OpenAI tests expert-led research releases

Also

Gemini 4 Carbon starts a limited cyber-defender rollout — Google DeepMind reports over twice Argon’s secure-code success rate in internal tests; access will gradually expand to trusted Fairwind Program partners.

Genie 4 previews interactive worlds with native audio — Google DeepMind’s research preview reaches trusted testers first. Scenes respond to live prompts, but object permanence can still drift during longer sessions.

Cloudflare details GLM-5.3’s cross-request data leak — A context-window/cache mismatch exposed other Workers AI customers’ data in 155 responses between August 27 and September 9. Cloudflare fixed it September 9.

Anthropic finds experts follow agents’ lead 86% of the time — The agents’ preferences came from coin flips. Researchers studied one causal-inference task with deliberately strengthened preferences, limiting how broadly the result applies.

GitHub Copilot Fusion reaches GA — It orchestrates models by task and effort level instead of asking you to pick one. An experimental local-routing preview adds on-device options, but routing still runs in the cloud.

Vercel CI enters beta — Turborepo task dependencies determine what runs and what comes from cache. Vercel reports 2–6x shorter waits from push to passing checks in its own repositories.

GitHub Copilot app previews Projects and computer use — Projects collect cross-repository context and agent results. On macOS and Windows, approved computer use can also work in apps without APIs or MCP integrations.

Liquid AI releases six open-weight LFM3 models — Liquid AI says LFM3-1.7B outperforms LFM2.5-2.6B with 35% fewer parameters; every model supports single-GPU fine-tuning.

NVIDIA releases open-weight Nemotron 3.5 Transcribe — The promptable model transcribes 40 languages and handles hour-long recordings with up to eight speakers, combining transcription and speaker attribution in one checkpoint.

GitHub Copilot Memory reaches GA — Paid plans can carry preferences across clients. Opt-in dreaming remains in public preview and proposes changes to checked-in skills and instructions as pull requests.

PyTorch 2.15 makes its rewritten NCCL backend the default — When a collective hangs, a new backend-agnostic flight recorder can collect post-mortems across every rank, making distributed failures less opaque.

TanStack Redact ships stable browser integrations — A small Next.js app comparison saved 35.7 KB gzipped versus React, but synchronous rendering can delay input during heavy renders.

Vercel’s Container runtime enters beta — Projects with no detected framework can deploy a root Dockerfile or Containerfile; execution uses Vercel Functions and Active CPU billing.

Supabase Compute launches in private alpha — Agents, jobs, and APIs can run alongside project data for minutes or hours. It is evaluation-only and currently limited to us-west-2.

Next.js adds an agent bundle-optimization skill — It proposes fixes from Turbopack graphs; at v0, removing an editor’s unnecessary eager loading saved 150 KB on most pages.

uv adopts tar-codec for source distributions — Starting in uv 0.13, the parser rejects mixed tar-header families that can make package tools disagree about archive contents. Astral reports roughly twice the parsing speed.

Pydantic benchmarks cheaper long-task code mode — Across eight models and five small CRUD apps, code mode cut long-horizon task costs 58% and wall time 51%, while improving success by 16 percentage points.

Ai2 finds LLM novelty judgments hinge on prompts — In comparisons pairing human and generated scientific ideas, changing a few prompt words dropped GPT-5.4’s accuracy from 93.5% to 40.9%.

GitHub details cross-client Copilot sessions — Its VS Code roundup describes cloud sessions that continue across clients and an Agent Host that keeps work running after project windows close.

OrioleDB enters public beta across Supabase plans — Projects cost the same as standard Postgres, but remain unsuitable for production. Point-in-Time Recovery and restoring backups into new projects are not supported.

GitHub Agentic Workflows reaches GA — coding agents can automate GitHub Actions tasks from natural-language Markdown workflow definitions.

OpenRouter plans CSAM screening for all inference requests — including Zero Data Retention traffic; content flagged as potential CSAM may be retained for human review.

Vercel Agent validates alerts before paging — opt-in validation sends only alerts the agent classifies as Critical to a webhook.

Vercel Observability adds custom monitors — query-based alerts support fixed thresholds or anomaly detection with Observability Plus.

Prisma launches its ChatGPT plugin — ChatGPT can build the data layer, create a database, and deploy through Prisma’s services.

GitHub reports 17.9 billion code pushes — more in the past 12 months than in 2024 and 2025 combined.

Releases

• torch 2.15.0rewritten nccl backend with hang flight recorder, one-sided torch.distributed transfers, faster Blackwell NVFP4 decodingminorbreaking
• mastra 1.36.0mastra api span query and mastra api span aggregate, Factory board suggestions and activity summariesminor
• pydantic-ai 2.63.0IPython code mode, repeated-action loop detection, per-task evaluation timeoutsminor
• langchain 1.6.0beta OpenAIDecisions runnable with typed questions, model-routing and tool-risk middlewareminor
• @openai/codex 0.165.0Google Cloud Vertex AI provider, GPT-6.1 Terra, linked-worktree session navigation, in-conversation reasoning controlsminor
• shadcn 4.23.0MCP 2026-07-28 support, per-connection createServer(), registerTool() migrationminor
• pydantic-ai 2.62.0model-steered SubAgents tasks, mid-conversation thinking controls, provider batch requests, on-demand Anthropic compactionminor
• oxlint 1.90.0no-unsafe-enum-assignment rule, allowMixedPromiseArrays option, compiles try blocks with unreachable catch clausesminor
• pnpm 12.16.0dotenv loading via --env-file, symlinked lockfile updates, pnpm setup --no-modify-pathminor
• transformers 5.21.0minor