“Microsoft's Thinking 1 model scored significantly lower on agentic coding benchmarks like TerminalBench 2.0 and SuiBench Pro compared to previous generation models from Anthropic and OpenAI.”