AI developer tools · What shipped

OpenAI released GPT-6.1 Sol on September 29, 2026, a mid-tier model achieving near-flagship performance at significantly reduced costs. On DeepSWE v1.1, GPT-6.1 Sol matches GPT-6 Astra's 74.8% accuracy while reducing cost per task by 80%. On OSWorld 2.0, it scores 71.4% compared to Astra's 73.5%, costing one-seventh as much per task. The model also outperforms Claude Opus 5.5 on professional document analysis (GDP.pdf: 32.0% vs 28.8%) and multi-step workflows (AutomationBench: 36.0% vs Claude Opus 5.5 at 35.4% with one-third the cost). Cached input pricing dropped 50% to $0.10 per million tokens. GPT-6.1 Sol is available immediately via the OpenAI API (gpt-6.1-sol identifier) and ChatGPT Work/Codex, with an ultrafast variant delivering 8x faster token generation rolling out to Codex.

Jevstiller is an open source project that distills Jev's outputs into a local student model, enabling faster and cheaper inference by handling familiar requests locally while routing uncertain queries upstream. The system achieves 98% agreement with Jev by learning from user requests on local hardware, reducing latency to as little as 15ms from device CPU instead of waiting 300ms for Jev's cloud responses. Jevstiller maintains agreement through continuous auditing that sends 2% of all requests to Jev regardless, allowing it to detect and quickly recover from model drift—demonstrated in testing when it recovered from a 90% to 9% local share drop within 49 minutes after detecting a simulated backend change.

OpenAI introduced GPT-6 Sol and GPT-6 Luna, bringing improvements to coding and agentic workflows while reducing API costs compared to GPT-5.6 predecessors, now available in GitHub Copilot. Anthropic launched Claude Opus 5.5 focused on long-running agentic coding tasks, 30% faster than Opus 5 with 40% lower costs and improved tool use and prompt injection resistance, also available through GitHub Copilot. GitHub added local sandboxing to the Copilot app, allowing developers to restrict agent access to files, network resources, and credentials on a per-project basis for safer local codebase operations.