AI Briefing#context-window
The latest AI and developer news about #context-window, with the original source and a short summary.
The latest AI and developer news about #context-window, with the original source and a short summary.
The Claude Code system prompt revised its file-reading and REPL rules.
Context Mode, built for AI coding agents, has been introduced as reducing context usage by 98%.
Anthropic has ended the 1M context beta for Sonnet 4·4.5 and now blocks requests exceeding 200K.
GitHub Copilot is switching to token-based billing and has sharply raised the pricing multipliers for top-tier models.
On a single RTX 3090, Qwen3.6-27B achieved a 218K context and stable tool calling.
On an RTX 5080 16GB, Qwen3.6-35B-A3B hit 30 t/s at 128K and 89 t/s for new tokens.
In an 18GB M3 Pro bench, Gemma4:e4b ranked first with 72%, while Qwen3.5:9b scored 68%.
Mistral has released Medium 3.5, a 128B dense model, along with the Vibe remote agent.
In an RTX 5090 Llama.cpp experiment, MiniMax-M2.7 processed at 22 tps even at 100k context but only generated at 3~4 tps.
MIT researchers have proposed Recursive Language Models, which split up context to explore it.
Released an AI memory that achieved Recall@5 59% on the LoCoMo-10 benchmark.
DeepSeek-V4 unveiled a 1M-token context and agent optimizations.
A bug that caused Opus 4.7's 1M context to be truncated to 200K has been fixed.
In a 108K-token JS file memory test, Qwen 3.6 significantly outperformed Gemma 4.
compressmcp compresses JSON keys in MCP responses to cut Claude context usage by 40%.
Cloudflare has unveiled the private beta of **Agent Memory**, which preserves memory across sessions.
Because of the new tokenizer in Opus 4.7, the same input burns through tokens up to 35% faster than 4.6.
Using turbo3, achieved about 40 tok/s running Qwen3.6 35B A3B at 260k ctx on a 12GB VRAM 3080.
Assume LLMs hallucinate, loop, and ignore tools — wrap them with subagents and scripts.
The solution to agent problems is not the prompt but the operating structure.
Canceled Copilot Pro+ and switched to Claude because of rate limiting.
Agents, skills, and MCP piled up, consuming 30,000 tokens before the session even started.
Anthropic's Claude Code prompt cache TTL change and large-context costs have fueled complaints about faster quota depletion.
It proposes a recursive development workflow that uses repo documents to prevent context rot in agent work.