Gemini 3.7 Flash: Coding and Agent Model at $0.75/1M Input Tokens
Google shipped Gemini 3.7 Flash on August 13, 2026 — three weeks after Gemini 3.6 Flash — positioning it as a drop-in upgrade for coding pipelines and long-running agents. The model card describes it as algorithmic improvements to the reasoning core rather than a new pretraining run, which means the March 2026 knowledge cutoff carries over unchanged. What changed is performance on software engineering and document-heavy workflows, and the price: at $0.75 per 1M input tokens and $3.75 per 1M output tokens, it arrives at roughly a third the blended cost of its nearest frontier rivals.
For developers already thinking about what the agentic era means for data science, 3.7 Flash is a clear pricing signal that always-on agent workloads are becoming economically viable without a Pro-tier budget.
Benchmark Profile
The gains concentrate in three areas: software engineering, document comprehension, and web development. On FrontierCode 1.1 Main, 3.7 Flash scores 43.6% against 34.4% for 3.6 Flash — the highest score in Google's comparison table. On DeepSWE v1.1, it reaches 65.3% versus 48.6% for the prior version. WebDev Arena Elo moves from 1538 to 1588, again the top result in the table.
Document and automation results are the larger surprise. GDP.pdf rises from 22.0% to 34.0%. AutomationBench goes from 17.0% to 30.4% — ahead of Claude Sonnet 5 at 10.7% and GPT-5.6 Terra at 23.6%. Long-context retrieval on GDM-MRCR v2 at 128k reaches 97.0%. Harvey LAB-AA, a complex legal workflow eval, reaches 90.7%.
The gaps in the other direction are equally clear. GPT-5.6 Terra leads on DeepSWE (69.6%), Terminal-bench 2.1 (87.4% vs 85.8%), Terminal-bench 3.0 (20.8% vs 14.9%), and OSWorld-2.0 (50.2% vs 47.9%). CharXiv Reasoning without tools is a minor regression: 84.5% for 3.7 Flash against 85.2% for 3.6 Flash. On the Artificial Analysis Intelligence Index, 3.7 Flash scores 56 — one point behind both GPT-5.6 Terra and Muse Spark 1.2 at 57. Teams evaluating AI agents for direct corpus interaction will notice that terminal and computer-use benchmarks still favour GPT-5.6 Terra.
Pricing
| Model | Input ($/1M tokens) | Output ($/1M tokens) | Blended $/1M (80/20 mix) | Artificial Analysis Intelligence Index |
|---|---|---|---|---|
| Gemini 3.7 Flash (introductory, until Dec 31, 2026) | $0.75 | $3.75 | $1.35 | 56 |
| Gemini 3.7 Flash (from Jan 1, 2027) | $1.50 | $7.50 | $2.70 | 56 |
| Muse Spark 1.2 | $1.25 | $4.25 | $1.85 | 57 |
| Claude Sonnet 5 | $2.00 | $10.00 | $3.60 | 55 |
| GPT-5.6 Terra | $2.00 | $12.00 | $4.00 | 57 |
The introductory rate expires December 31, 2026, after which input doubles to $1.50 and output doubles to $7.50. At today's rate, the blended cost at an 80/20 input-output mix is $1.35 per 1M tokens — $2.25 cheaper than Sonnet 5 and $2.65 cheaper than GPT-5.6 Terra at the same mix. Even at the post-January list price, 3.7 Flash stays cheaper than both frontier alternatives. For teams considering how to automate LLM prompt optimization in production, the pricing structure makes repeated inference loops substantially less punishing than on rival models.
Context, Output Cap, and Modality
The 1M-token context window and 64K-token output limit are unchanged from 3.6 Flash. Input modalities cover text, images, audio, and video. The model supports customizable thinking configurations that let developers trade response quality against latency and cost — relevant for agentic coding workflows where some steps require deep reasoning and others do not.
Google's own benchmark set targets legal, financial services, biosciences, and enterprise operations as primary verticals, which the Harvey LAB-AA (90.7%), GDP.pdf (34.0%), and AutomationBench (30.4%) results are designed to support. Long-context retrieval at 97.0% on GDM-MRCR v2 at 128k makes the 1M-token window practically usable for PDF-to-structured-data pipelines and multi-document legal analysis.
Availability
Access is API and enterprise only, with no open weights released. Hosted surfaces include the Gemini API, Google AI Studio, Google Antigravity, Android Studio, the Gemini Enterprise Agent Platform, and the Gemini Enterprise app. Consumer access runs through Gemini Spark on Google AI Pro and Ultra plans. There is no self-hosted or air-gapped deployment path, which excludes teams with strict data-residency requirements regardless of the pricing advantage.
The release cadence itself is notable: 3.7 Flash landed three weeks after 3.6 Flash, continuing a compression of the Flash release cycle that has seen three iterations ship since Google I/O in May 2026. At $0.75 per 1M input tokens through the end of the year, Google is using price to drive adoption before the list rate doubles — developers evaluating the model for production agent workloads should factor that January 2027 reset into any cost modelling they do today.