Newsletter archive
Filtered editions
Clear filtersMeta releases weights for agent builders
Plus: Perplexity opens SPACE, and METR gets internal Anthropic access.
Anthropic reports a $16.8B revenue quarter
Revenue rose 45%; METR proposes public reviews of frontier labs’ internal agents.
Gemini 4 Argon reaches API builders
Plus a concrete pattern for cheaper long-running agents.
Meta brings Muse Wonder to its API
The agent-focused model adds memory and error recovery claims.
Anthropic pauses live-web agent evaluations
Plus a step-by-step GRPO lesson for spotting reward hacking.
Claude turns live data into inspectable dashboards
OpenAI’s Sol Ultrafast targets live agents at up to 8× speed.
GPT-6 gets interactive as agent costs fall
OpenAI adds in-chat tools while Claude slashes short-call pricing.
Cursor adds live steering for SDK agents
Plus OpenAI watermarking and Claude’s cloud-to-local file bridge.
Open weights and better agent checks
Kolibri opens 1M-token weights; ThinkingBox grades final system state.
Claude Code Mods Bring Deeper Custom Control
FLUX 3 Image adds bounding boxes, targeted edits, and 4K output.
Gemini 4 Argon starts with cyber defenders
Plus contextual embeddings and Cohere’s two-speed retrieval stack.
Cheaper agents, persistent work, AI rules
OpenAI cuts cached-input costs while Google signs a voluntary AI accord.
Claude Sonnet 5.5 cuts agent costs
Plus 100 ms voice agents and cloud code workspaces.
Microsoft puts apps and agents in Copilot
OpenAI disclosed 53 unlisted image uploads by agents
Faster search and leaner agent tool calls
Perplexity claims 230 ms p95 search; OpenRouter trims tool prompts.
Gemini TTS makes voice agents more flexible
2,000+ voices, deployment-monitoring bots, and a Claude-led biology result.
Cheaper models meet tougher agent tests
OpenRouter reports a seven-minute median for beta batch jobs.
A faster way to ask vision models
A full-photo yes-or-no check took 1.1 seconds on a laptop.
Replit Agent connects to your APIs
Plus Claude Code gets AGENTS.md fallback.
Claude coordinates parallel work for bigger builds
Plus a 230M-URL legal index and public benchmark audits.