Google released Gemini 3.7 Flash on August 13, 2026, calling it the company’s most intelligent workhorse model yet for coding and agents. The model posts double-digit gains over Gemini 3.6 Flash on production coding and automation tests while shipping at an introductory rate of $0.75 per million input tokens and $3.75 per million output tokens through December 31.
That price is half the original 3.6 Flash rate. For teams running multi-step agents that burn tokens on planning, tool calls and retries, the second-order effect is immediate: volume workloads that once looked expensive become practical.
The release pairs those unit economics with stronger first-pass results on coding and workflow benches. Together they change how teams size agent traffic, cache strategy and retry budgets before the calendar flips to standard rates.
What 3.7 Flash Improves
Tulsee Doshi, Senior Director of Product Management for the Gemini team, wrote in the official Gemini 3.7 Flash announcement that the release follows developer feedback just three weeks after 3.6 Flash. Gains concentrate on software engineering, knowledge work and web development.
On FrontierCode 1.1 Main the model reaches 43.6 percent versus 34.4 percent for 3.6 Flash. DeepSWE v1.1 rises to 65.3 percent from 49.0 percent or 48.6 percent depending on the exact prior figure reported. WebDev Arena Elo climbs to 1588 from 1538. AutomationBench, which tests real business workflows, jumps to 30.4 percent from 17.0 percent. GDP.pdf document comprehension moves to 34.0 percent from 22.0 percent.
Those lifts cluster where agents spend the most tokens: writing and repairing code, following multi-step business flows, and reading long documents. A higher first-pass hit rate shortens the plan-call-inspect loop that dominates production spend.
- FrontierCode 1.1 Main: 34.4% to 43.6%
- DeepSWE v1.1: about 49% to 65.3%
- WebDev Arena Elo: 1538 to 1588
- AutomationBench: 17.0% to 30.4%
- GDP.pdf: 22.0% to 34.0%
The model supports a 1,048,576-token input window and 65,536-token output, with thinking levels of low, medium (default) and high. It accepts text, image, video, audio and PDF. Function calling, code execution, computer use (preview), search grounding and structured outputs are all supported.
Keeping the same context ceiling and tool surface as the prior Flash generation means teams can swap models without rewriting retrieval pipelines or tool schemas. The intelligence jump arrives inside a familiar envelope.
The Price Cut Changes Agent Math
Introductory pricing lasts until December 31, 2026. On January 1, 2027 standard rates of $1.50 input and $7.50 output take effect for both 3.7 and 3.6 Flash. Context caching sits at $0.075 per million tokens during the intro window.
Agents rarely finish in one shot. They plan, call tools, inspect results, recover from errors and loop. Each extra step multiplies token spend. A model that reduces failed loops and retries therefore cuts total cost even before the sticker price falls. Browser Use co-founder and CTO Gregor Zunic reported the 3.7 Flash agent ran 35 percent cheaper than 3.6 Flash in their tests, with higher prompt-cache hits and fewer tool errors.
- Intro input: $0.75 per 1M tokens through year-end
- Intro output: $3.75 per 1M tokens (includes thinking tokens)
- Standard after Jan 1 2027: $1.50 / $7.50
- Context window: 1M input, 64k-65k output
Compare that with the figures Google lists for rivals on its own table: Claude Sonnet 5 at roughly $2 / $10 and GPT-5.6 Terra at $2 / $12. Even after the intro period ends, 3.7 Flash undercuts those list prices. Teams watching token cost pressures already slowing some big AI bets now have a clearer path to keep agent volume high without the same bill shock.
Output pricing matters more than input for agent stacks because thinking tokens and tool-mediated replies land on the output meter. The intro $3.75 rate, which includes thinking tokens, keeps deep reasoning affordable while teams tune prompts and cache hit rates.
Context caching at $0.075 per million tokens during the intro window further rewards stable system prompts and repeated tool schemas. Higher cache hits, the pattern Zunic described, compound the headline discount.
Where the Numbers Stack Up
Google published a broad comparison on the DeepMind benchmark table for 3.7 Flash. Selected results:
| Benchmark | Gemini 3.7 Flash | Gemini 3.6 Flash | Claude Sonnet 5 | GPT-5.6 Terra |
|---|---|---|---|---|
| FrontierCode 1.1 Main | 43.6% | 34.4% | 42.7% | 41.3% |
| DeepSWE v1.1 | 65.3% | 48.6% | 53.8% | 69.6% |
| Code Arena WebDev Elo | 1588 | 1538 | 1541 | 1523 |
| AutomationBench | 30.4% | 17.0% | 10.7% | 23.6% |
| GDP.pdf | 34.0% | 22.0% | 28.0% | 24.7% |
| Terminal-bench 2.1 | 85.8% | 78.0% | 80.4% | 87.4% |
| Input price $/1M (intro) | $0.75 | $0.75 | $2.00 | $2.00 |
3.7 Flash leads or matches closely on several coding and workflow tests while trailing on some longer-horizon agent suites. The pattern favors reliable multi-step execution at Flash latency and cost.
AutomationBench is the clearest gap versus both the prior Flash model and the listed rivals. Real business workflows are exactly the traffic that burns retries when a step fails, so a jump from 17.0 percent to 30.4 percent lands directly on operating cost.
FrontierCode and WebDev Arena Elo put the model at or above the mid-tier frontier names Google listed, while intro input pricing stays at less than half of those $2 list rates. Teams that standardize on one workhorse for coding agents can trade a narrow loss on one suite for broad gains on others without leaving the Flash price band.
Thinking Levels and Developer Controls
Developers set thinking_level to low, medium or high. Low trims latency for chat or incident response. Medium is the default and recommended for most coding and agent work. High extends reasoning and tool use for hard math or complex code at higher token cost.
The 1M token context and tool specs match the prior Flash generation. Computer use remains in preview. The model is now the default behind the Antigravity agent in Gemini Managed Agents and the Google Antigravity SDK. Gemini Spark for Google AI Pro and Ultra subscribers also switched to 3.7 Flash the same day.
Routing by thinking level lets one deployment serve mixed traffic. Support chat can stay on low. Core coding agents hold medium. Overnight batch jobs that refactor hard modules can step up to high without a separate model ID.
On Box’s evaluation of real enterprise knowledge work, Gemini 3.7 Flash was both more accurate and significantly faster than the prior model, with its largest gains on the most challenging analytical tasks.
Yashodha Bhavnani, VP of AI Products at Box, said that kind of progress expands what AI can take on for organizations.
Who Feels the Shift First
Individual developers and startups building coding assistants, support agents or internal automation gain the most immediate runway. Lower per-token cost plus fewer retries lets them serve more users before margins tighten. Enterprise knowledge-work teams already testing on Box, Databricks, Harvey and Hebbia reported measurable lifts in accuracy and precision at lower effective cost.
Rivals that price mid-tier models near $2 input face pressure to match either quality or unit economics. Google’s own table shows 3.7 Flash competitive with mid-sized frontier models on several agent and coding suites while staying well below their list prices during the intro window.
For startups, the intro window is a planning horizon as much as a discount. Shipping product features against $0.75 / $3.75 math, then re-checking margins at $1.50 / $7.50, is simpler when the same model ID carries both eras. Enterprises already on Box and similar stacks can point to accuracy and speed gains when they ask finance to fund wider rollouts.
Flash Line Moves Faster Than Before
Gemini 3.6 Flash arrived July 21, 2026 at the then-standard $1.50 / $7.50. Three weeks later 3.7 Flash lands with higher scores and the temporary half-price offer that also applies retroactively to 3.6. The cadence is aggressive. Earlier Flash releases often stretched longer between meaningful intelligence jumps.
- July 21, 2026: Gemini 3.6 Flash ships at $1.50 input / $7.50 output.
- August 13, 2026: Gemini 3.7 Flash launches with double-digit bench gains and intro rates of $0.75 / $3.75 that also cover 3.6.
- December 31, 2026: Introductory pricing window closes.
- January 1, 2027: Standard rates of $1.50 / $7.50 apply to both 3.7 and 3.6 Flash.
Availability spans the Gemini API, Google AI Studio, Vertex AI / Gemini Enterprise Agent Platform, the Gemini app and Antigravity. Full details sit on the current Gemini API pricing page. Safety updates cover CBRN and cyber offense domains per Google’s frontier safeguards.
Retroactive discounting on 3.6 Flash lowers the cost of staying put while teams validate 3.7. That reduces switching friction and keeps the prior model from becoming a stranded tier during the evaluation stretch.
Production Reality Check
Headline benches never equal every production stack. Prompt design, retrieval quality, tool reliability and human oversight still decide outcomes. The temporary nature of the intro pricing means teams should model both the $0.75 / $3.75 window and the post-2026 rates before locking long-term architecture.
Still, the combination of stronger first-pass code, better multi-step planning and halved inference cost removes a major barrier that kept many agent projects in pilot. High-volume loops no longer look like a luxury reserved for the largest budgets.
Computer use remaining in preview is a practical boundary. Agents that depend on full desktop control should keep human review paths and fallbacks until that surface leaves preview. Function calling, code execution, search grounding and structured outputs are already in the supported set for production designs that stay inside those tools.
Half Price Meets Stronger First Passes
The cost story is easy to oversimplify as a sticker cut alone. Zunic’s 35 percent cheaper agent run tied the savings to higher prompt-cache hits and fewer tool errors, not only to the $0.75 input rate. Quality and price move together.
When AutomationBench nearly doubles and FrontierCode jumps by roughly nine points, fewer repair cycles follow. Each avoided loop skips another round of planning tokens, tool payloads and thinking output. That is why volume workloads tip from expensive to practical even before finance models the January 2027 step-up.
Intro output at $3.75, with thinking tokens included, keeps the high thinking level usable for hard cases without an immediate bill spike. Teams can reserve high for the hardest jobs, hold medium as the default, and still land under rival list prices on Google’s comparison table.
Rivals Face a Tighter Mid-Tier Band
Google’s published table places 3.7 Flash against Claude Sonnet 5 and GPT-5.6 Terra on both scores and list pricing. On AutomationBench, GDP.pdf and WebDev Arena Elo, 3.7 Flash leads the group in the figures shown. On DeepSWE and Terminal-bench it trails GPT-5.6 Terra while still beating 3.6 Flash by a wide margin.
Input list prices for those rivals sit near $2 per million tokens, against 3.7 Flash intro input at $0.75 and a post-intro rate of $1.50. Even after the promotional window, the Flash tier undercuts the mid-tier names Google listed. That gap is the pressure point for anyone selling a similarly positioned coding or agent model.
Buyers already weighing token cost pressures now have a concrete alternate path: stay inside Flash latency and pricing, pick up the measured gains on coding and workflow suites, and revisit architecture only if a longer-horizon agent suite demands a different tradeoff.
Frequently Asked Questions
What is the exact introductory price for Gemini 3.7 Flash?
Through December 31, 2026 the paid tier charges $0.75 per million input tokens and $3.75 per million output tokens, including thinking tokens; context caching is $0.075 per million. From January 1, 2027 those rates double to $1.50 and $7.50.
How large is the context window on Gemini 3.7 Flash?
The model accepts up to 1,048,576 input tokens and can generate up to 65,536 output tokens, matching the prior Flash generation while adding tunable thinking levels.
Which coding benchmarks improved most versus 3.6 Flash?
FrontierCode 1.1 Main rose from 34.4 percent to 43.6 percent, DeepSWE v1.1 from roughly 49 percent to 65.3 percent, and WebDev Arena Elo from 1538 to 1588, according to Google’s published figures.
Can developers control how much the model reasons?
Yes. The thinking_level parameter accepts low for fast responses, medium as the balanced default, or high for maximum tool use and multi-step depth at higher token cost.
Where is Gemini 3.7 Flash available today?
It is generally available via the Gemini API, Google AI Studio, Google Cloud Vertex AI / Gemini Enterprise surfaces, the Gemini app for Spark, and as the default model for Antigravity agents.
