Google Launches Gemini 3.7 Flash: Hybrid Reasoning for Coding, Agents, and Knowledge Work
Gemini 3.7 Flash arrives August 13, 2026 with native hybrid reasoning, controllable thinking budgets, frontier coding benchmarks, and 50% introductory pricing through December.
A new flagship for hybrid reasoning
On August 13, 2026, Google released Gemini 3.7 Flash — its flagship hybrid reasoning model built for coding, agents, and complex knowledge work. Unlike earlier Flash tiers that traded depth for speed, 3.7 Flash ships with native controllable reasoning baked into the API: you dial thinking depth up or down per request instead of picking a separate “reasoning model.”
The launch targets three audiences at once: developers wiring production agents, teams running long-horizon coding tasks, and enterprises that need reliable multimodal analysis over massive document sets. Google positions 3.7 Flash as the model that finally makes one endpoint work for both sub-100ms chat replies and multi-minute deep research runs.
Native hybrid reasoning and the thinking budget
The centerpiece is thinking_config.thinking_budget — a single integer that controls how many internal reasoning tokens the model may spend before answering.
thinking_budget | Behavior | Best for |
|---|---|---|
| 0 | Ultra-low latency fast mode (sub-85ms first token) | Live chat, autocomplete, high-QPS routing |
| 256–1024 | Light multi-step planning | Tool routing, simple refactors, RAG synthesis |
| 4096–16384 | Deep multi-step reasoning | Agent loops, complex debugging, research reports |
| 32768–65536 | Maximum depth | Long-horizon coding, multi-file architecture changes |
Set the budget in the Gemini API like this:
from google import genai
from google.genai import types
client = genai.Client()
response = client.models.generate_content(
model="gemini-3.7-flash",
contents="Refactor this React component to use hooks and add error boundaries.",
config=types.GenerateContentConfig(
thinking_config=types.ThinkingConfig(
thinking_budget=8192 # 0 = fast mode; up to 65536 for deep reasoning
)
),
)
print(response.text)Because reasoning is native rather than bolted on, latency scales smoothly: a thinking_budget of 0 behaves like a classic Flash model, while higher values unlock chain-of-thought quality without switching model IDs. Billing separates visible output tokens from internal thinking tokens, so you only pay for the depth you request.
Benchmark highlights
Google published head-to-head numbers against Gemini 3.6 Flash, Claude Sonnet 5, and GPT-5.6 Terra. 3.7 Flash leads on every eval listed:
| Benchmark | Gemini 3.7 Flash | Gemini 3.6 Flash | Claude Sonnet 5 | GPT-5.6 Terra |
|---|---|---|---|---|
| FrontierCode 1.1 | 43.6% | 34.4% | 39.1% | 37.8% |
| DeepSWE v1.1 | 65.3% | 49.0% | 56.2% | 53.4% |
| WebDev Arena (Elo) | 1588 | 1538 | 1552 | 1544 |
| GDP.pdf | 34.0% | 22.0% | 28.5% | 26.8% |
| AutomationBench | 30.4% | 17.0% | 23.1% | 21.5% |
| GDM-MRCR v2 | 97.0% | 91.2% | 94.5% | 93.1% |
| HLE-Verified | 53.6% | 44.8% | 48.2% | 46.7% |
Two numbers stand out for builders:
- DeepSWE v1.1 at 65.3% — a 16-point jump over 3.6 Flash, signaling that agentic coding is the primary training target.
- GDM-MRCR v2 at 97.0% — near-perfect long-context retrieval across a 2.5M-token window, critical for document-heavy workflows.
For context on the previous generation, see our Gemini 3.6 Flash vs 3.5 Flash comparison.
Pricing and 50% introductory discount
Google is running an aggressive launch promotion:
| Intro price (through Dec 31, 2026) | Standard price (from Jan 1, 2027) | |
|---|---|---|
| Input | $0.75 / 1M tokens | $1.50 / 1M tokens |
| Output | $3.75 / 1M tokens | $7.50 / 1M tokens |
| Context caching | $0.075 / 1M tokens | $0.075 / 1M tokens |
That is a 50% discount on both input and output for the remainder of 2026. Context caching — useful when you reuse large system prompts or document corpora — stays at $0.075 per million tokens regardless of the promo window.
Combined with fewer tool-call round trips at higher thinking budgets, effective cost per completed agent task can drop well below the per-token sticker price. Our API cost estimate guide walks through the math for typical workloads.
Multimodal, web dev, and agent capabilities
Beyond benchmarks, 3.7 Flash ships with production-ready capabilities that matter in real deployments:
- 2.5M-token context window — ingest entire codebases, contract bundles, or research corpora in one pass.
- 245 output tokens/s — fast enough to stream UI code and long reports without throttling user experience.
- 99.7% JSON tool execution precision — reliable structured outputs for agent tool chains.
- Computer use — built-in client-side automation for browser and desktop workflows.
- UI generation fidelity — higher layout accuracy on WebDev Arena (1588 Elo) translates to cleaner generated frontends out of the box.
These features are available today through the Gemini API, Google AI Studio, Gemini Enterprise, and Google Antigravity.
Developer guide and migration tips
If you are upgrading from 3.6 Flash or 3.5 Flash-Lite, start with these patterns:
- Replace model ID — swap
gemini-3.6-flashforgemini-3.7-flashin existing calls; the API surface is backward compatible. - Set thinking budgets per task type — use
0for user-facing chat,1024for RAG Q&A,8192+for coding agents, and32768only for overnight batch jobs where latency is irrelevant. - Enable context caching — if your system prompt or document set exceeds 32K tokens and repeats across requests, caching cuts input cost by 90%.
- Tune tool schemas — the 99.7% JSON precision reward is strict schemas; loose or ambiguous tool definitions still fail at the application layer.
- Monitor thinking token usage — log
usage_metadatato understand which agent steps actually need deep reasoning versus fast mode.
For agentic workflow patterns, see our 3.5 Flash-Lite agentic workflows guide — the thinking budget concept extends naturally to 3.7 Flash with finer granularity.
Bottom line
Gemini 3.7 Flash is the first Flash-tier model where speed and depth are not mutually exclusive. The controllable thinking_budget lets one model serve latency-critical chat and deep agentic coding alike, and the launch pricing makes experimentation cheap through the end of 2026.