Paul CharlesworthA place for Paul Charlesworth to think out loud and write explainers for people to learn from.https://paulcharlesworth.net/GPT 5.5 leads the DeepSWE benchmarkhttps://paulcharlesworth.net/blog/deepswe/https://paulcharlesworth.net/blog/deepswe/DeepSWE is a new software engineering benchmark built around fresh, diverse, real-world tasks with handwritten verification.Sun, 31 May 2026 00:00:00 GMTaisoftware-engineeringbenchmarksComposer 2.5 sits with the frontier coding agentshttps://paulcharlesworth.net/blog/composer-2-5-coding-agent-benchmarks/https://paulcharlesworth.net/blog/composer-2-5-coding-agent-benchmarks/Cursor Composer 2.5 ranks third on Artificial Analysis' coding agents benchmark—and costs a fraction of frontier models.Sat, 23 May 2026 00:00:00 GMTaicursorsoftware-engineeringThe Falling Cost of Inferencehttps://paulcharlesworth.net/blog/the-falling-cost-of-inference/https://paulcharlesworth.net/blog/the-falling-cost-of-inference/The latest models are getting more expensive, but the absolute cost of inference is falling rapidly.Thu, 14 May 2026 00:00:00 GMTaiAI Proves You Don't Know What You Wanthttps://paulcharlesworth.net/blog/ai-proves-you-dont-know-what-you-want/https://paulcharlesworth.net/blog/ai-proves-you-dont-know-what-you-want/Why weak AI prompts often reveal unclear requirements, not weak tools.Mon, 04 May 2026 00:00:00 GMTaisoftware-engineering