Introduction
Gemini 3.7 Flash was officially launched on August 13, 2026. It is the first model released after Koray Kavukcuoglu took office as the new head of Google DeepMind. The Flash series has achieved a rapid iteration cycle of nearly monthly updates; Gemini 3.6 Flash launched in mid-July, while 3.7 Flash went online only three weeks afterward.
The new model delivers systematic improvements in coding and Agent capabilities. Benchmark results show FrontierCode 1.1 rose from 34.4% to 43.6%, DeepSWE v1.1 increased from 49.0% to 65.3%, and AutomationBench jumped from 17.0% to 30.4%. The model supports a 1 million-token context window and a maximum output length of 64K tokens, with three adjustable reasoning tiers: low, medium and high. Its official model ID is gemini-3.7-flash.
Under the limited-time promotion valid until December 31, 2026, the pricing is set at $0.75 per million input tokens and $3.75 per million output tokens, half of the original price of Gemini 3.6 Flash. Standard rates will double starting in 2027.
Google officially recommends the brand-new Interactions API as the primary access method. This interface abandons legacy parameters such as temperature, top_p, and top_k, marking a critical shift toward native Agent-oriented architecture.
Three Core Improvements in the 3.6 → 3.7 Flash Upgrade
Google has not disclosed detailed parameter scales or training data volumes. The company states performance gains stem from optimized training strategies and data curation. Three major capability upgrades are observed:
- Enhanced coding capacity: Better performance in code debugging, Web application generation, and higher fidelity when translating design drafts into front-end source code.
- Higher stability for Agent loops: More rigorous multi-step planning, lower failure rates for tool invocation, and active clarification of ambiguous requirements instead of falling into infinite loops.
- Progress in knowledge-intensive tasks: Harvey LAB-AA (legal & professional knowledge QA) reaches 90.7%, surpassing Claude Sonnet 5 (90.1%) and GPT-5.6 Terra (85.2%).
Benchmark Results: Leading in Coding, Yet Not the Top Choice for DeepSWE
The official test data covers self-comparison between 3.6 Flash and 3.7 Flash, alongside horizontal benchmarks against mainstream competitor models.
Self-improvement: Gemini 3.6 Flash vs Gemini 3.7 Flash
| Benchmark Dataset | Gemini 3.6 Flash | Gemini 3.7 Flash | Gain |
|---|---|---|---|
| FrontierCode 1.1 Main | 34.4% | 43.6% | +9.2 pp |
| DeepSWE v1.1 | 49.0% | 65.3% | +16.3 pp |
| WebDev Arena (Elo) | 1538 | 1588 | +50 |
| AutomationBench | 17.0% | 30.4% | +13.4 pp |
Horizontal Comparison with Mainstream Models
| Benchmark Dataset | Gemini 3.7 Flash | Claude Sonnet 5 | GPT-5.6 Terra |
|---|---|---|---|
| FrontierCode 1.1 Main | 43.6% | 42.7% | 41.3% |
| DeepSWE v1.1 | 65.3% | 53.8% | 69.6% |
| WebDev Arena (Elo) | 1588 | 1541 | 1523 |
| AutomationBench | 30.4% | 10.7% | 23.6% |
| Terminal-bench 2.1 | 85.8% | — | 87.4% |
| Harvey LAB-AA | 90.7% | 90.1% | 85.2% |
| HLE-Verified | 53.6% | 31.0% | 51.1% |
| OSWorld-2.0 | 47.9% | — | 50.2% |
Overall performance landscape:
- Gemini 3.7 Flash leads FrontierCode and AutomationBench (coding & multi-step automation).
- GPT-5.6 Terra achieves the highest score on DeepSWE (software engineering Agent) at 69.6%, followed by Gemini 3.7 Flash at 65.3%, ahead of Claude Sonnet 5’s 53.8%. For reference, DeepSeek V4 Pro scores 62.7% on DeepSWE.
- GPT-5.6 Terra maintains narrow advantages on Terminal-bench and OSWorld for terminal & desktop operation tasks.
Pricing Strategy: Half-price Promotion until End of 2026
| Type | Promotional Rate (Valid until 2026-12-31) | Standard Rate Starting 2027 |
|---|---|---|
| Input Tokens | $0.75 / Million | $1.50 / Million |
| Output Tokens | $3.75 / Million | $7.50 / Million |
The promotional price is approximately half of Gemini 3.6 Flash’s original pricing. Effectively, users receive a one-year half-price Pro tier before the end of 2026. Rates will automatically revert to standard pricing in 2027 without manual adjustment, and charges will follow official API billing records.
Independent testing using the open-source Browser Use Agent framework indicates Gemini 3.7 Flash delivers a roughly 35% lower effective cost compared to 3.6 Flash, driven by improved token efficiency combined with adjusted pricing. As a reference, Gemini 3.6 Flash shares the same promotional unit price as 3.7 Flash; the key difference lies in the promotion expiration date, with 3.7 Flash locked to December 31, 2026.
API Integration: Interactions API Becomes the Official Primary Method
Google simultaneously released the Interactions API alongside Gemini 3.7 Flash, designated as the official recommended access interface for the new model.
Core Differences Between Legacy Chat Completions and New Interactions API
| Item | Legacy Method (Chat Completions) | New Method (Interactions API) |
|---|---|---|
| Parameter Support | Supports temperature, top_p, top_k |
These parameters are removed and unsupported |
| Reasoning Control | thinking_budget (numeric value) |
thinking_level (low / medium / high) |
| Multi-turn Dialogue | Client-side maintains message array | Server-side manages conversation via previous_interaction_id |
| Pre-filling | Pre-fill dialogue turns supported | Pre-filling capability removed |
Three Tiers of thinking_level
| Tier | Suitable Scenarios | Token Consumption |
|---|---|---|
| low | Quick Q&A, real-time dialogue, data format conversion | Minimum |
| medium (Default) | Code debugging, Agent workflows, multi-step planning | Moderate |
| high | Complex mathematics, competitive programming, architecture design | Maximum |
Teams running legacy systems that still require Chat Completions format can continue using the old interface. However, official documentation prioritizes the Interactions API for all new development. When managing multiple LLM endpoints from different vendors, Treerouter can serve as a unified gateway to standardize heterogeneous API schemas and simplify migration workloads.
Side-by-Side Comparison: Gemini 3.7 Flash vs DeepSeek V4 Pro
Both models launched in August 2026 and focus heavily on Agent capabilities, enabling direct horizontal evaluation.
| Dimension | Gemini 3.7 Flash | DeepSeek V4 Pro |
|---|---|---|
| Context Window | 1M tokens | 1M tokens |
| Maximum Output Length | 64K tokens | 384K tokens |
| DeepSWE v1.1 | 65.3% | 62.7% |
| FrontierCode 1.1 | 43.6% | — |
| Reasoning Modes | low / medium / high | Non-Think / High / Max |
| Input Price (Promotion) | $0.75 / M | ~$0.43 / M |
| Output Price (Promotion) | $3.75 / M | ~$0.87 / M |
| API Protocol | Native Interactions API + OpenAI Compatible | OpenAI & Anthropic dual compatibility |
Selection reference:
- DeepSeek V4 Pro features lower token pricing and supports far longer maximum output (384K vs 64K).
- Gemini 3.7 Flash holds advantages in FrontierCode, AutomationBench and Harvey LAB-AA benchmarks, with native ecosystem integration for Agent platforms including Antigravity and Android Studio.
Release Background: New Leadership + Acceleration Pressure
Two major external factors shape this product launch:
- Koray Kavukcuoglu took charge of Google DeepMind. As former chief scientist, he holds complete decision power over the Gemini roadmap. Gemini 3.7 Flash is the first official model released under his leadership, with clear positioning centered on coding and Agent workloads, supported by comprehensive benchmark data, reflecting a more engineering-oriented product direction.
- Sergey Brin’s direct push. Multiple media reports confirm Brin requested the Gemini team to speed up iteration rhythms. The launch of 3.7 Flash only three weeks after 3.6 Flash directly reflects this pressure.
Google’s flagship high-end model (tentatively following the Gemini 3.5 Pro route) still lacks a confirmed release date. The Flash series continues to accumulate user traction by targeting practical workloads with balanced cost and performance.
Frequently Asked Questions
Q: Can I directly replace gemini-3.6-flash with gemini-3.7-flash by switching model ID?
Direct replacement will trigger errors if your code calls deprecated parameters such as temperature, top_p or top_k. Developers must follow Google’s official migration guide and update code before switching.
Q: The promotional price of 0.75/3.75 matches Gemini 3.6 Flash. Why is this launch regarded as a price cut? The promotional window is limited. After December 31, 2026, Gemini 3.7 Flash will revert to 1.50/7.50 pricing. In effect, users can access upgraded model performance at the original promotional rate of Gemini 3.6 Flash for the rest of the year.
Q: Is the 1M-token context window of Gemini 3.7 Flash reliable for Agent tasks? Official internal GDM-MRCR v2 long-context benchmark achieves 97.0% accuracy under 128K context. Google has not published independent public test data verifying uniform attention quality across the full 1M-token range.
Q: If I run Gemini, DeepSeek and other LLMs simultaneously, is there a unified management solution?
Teams can deploy a centralized routing layer to manage multi-model services. With unified gateway access, developers only need to adjust the model field in requests without maintaining separate API keys for every provider.
Q: Does Gemini 3.7 Flash support video and audio input? Google AI Studio documentation confirms multi-modal support including text, images, video and audio. Specific restrictions such as single video duration limits are defined in the official DeepMind developer reference.
Summary
The core value of Gemini 3.7 Flash lies in its prominent coding and Agent performance. It ranks first among mainstream models on FrontierCode 1.1 (43.6%) and AutomationBench (30.4%). Its DeepSWE score reaches 65.3%, outperforming Claude Sonnet 5, though falling slightly behind GPT-5.6 Terra.
The limited-time discounted pricing offers a valuable window to evaluate Google’s Agent capabilities before year-end. The rollout of Interactions API signals Google’s strategic shift toward native Agent architecture, and the removal of traditional sampling parameters marks a clear migration roadmap for developers.
All information in this article is current as of August 14, 2026, and subject to official announcements from Google AI.
Reference Links
- Official Gemini 3.7 Flash model page
- Google DeepMind Model Center
- Gemini Enterprise Agent Platform documentation
