Google launches Gemini 3.7 Flash: faster coding agents at $0.75 per million tokens

Google has released another Flash model, and it came fast. Gemini 3.7 Flash launched on August 13, just three weeks after Gemini 3.6 Flash. Google calls it its "most intelligent workhorse model yet for coding and agents."

The pitch has two parts. Google says 3.7 Flash makes big gains in software engineering, knowledge work and web development. It also launches at an introductory price that, in Google's words, is "half the original 3.6 Flash cost per million tokens."

That wording matters, and we explain why below. First, what shipped.

What Google announced

Gemini 3.7 Flash is available now in the Gemini API through Google AI Studio, in the Gemini Enterprise Agent Platform, in the Gemini Enterprise app and in Google's Antigravity coding tool. It also now powers Gemini Spark, the personal agent in the Gemini app.

According to the DeepMind model card, 3.7 Flash is built on 3.6 Flash, with "algorithmic improvements to its core reasoning foundation" and support for agentic video understanding. It accepts text, images, audio and video, has a context window of up to 1 million tokens and outputs text of up to 64K tokens.

Google says 3.7 Flash grew directly out of developer feedback on 3.6 Flash. The company also says it adapts better to roadblocks, asks for clarification when needed and follows instructions more closely. It "thinks more diligently," putting more effort into multi-step planning and tool calls.

The benchmark jump over 3.6 Flash

For a release only three weeks after the last one, Google's reported gains are large. These are Google's own figures from the launch post:

The DeepSWE result stands out, rising by more than 16 points. Google says the model reaches correct code on the first try more often and writes more production-ready code. It also says 3.7 Flash builds more complete apps in fewer prompts and closely matches reference designs, whether the reference is a screenshot or a full design system.

One small inconsistency is worth noting. The model card lists 3.6 Flash at 48.6% on DeepSWE, while the blog post says 49.0%. The difference is minor, but it shows how much these numbers move depending on the run.

How it compares with rivals

The model card also compares 3.7 Flash with competitors in the same price range: Anthropic's Claude Sonnet 5, OpenAI's GPT-5.6 Terra and Meta's Muse Spark 1.2. All results are Google-reported.

The overall picture is a mixed win. Google's own table puts 3.7 Flash ahead on production code quality, business automation, document comprehension, long-video understanding and long-context retrieval. GPT-5.6 Terra still leads on DeepSWE, Terminal-bench and computer use. All three rivals score higher on GDPVal-AA knowledge work, with Muse Spark 1.2 on top.

Key takeaway: Gemini 3.7 Flash doesn't beat every rival on every test, but by Google's numbers it gets close to GPT-5.6 Terra on coding for well under half the per-token price.

Pricing context matters here. Google lists Claude Sonnet 5 at a temporarily discounted $2/$10. GPT-5.6 Terra's $2/$12 reflects the price cut OpenAI made on July 30.

The price: read the footnote

Google's list price for 3.7 Flash is $0.75 per million input tokens and $3.75 per million output tokens. Output pricing includes thinking tokens. That rate is introductory and runs through December 31, 2026. From January 1, 2027, the price doubles to $1.50 input and $7.50 output.

Here's the detail. When 3.6 Flash launched on July 21, it cost $1.50/$7.50. Google's 3.7 Flash model card now lists both 3.6 Flash and 3.7 Flash at $0.75/$3.75, with a shared footnote saying the introductory price for both models ends on December 31. Google Cloud's Agent Platform pricing page says the same.

So "half the original 3.6 Flash cost" is accurate, but moving an existing 3.6 Flash workload to 3.7 Flash won't lower your per-token bill today. The two now cost the same. The case for switching is capability, plus whatever token savings the new model gets on your tasks.

Early customers are reporting savings through efficiency. On Google DeepMind's model page, Browser Use co-founder and CTO Gregor Zunic says its 3.7 Flash agent "was 35% cheaper than 3.6 Flash, with a +8% observed prompt-cache hit rate and fewer tool errors." That is one company's workload, but it shows where savings are likely to come from.

Gemini Spark gets the upgrade

The consumer side of this launch is Gemini Spark. Google introduced Spark at I/O in May as a personal AI agent that runs in the background 24/7 and acts on your behalf under your direction. Spark now uses 3.7 Flash.

Google says the new model makes Spark more efficient at knowledge work, with better tool use in Google Workspace apps and higher accuracy on complex, multi-step workflows. Its examples are everyday office work: consolidating files, drafting emails and updating status documents.

Google's launch post says Spark is available to Google AI Pro and Ultra subscribers in more than 160 countries. On July 30, Google expanded Pro access to those markets and added Chrome auto-browse, which lets Spark use your logged-in accounts to handle web errands. Google says auto-browse is launching in the US first.

What developers need to know

If you are moving from 3.5 Flash or 3.6 Flash, a few steps come first. Google Cloud's developer guide for 3.7 Flash lists these changes:

1. Point requests at the model ID gemini-3.7-flash.
2. Control reasoning with thinking_level (LOW, MEDIUM by default, or HIGH) instead of the older thinking_budget field. MINIMAL is not supported on this model.
3. Remove sampling parameters the model no longer accepts, including temperature, top_p, top_k, frequency_penalty, presence_penalty and candidate_count.
4. Remove prefilled model turns from multi-turn chat histories, and make sure function responses match their calls by ID and name.
5. Re-test your prompts and cost profile at your chosen thinking level before switching production traffic.

Losing temperature and top_p will annoy teams that tuned them carefully. Google is betting that the thinking level is the only control most agent builders need.

The model card lists a March 2026 knowledge cutoff, with the caveat that some domains may only reflect knowledge up to January 2025, in line with the rest of the Gemini 3 family.

Safety notes

Google says 3.7 Flash ships with updated safeguards against misuse in chemical, biological, radiological and nuclear (CBRN) domains and in cyber offense.

Under its Frontier Safety Framework, Google found that the model did not reach any critical capability level. But the model card says 3.7 Flash reached the "alert threshold" for both the CBRN uplift and cybersecurity critical levels. Google concluded it stays below those levels and says it continues to deploy mitigations. That is candid disclosure, and it's worth tracking as Flash models keep getting more capable.

Why it matters

Google is releasing Flash updates faster than most rivals update their mid-tier models. Two Flash releases in three weeks, each promising agent gains, means teams building on Gemini have to evaluate new models more often.

It also shows where the competition is. Flash-class models are the ones that run agent workloads at scale, and cutting their price to $0.75/$3.75 puts pressure on Anthropic and OpenAI's mid-tier pricing. Meanwhile, Gemini 3.5 Pro still hasn't shipped. Google said at I/O it would arrive the following month, and in July said only that it was testing with partners.

The price increase on January 1 is the main thing to plan for. Anyone building a 2027 budget on 3.7 Flash should use $1.50/$7.50, not the launch price.

FAQ

How much does Gemini 3.7 Flash cost?

In the Gemini API, it costs $0.75 per million input tokens and $3.75 per million output tokens (thinking tokens included) through December 31, 2026. From January 1, 2027, it rises to $1.50 and $7.50.

Is Gemini 3.7 Flash cheaper than 3.6 Flash?

Only compared with 3.6 Flash's original launch price. Google's model card and Cloud pricing page now list both models at the same $0.75/$3.75 introductory rate, so switching doesn't lower your per-token price. The gains come from capability and possibly fewer tokens per task.

Where can I use Gemini 3.7 Flash?

It is available in the Gemini API through Google AI Studio, Gemini Enterprise Agent Platform, the Gemini Enterprise app and Google Antigravity. Consumers get it through Gemini Spark if they have a Google AI Pro or Ultra subscription.

Does it beat GPT-5.6 Terra and Claude Sonnet 5?

On some tests. Google's model card shows 3.7 Flash ahead on production code quality, business automation and document comprehension, and behind GPT-5.6 Terra on DeepSWE, Terminal-bench and computer use. These are vendor-reported results.

What changes for existing Gemini integrations?

Use the gemini-3.7-flash model ID, switch to thinking_level for reasoning control, and remove unsupported sampling parameters such as temperature and top_p. Re-run your evaluations before moving production traffic.

The bottom line

Gemini 3.7 Flash is a big update to Google's workhorse model, with double-digit gains over 3.6 Flash on several coding and automation tests. It ships at a price that undercuts its closest rivals.

Just be clear about what the price cut is. Google also moved 3.6 Flash to the same introductory rate, and both double on January 1. For developers, the reason to switch is better agents. For Spark subscribers, the upgrade happens automatically.