engblogs

How to run 100k battles of untrusted agent code at scale by dynamically adding per-GPU vLLM replicas behind a single Caddy proxy, with a runtime-concurrency limiter and a dedicated rate-limiter to keep rounds fair, contained, and on schedule.

A concise technical guide to chess insults, outlining the meanings, origins, and usage of terms like sandbagger, patzer, fish, woodpusher, and flagger with practical examples.

A concise taxonomy of chess-insult terminology with etymologies, practical examples, and implications for rating integrity and online etiquette.

Gemini Robotics ER 2 acts as a high-level embodied reasoning brain that fuses real-time video understanding, task orchestration, and multi-robot collaboration with tool use and safety to enable proactive, multi-step robotic workflows.

How GPT-5.6 unites frontier intelligence with frontier efficiency through end-to-end stack optimizations—advances in inference, load balancing, speculative decoding, kernel and cache optimizations, and an agentic harness to deliver more work per token at lower cost.

Free access to frontier AI tools for 100,000 academic researchers, featuring GPT-5.6 and Codex, to accelerate scientific discovery across sciences, math, and engineering through enhanced collaboration, training, and research workflows while prioritizing privacy and governance.

Two harness settings—retained reasoning and compaction—boost ARC-AGI-3 scores for GPT-5.6 Sol by roughly 3x and cut output tokens by about 6x by preserving past thinking and streamlining context across turns.

GPT-5.6 dramatically improves intelligence-per-token and lowers operating costs through end-to-end optimizations of inference, kernels, and its agentic harness, including load balancing, speculative decoding, caching, and tooling integration.

Free access to frontier AI tools for 100,000 academic researchers to accelerate discovery and streamline workflows with ChatGPT, Codex, and GPT-5.6, enabling collaborative, privacy-conscious research across disciplines.

GPT-5.6 combines frontier intelligence with frontier efficiency by optimizing inference, load balancing, speculative decoding, caching, and its agentic harness to deliver more work per token at lower cost across diverse workloads.

OpenAI accelerates scientific discovery by giving 100,000 academic researchers free access to frontier AI tools, including GPT-5.6 Sol Pro, to boost research workflows, enable cross-institution collaboration, and safeguard privacy with data not used to train models by default.

Enabling retained reasoning and compaction in the ARC-AGI-3 harness dramatically boosted GPT-5.6 Sol's scores, highlighting the impact of API settings and harness design on 2D puzzle benchmarks.