Daily cross-vendor scan — APIs, models, SDKs, pricing, deprecations. Source-linked.
Cursor 2.2 adds Debug Mode (instruments your app with runtime logs to find root causes), a browser sidebar visual editor, Plan Mode with inline Mermaid diagrams, automatic judging across parallel agent runs, and pinned chats in the agent sidebar.
Meta's Superintelligence Labs released Muse Spark, a multimodal reasoning model whose Contemplating mode runs parallel reasoning agents — 58% on Humanity's Last Exam, 38% on FrontierScience Research, live at meta.ai with a private API preview opening.
OpenAI set Aug 1 2026 as the EOL for the legacy Chat Completions streaming format; migrate to the Responses API streaming events (event-typed SSE).
Google shipped multi-token prediction drafters for Gemma 4 — speculative decoding with shared KV cache; Google reports roughly 2x tokens/sec for the 26B model on an RTX PRO 6000 and ~2.2x on Apple Silicon edge models.
OpenAI launched three voice API models: GPT-Realtime-2 for voice reasoning, GPT-Realtime-Translate (70+ input languages, 13 output), and GPT-Realtime-Whisper for live speech-to-text.
Google added Gemini 3.1 Pro and Gemini 3 Flash to the Gemini API as preview models, alongside Gemini 3.1 Flash-Lite, Veo 3.1, and Gemini Robotics — no GA dates, no published pricing.
xAI shipped Connectors on Grok Web — read/write for SharePoint, Outlook, OneDrive, Google Workspace, Notion, GitHub, and Linear, plus a Bring-Your-Own MCP slot; mobile coming soon.
Anthropic doubled Claude Code rate limits on Pro, Max, Team, and Enterprise plans, removed peak-hour reductions, and raised Opus API limits — backed by a new 300MW / 220,000-GPU Colossus 1 deal with SpaceXAI.
Google added event-driven webhook support to the Gemini API for long-running and batch jobs, replacing the older poll-for-status flow.
Anthropic shipped ten Claude finance agent templates as Cowork and Managed Agent plugins, plus Claude add-ins for Excel, PowerPoint, and Word; Outlook is coming soon.
Mistral released Medium 3.5, an open-weight 128B dense model with a 256k context — Mistral reports 77.6% on SWE-Bench Verified — and Vibe coding agents can now run remotely from the CLI or Le Chat.
DeepSeek released V4, an open-weight model with a million-token context, hybrid CSA+HCA attention, interleaved tool-call thinking, and a new RL sandbox called DSec — live on web, app, and API.