Anthropic's Message Threads beta cuts agent payloads 800KB to 1.8KB
Send only new turns and a thread ID; first-token latency falls ~35% on huge contexts. Plus GitHub's one-click MCP connectors and DeepSeek's new open weights.
Copy markdown
Send the delta, not the transcript
Message Threads, now in public beta on the Messages API, caches your conversation server-side: the first call opens a thread, your client keeps the response ID, and every later turn sends only the new messages plus that ID. In long Claude Code sessions the median request dropped from ~800KB to 1.8KB and time-to-first-token fell ~35% on 500K-1M-token prompts. It's live on the native API, AWS, Microsoft Foundry, and Google Cloud.
The 5-minute catch - keep your own copy
A thread expires five minutes after its last request, and its state is encrypted in memory and never written to disk, so if one lapses your client must resend the full conversation. Billing is unchanged too: Claude still processes the entire context each turn, so you're buying lower latency and bandwidth, not cheaper tokens.
GitHub wires 26 tools into Copilot with one sign-in
GitHub-verified MCP connectors are in public preview: connect Slack, Figma, Notion, Sentry, Linear, Cloudflare, Vercel or Atlassian Rovo once and they work across the Copilot app, CLI, VS Code, cloud sessions, and GitHub Mobile - no per-tool MCP wiring. Open to every plan, though Business and Enterprise need an admin to flip on the Editor preview policy.
DeepSeek V4.1 Flash ships open, under MIT
DeepSeek's newest Flash flagship landed with downloadable open weights under an MIT license and is already hosted on third-party providers with public per-token pricing - so you can self-host it or call it through a gateway today without waiting on DeepSeek's own endpoint.
Hot-swap LoRAs on a local Ryzen AI NPU
AMD's developer guide shows how to switch LoRA adapters at runtime on a single INT4 base model kept resident on the Ryzen AI NPU: adapters are quantized against the frozen base with AMD Quark, then flipped with one set_active_adapter call. It's a cheap way to run several fine-tuned personas (Qwen2.5-3B, Qwen3-4B) on one laptop without reloading the base model.