Ai2 released AstaBrief 8B, an open-weights model designed to generate cited scientific reports faster and run locally. The team reports a 3.5× reduction in end-to-end generation time (51.1s vs 178.5s) and describes training choices and data filters intended to improve citation grounding.
Chatham Financial says it uses Codex and GPT‑5.6 to build internal tools, power the Chatham Onyx platform, and redesign trade validation workflows. Early results show trade review times fell from about 30 minutes to under 4 minutes.
Stardrift for iOS is an AI travel assistant that creates day-by-day trip plans and an organized wishlist. Leila Clark says the team has helped 10,000+ users; the app pulls booking details, recommends nearby places and highlights travel logistics. Launched in 2026.
The team behind Olmo-core 3 released an open training framework on October 1, 2026 intended to scale mixture-of-experts (MoE) training into the trillion-parameter range while keeping efficiency. Benchmarks report expert pools up to 128, MXFP8 gains ~21%, and tests at 1.2T parameters.
ProvenanceGuard is a post-generation verifier that keeps tool and source IDs through claim decomposition, support checking and attribution. In tests on medical agent traces, it flagged 138 of 139 unsupported claims and identified the correct source about 86% of the time.
OpenAI introduced GPT-6.1 Sol as an upgrade to GPT-6 Sol that narrows the gap with GPT-6 Astra on coding, professional workflows and computer-use tasks while reducing several token prices. Cached input is priced at $0.10 per million tokens.
The Open TTS Leaderboard (published Sep 30, 2026) introduces an objective, scalable framework to benchmark open-source multilingual TTS and voice cloning. Metrics include WER/CER via Qwen3 ASR, inference speed (RTFx, TTFA) and speaker SIM with WavLM embeddings.
At Dev Day, Sam Altman introduced the Decisions API, which he said narrows a model to a fixed set of options to make it faster while preserving image understanding and safety. Observers compare it to TypeSafe AI’s Jev; developers are testing similar decision-focused models.
Holo4 ships in 27B and 35B-A3B sizes and Holotron4 Nano updates a Nano agentic stack. The authors say the models operate across GUIs, code, MCP and APIs and report competitive scores on OSWorld 2.0 and AutomationBench at lower cost per task.
The launch team says RemoteConsole is a mobile SSH terminal designed for developers on the go. It lets users reattach to persistent tmux sessions running Claude Code, Codex, Aider or Gemini CLI and includes an SFTP browser and NAT traversal without a VPN.