8/26/2026, 1:04:30 PM · foundation-models

Google Releases Gemini 3.7 Flash, Scoring 65.3% on DeepSWE v1.1 Coding Benchmark

Google DeepMind's Gemini 3.7 Flash, released August 13, 2026, posts a 16-point gain over its three-week-old predecessor on a contamination-resistant software engineering benchmark while cutting introductory token pricing in half.

Overview

Google DeepMind released Gemini 3.7 Flash on August 13, 2026, <cite index="19-3">positioning it as a fast, low-cost model for coding and agent workloads.</cite> <cite index="19-4">The model arrived just 23 days after Gemini 3.6 Flash, retains the same 1,048,576-token context window, and launches at an introductory price of $0.75 per million input tokens and $3.75 per million output tokens.</cite> <cite index="26-3,26-4">Those introductory rates expire on December 31, 2026, after which standard pricing of $1.50 per million input tokens and $7.50 per million output tokens applies.</cite>

Benchmark Performance

The headline figure is a sharp gain on DeepSWE, a benchmark created by startup Datacurve. <cite index="13-1,13-2,13-3">DeepSWE evaluates coding agents on 113 original, long-horizon software engineering tasks spanning 91 active open-source repositories in five languages; tasks are written from scratch and never upstreamed to limit training contamination, and each is graded by a hand-written program-based verifier that checks observable behavior rather than a specific implementation.</cite> <cite index="13-4">Agents run autonomously in sandboxed containers without internet access.</cite>

<cite index="25-1,25-2">According to Google's official announcement, 3.7 Flash shows strong gains over 3.6 Flash in coding tasks such as debugging and issue resolution, achieving improved performance in generating production-ready code as measured by FrontierCode 1.1 Main (43.6% vs. 34.4%) and DeepSWE v1.1 (65.3% vs. 49.0%).</cite> <cite index="25-4,25-5,25-6">In web development, the model generates more functional layouts and feature-complete apps in fewer prompts, showing high design adherence whether the reference input is a screenshot, image, or full design system, and outperforms 3.6 Flash on Arena.ai's WebDev Arena with an Elo score of 1,588 versus 1,538.</cite>

<cite index="19-10">Additional scores listed on the model card include 97.0% on GDM-MRCR v2 at 128k context and 14.9% on Terminal-bench 3.0</cite> — the latter illustrating how scores fluctuate significantly depending on which benchmark edition is cited.

Architecture and Release Rationale

<cite index="27-8">Google's release notes are explicit that the gains came from algorithmic innovations rather than a larger model or a longer context window.</cite> <cite index="26-15">The knowledge cutoff date for Gemini 3.7 Flash is March 2026.</cite> <cite index="26-5">Google identifies agentic workflows, coding tasks, and enterprise workflows as primary use cases.</cite>

Context and Caveats

The rapid release cadence — a meaningful benchmark jump in under a month — reflects an accelerating cycle in Large Language Model (LLM) development at the flash tier. <cite index="27-10">A flash-tier model, the cheap and fast tier Google positions below its Pro-tier models, closing that much distance on real software engineering tasks in three weeks points to a shift in where the actual bottleneck in frontier AI development lies.</cite>

On the broader leaderboard, context matters. <cite index="20-2">On the hardest agentic tests, competing models such as GPT-5.6 Terra still lead: DeepSWE v1.1 at 69.6% against Gemini 3.7 Flash's 65.3%.</cite> Benchmark scope is also a limiting factor: <cite index="14-7,14-8">DeepSWE draws exclusively from open-source repositories with 500-plus stars, results may not generalize to proprietary codebases, and widely used languages such as C++ and Java are absent entirely.</cite> <cite index="14-10,14-11">Datacurve is a startup with its own commercial interests, though the company's decision to publish the full dataset, all agent trajectories, and the evaluation harness on GitHub mitigates this concern considerably.</cite>

<cite index="21-5">Gemini 3.7 Flash is rolling out to Spark and is available across Google's developer and enterprise AI products.</cite>

Sources

  1. [1]
    Dissecting Embodied Abilities in Multimodal Language Models through Skill-level Evaluation and Diagnosis
  2. [2]
    Improved Gemini 2.5 Flash and Flash-Lite
  3. [3]
    AudioMarathon: A Comprehensive Benchmark for Long-Context Audio Understanding and Efficiency in Audio LLMs
  4. [4]
    Continuing to bring you our latest models, with an improved Gemini 2.5 Flash and Flash-Lite release - Google Developers Blog
  5. [5]
    Gemini 2.5 Flash - Intelligence, Performance & Price Analysis | Artificial Analysis
  6. [6]
    The ORCA Benchmark: Evaluating Real-World Calculation Accuracy in Large Language Models
  7. [7]
    Google I/O 2025: Updates to Gemini 2.5 from Google DeepMind
  8. [8]
    Gemini 2.5 Flash | Gemini Enterprise Agent Platform | Google Cloud Documentation
  9. [9]
    Gemini 3.5 Flash - Model Card — Google DeepMind
  10. [10]
    Gemini 3.7 Flash — Google DeepMind
  11. [11]
    DeepSWE: Measuring Frontier Coding Agents on Original, Long-Horizon Engineering Tasks
  12. [12]
    DeepSWE: Measuring Frontier Coding Agents on Original, ...
  13. [13]
    DeepSWE | Epoch AI
  14. [14]
    DeepSWE blows up the AI coding leaderboard, crowns GPT-5.5, and finds Claude Opus exploiting a benchmark loophole | VentureBeat
  15. [15]
    DeepSWE Leaderboard
  16. [16]
    SBSC: Step-By-Step Coding for Improving Mathematical Olympiad Performance
  17. [17]
    Frontier Coding Agents Use Metaprogramming to Adapt to Unfamiliar Programming Languages
  18. [18]
    GitHub - datacurve-ai/deep-swe: Measuring frontier coding agents on original, long-horizon engineering tasks · GitHub
  19. [19]
    Google releases Gemini 3.7 Flash for coding and agents
  20. [20]
    Gemini 3.7 Flash: A Big Coding Jump in a Three-Week Point Release
  21. [21]
    Gemini 3.7 Flash Cuts Launch Price 50% And Hits 65.3% DeepSWE
  22. [22]
    Gemini (language model)
  23. [23]
    Google Gemini
  24. [24]
    Gemini 3.7 Flash: our most intelligent workhorse model
  25. [25]
    Gemini 3.7 Flash - Model Card — Google DeepMind
  26. [26]
    Gemini 3.7 Flash Explained: How Google's New Coding Workhorse Jumped 16 Points on Real Software Engineering Benchmarks