Google DeepMind released an updated version of Gemini 2.5 Flash this week that significantly improves performance on coding benchmarks, with the model now leading several widely-used evaluations including SWE-bench Verified, a test of the ability to resolve real GitHub issues without human assistance.
The updated Flash variant scores 68.4% on SWE-bench Verified, up from 54.2% in the April 2026 release, surpassing the previous leader Claude Sonnet 4 by 2.1 percentage points in controlled comparisons. On HumanEval+, which measures correctness of generated code including edge cases, Gemini 2.5 Flash now achieves 91.6% pass@1 accuracy.
The performance improvements are attributed to three changes: an extended code-specific pre-training corpus incorporating 18 months of additional open-source repository data, a reinforcement-learning fine-tuning stage using AI-generated unit test results as reward signals, and a new "diff mode" in the API that returns minimal patch diffs rather than full file rewrites, reducing output token count by 60% for editing tasks.
For developers, the practical impact is significant. The model's speed and cost profile — Flash is priced at 75 cents per million input tokens and $3 per million output tokens — makes it economically viable for use in continuous integration pipelines where every commit triggers AI-assisted code review. Several large development teams have reported integrating Gemini 2.5 Flash into GitHub Actions workflows to surface potential bugs before human reviewers see the pull request.
Google is also expanding Google AI Studio to include a "coding sandbox" environment where developers can test multi-file code generation with live execution, aiming to compete directly with Cursor's AI-native editor which has gained strong traction in 2026.
The announcement is part of Google's broader strategy to reassert itself in the AI developer tools market after losing significant mindshare to OpenAI and Anthropic over the past two years. The Gemini API now processes over 3 trillion tokens per month across all variants, a 400% increase from July 2025, suggesting the strategy is gaining traction even as the company faces intense competition.
DeepMind researchers noted that Flash was specifically optimised for latency, with median time-to-first-token under 300ms on standard API access, making it suitable for real-time coding assistants where perceived responsiveness matters as much as raw accuracy.