Gemini 3.7 Flash Is Half Price Until New Year's Day

Google released Gemini 3.7 Flash on Thursday, and the most interesting number attached to it isn't a benchmark — it's the price tag, which has an expiration date. Developers pay 75 cents per million input tokens and $3.75 per million output tokens right now. That rate runs through December 31. On January 1, 2027, it doubles to $1.50 and $7.50.

The model is aimed at two jobs Google is betting companies will pay real money for: writing and repairing software, and powering AI agents — programs that plan a task, choose their own tools and grind through several steps without a person approving each click. According to Reuters, Google is pitching it as the low-cost option for businesses building exactly that kind of hands-off automation.

The gains are not small. On DeepSWE v1.1, a software-engineering test, the new model scored 65.3% against 49.0% for the version it replaces. And that version, Gemini 3.6 Flash, was three weeks old. Even if you never touch an API, this matters: cheap agent models are what make a 24/7 assistant viable inside a $19.99-a-month consumer subscription instead of a $200 one.

What Gemini 3.7 Flash Actually Does Better

Google published a short scorecard comparing the two releases. Every figure below is the company's own testing, not an independent lab's, which is the standard caveat for every model launch from every vendor this year.

• FrontierCode 1.1 Main: 43.6%, up from 34.4%
• DeepSWE v1.1 (issue resolution): 65.3%, up from 49.0%
• WebDev Arena, an Elo rating for building web interfaces: 1588, up from 1538
• Document reasoning: 34.0%, up from 22.0%
• AutomationBench, for multi-step tool use: 30.4%, up from 17.0%

Translated into plain terms: it fixes more real bugs, builds more usable front-end layouts on the first try, and reads dense documents with fewer mistakes. Google says it can produce feature-complete apps in fewer prompts, which is the metric developers actually feel — fewer rounds of "no, not like that."

What This Costs If You Actually Run It

One Coding Job, Priced Out

Token pricing is abstract until you do the arithmetic. Say an agent reads 200,000 tokens of your codebase and writes back 20,000 tokens of patches and explanation. At today's rate that's about 15 cents in, 7.5 cents out — roughly 23 cents per run. Fire that off 100 times a day across a small engineering team and you're near $23 a day, or about $680 a month.

The January 1 Jump

Now run the same math on the permanent rate. That $680 becomes roughly $1,360 a month on New Year's Day, with zero change in how much work you're doing. Any US team drafting a 2027 budget off Q4 invoices will be short by exactly half. Teams that lock in annual cloud commitments this quarter should be modeling the post-discount number, not the promotional one.

The Number Google Isn't Putting on the Poster

Look again at AutomationBench: 30.4%. That's nearly double the previous score, and it's the benchmark closest to what Google is selling — an agent that handles a whole workflow. It also means the model misses roughly seven of every ten of those tasks. Agent benchmarks are deliberately brutal, so a low score isn't a scandal. But it's a useful reality check against demo videos where everything runs clean.

The practical read: this thing is genuinely good at bounded, well-described work, like fixing a specific failing test. It is not yet reliable enough to leave alone with your billing system overnight. Keep a human on the approval step for anything that touches money, customer data or production deploys.

Why New Models Now Land Every Three Weeks

Three weeks between major releases would have been unthinkable two years ago. It's now the cadence, and the reason is competitive pressure at the mid-tier — the workhorse models that handle the boring bulk of enterprise traffic, where cost per token decides who wins the contract. Halving the price for four months is a customer-acquisition move dressed as a launch discount.

Google has been running the same play on consumers. It cut its top AI Ultra plan from $249.99 to $99.99 a month at its developer conference in May, then opened Gemini Spark, its background agent, to $19.99 Pro subscribers in the US in July. Gemini 3.7 Flash is now the engine underneath Spark for AI Pro and Ultra subscribers in more than 160 countries.

Where to Try It and What to Watch

Developers can reach it through Google AI Studio, the Gemini API, Android Studio and Antigravity, Google's agent-building environment. Enterprises get it in the Gemini Enterprise Agent Platform, per TestingCatalog. There's no waitlist — it's live now.

If you're evaluating it, do the benchmarking during the discount window and record what a real workload costs you, not what a demo costs. Then double that figure and decide whether it still clears your bar. Put December 31 on the calendar, and watch whether Google extends the promotional rate — if a rival ships something cheaper first, it probably will.