Google shipped Gemini 3.7 Flash, its new mid-tier model for coding and agentic workflows, featuring a 50% introductory price cut and significant benchmark gains over version 3.6. The release positions the model as a workhorse for development, with immediate integration across the Gemini API, Android Studio, and external tooling like Cline.
- Google reports DeepSWE scores of 65.3% (vs 49.0%) and FrontierCode 43.6% (vs 34.4%), with an introductory price of $0.75 / $3.75 per 1M tokens.
- Artificial Analysis places Gemini 3.7 Flash at #56 on its Intelligence Index, noting ~340 output tok/s and improved discipline in tool loops.
- DeepSeek open-sourced DeepSeek Harness under the MIT license as a developer preview, emphasizing composable plugins and KV-cache-aware history semantics.
- Arcee released NAC under Apache 2.0 for long-running asynchronous work, while OpenAI previewed Ultrafast mode for GPT-5.6 Sol on Cerebras hardware at 750 tokens/sec.
The industry is shifting focus toward agent runtimes and custom evaluation platforms as model latency decreases and tool interaction complexity increases.