An Nvidia research team argues that ever-larger context windows are often overkill and expensive. In many tests, periodic summarization combined with targeted lookups performed about as well as stuffing entire documents into the prompt, suggesting the industry should rethink the 'longer is better' assumption.
Context Window
上下文窗口
Maximum tokens a model can process in a single inference, determining its capacity for large codebases.
Curated Intelligence
Microsoft released MAI Cyber 2 Flash, a dedicated cybersecurity model that edges out GPT-6 Astra on the CyberGym-E2E benchmark with a 92.3% task success rate versus 91.5%, at under 2% of the per-task cost. It features a 1M-token context window, four times its predecessor, and will be exclusive to Microsoft security products, auto-replacing existing MAI configurations from mid-December.
Amazon published three composable optimization strategies for Amazon Bedrock AgentCore handling large JSON payloads, cutting token consumption by up to 71% and average latency by up to 80% on a representative 23 MB benchmark payload. The techniques target context bloat caused by repeated full-payload injection on every tool call in multi-turn agent sessions while preserving full reasoning capability.
Tokendoctor is a locally run cost-diagnostics and optimization tool for Claude Code that replays a user's real sessions to surface overspend and apply reversible fixes. On the author's six weeks of actual use it cut main-thread cost per request by 38% and context per request from 412K to 219K, though gains depend heavily on session length and model configuration.
Cloudflare has launched a beta of Auto Router inside AI Gateway, using an edge-side multi-head classifier to score each request on complexity, ambiguity, risk and context dependence, then selecting a model via an expected-quality-minus-adaptive-cost-penalty formula. Its own benchmark shows roughly 10 points higher reliability at about one-third the cost, but practitioners question whether risk scoring based on only the last few turns and 8KiB of context can hold up in long agent sessions.
Silicon Valley startup AFK AI says its product Comma already implements a single-router, multi-worker multi-agent architecture, using continuous context and long-term memory to keep agents running 24/7 across many parallel tasks, with code open-sourced. The piece frames this as converging with DeepSeek Harness's Super-agent vision around human attention as the bottleneck, but offers mostly conceptual narrative without verifiable architecture details or benchmark results.
AFK AI says its product Comma already implements the single-Router, multi-Worker super-agent architecture later outlined by DeepSeek's Harness team, adding persistent memory, continual learning and 24/7 uptime to address the single-session context bottleneck. The system is publicly available with open-sourced code, and the piece frames human attention—capped at roughly 3-5 active sessions—as the real constraint on agent concurrency.
Nanbeige, the foundation model team at Chinese recruitment platform BOSS Zhipin, has released Nanbeige4.3-33B, a dense 33B reasoning model under Apache 2.0 that scores 20 on the Artificial Analysis Intelligence Index v5, matching Qwen3.8 27B (xhigh). It offers a 524K-token context window and runs locally on a laptop, with terminal use and physics reasoning as relative strengths but weaker knowledge benchmarks than Qwen3.8 27B.
OpenRouter fixed a bug in its stealth reasoning model Crouton Alpha that caused agents to stop early, typically right before a tool call; agents should now run tasks to completion. The model was previously announced as free to use with a 1M-token context, text/image/audio/video input, and no training on user data.
Baidu released ERNIE-5.5-Flash-Base on Hugging Face, a mixture-of-experts foundation model with 280B total parameters and 14B active per token, pre-trained on 33.5 trillion tokens with a 128K training context that can extend to one million tokens during fine-tuning and inference. It ships deliberately without fine-tuning or post-training, positioning the raw weights for developer customization, with a post-trained version to follow.