<?xml version="1.0" encoding="UTF-8"?><rss version="2.0"><channel><title>devtake.dev — GPT-5.6 Sol</title><description>Articles on devtake.dev covering GPT-5.6 Sol.</description><link>https://devtake.dev/</link><language>en-us</language><item><title>Alibaba&apos;s Qwen3.8-Max beats Fable 5 on Terminal-Bench, and the weights go public next week</title><link>https://devtake.dev/article/qwen-3-8-max-open-weights/</link><guid isPermaLink="true">https://devtake.dev/article/qwen-3-8-max-open-weights/</guid><description>Qwen3.8-Max is a 2.4-trillion-parameter MoE that tops Claude Fable 5 on Terminal-Bench 2.1 and trails it badly on SWE-bench Pro. It&apos;s the first open Max-tier Qwen.</description><pubDate>Mon, 03 Aug 2026 09:20:00 GMT</pubDate><category>ai</category><category>qwen</category><category>llm</category><category>ai-models</category><category>open-weights</category><category>moe</category><category>benchmarks</category><category>alibaba</category><category>china</category><author>dieter-morelli</author></item><item><title>Ten decade-old math problems fell to an unreleased OpenAI model, for about $2,000 of tokens each</title><link>https://devtake.dev/article/openai-astra-math-lean-proofs/</link><guid isPermaLink="true">https://devtake.dev/article/openai-astra-math-lean-proofs/</guid><description>OpenAI published ten results in math and theoretical CS from an internal build of Astra, with Lean 4 certificates for every proof. What that verification does and doesn&apos;t settle.</description><pubDate>Mon, 03 Aug 2026 09:05:00 GMT</pubDate><category>ai</category><category>openai</category><category>llm</category><category>ai-models</category><category>mathematics</category><category>formal-verification</category><category>astra</category><category>benchmarks</category><author>dieter-morelli</author></item></channel></rss>