Their scores on SWE bench are very close because the benchmark is nearly saturated. Gemini 3 beats Sonnet 4.5 on TerminalBench 2.0 by a nice margin (54% vs. 43%), which is also agentic coding (CLI instead of python).