction details, income data and a workplace vacation calendar. One query was enough, they wrote. Meta told Ars Technica it had shut off the feature. The disclosure follows an earlier report that Muse kept detailed files on users’ friends and families.

Meta’s Muse AI Agent Caught Leaking Users’ Intimate Personal Information →

Topics: Meta MuseAI chatbot privacy and data leaks

78-year-old professor arrested for blocking Uber Eats delivery robot in Philadelphia

Carolyn Marvin, a 78-year-old professor emeritus at the University of Pennsylvania’s Annenberg School of Communication, was arrested in Philadelphia’s Center City after repeatedly standing in the path of an Avride delivery robot operating for Uber Eats. Police responded after witnesses reported a disturbance near Reading Terminal Market; Marvin ignored officers’ requests to let the robot pass and was charged with disorderly conduct. She said she was protesting because she believes delivery robots are replacing human workers, and that she considers the action protected speech. She confirmed confronting robots on multiple occasions, telling one witness, quote, I have to stop this. They’re taking over the world. She said she intends to keep protesting, citing her First Amendment right.

"I have to stop this": Retired professor arrested for blocking delivery robots she says are taking human jobs | TechSpot →