Gemini 3.7 Flash Costs $0.75 Now. January 1, It Doubles.

TL;DR

Google shipped Gemini 3.7 Flash on August 13, three weeks after 3.6 Flash. The coding numbers are a real jump: DeepSWE v1.1 goes 49.0% to 65.3%, AutomationBench 17.0% to 30.4%. It lands at $0.75 per million input tokens, against $2.00 for both Claude Sonnet 5 and GPT-5.6 Terra. Two things the launch post does not put in bold: that price is introductory and doubles on January 1, 2027, and Google's own API docs list temperature, top_p, top_k and candidate_count as no longer supported on this model. The sampling knobs are gone.

Three weeks is the release cycle now

Gemini 3.6 Flash is not old. It shipped in late July. Twenty-one days later Google replaced it with a model that beats it on essentially every published eval, at half the list price of what 3.6 Flash was charging.

If you maintain a model-selection spreadsheet, it went stale while you were reading this paragraph. That is the actual operating condition for 2026: the mid-tier is now the contested tier, and the vendors are iterating there faster than most teams can finish a migration.

The gains are concentrated in coding and agents

Google published five head-to-head comparisons against 3.6 Flash. Every one of them is a software-engineering or automation eval, which tells you exactly who this model is aimed at.

WebDev Arena Elo moves from 1538 to 1588. On long context, Google reports 97.0% on GDM-MRCR v2 at 128k and 85.4% on LVBench for long video. The context window stays at 1M tokens with 64k max output, and the model id is gemini-3.7-flash.

Artificial Analysis independently scores it 56 on its Intelligence Index, at 340.1 output tokens per second and a 9.83 second time to first token. Fast on throughput, sluggish on the first byte, which matters more than people admit when you are the one watching a cursor blink.

The price is a promotion, not a price

$0.75 input and $3.75 output is genuinely aggressive for a 1M-context model with these coding scores. It is also, per Google's own documentation, valid through December 31, 2026. On January 1, 2027 it becomes $1.50 and $7.50. Exactly double, both sides.

Worth noting what the comparison class actually did this month. Anthropic made Sonnet 5's introductory $2/$10 permanent and cancelled a scheduled increase to $3/$15. Google went the other way and put an explicit expiry on the sticker. Same promotional mechanic, opposite endings, and only one of them requires you to diary a date.

Artificial Analysis puts the blended rate at $0.58 per million on a 7:2:1 cache-hit, input, output mix. That is a genuinely cheap agent loop today. Model your unit economics on the January number instead, or you will be repricing your product during the holidays.

You do not get to set temperature any more

This is the change that will actually break code, and it is buried in the docs rather than the blog post. On gemini-3.7-flash, Google lists temperature, top_p, top_k and candidate_count as no longer supported. Separately, thinking_budget is replaced by a thinking_level enum: low, medium (the default) and high. Passing minimal returns an error.

The swap from a token budget to a three-value enum is the shift worth understanding. thinking_budget let you say "spend up to 8,192 tokens reasoning about this." thinking_level lets you say "think a normal amount." It is the difference between setting your oven to 412 degrees and choosing low, medium or high. You lose precision you probably were not using well, and you gain a knob the vendor can retune underneath you without breaking your integration.

That last part cuts both ways. Google can improve what "medium" means in a later checkpoint and your app inherits it for free. Google can also change what "medium" costs you, and your app inherits that for free too.

Practical version: grep your codebase for those four parameter names before you point anything at gemini-3.7-flash. If you have a shared LLM client that sets temperature: 0 by default for deterministic extraction, and most of them do, that is your first breakage.

One benchmark went backwards

Not everything improved. On CharXiv, the chart-reasoning eval, 3.7 Flash scores 84.5% without tools against 85.2% for 3.6 Flash, and 88.7% with tools against 89.4%. Small regressions, 0.7 points each, in the same direction on both configurations.

If your workload is chart-heavy document extraction, dashboards, financial PDFs, scientific figures, that is the one place a blind upgrade could cost you accuracy. Run your own eval before you flip the model id.

How it stacks up on the composite

On the Artificial Analysis Intelligence Index, 3.7 Flash scores 56. GPT-5.6 Terra and Muse Spark 1.2 sit at 57, Claude Sonnet 5 at 55. In other words: it did not take the crown, it took the price. That is a legitimate strategy in a tier where the intelligence spread is two points and the price spread is nearly three times.

The caveat that applies to every launch-day post, including this one: the head-to-head numbers against 3.6 Flash are Google's own, run by Google, published by Google. The Artificial Analysis figures are independent. Treat the rest as a vendor's best day until someone else reproduces it.

Where you can actually use it

Google lists availability through AI Studio, Android Studio and Antigravity for developers, the Gemini Enterprise Agent Platform for companies, and Gemini Spark for AI Pro and Ultra subscribers across 160-plus countries. No open weights, as expected. This is an API model.

Key Takeaways

• Real coding gains in three weeks. DeepSWE v1.1 49.0% to 65.3%, AutomationBench 17.0% to 30.4%, FrontierCode 1.1 34.4% to 43.6%, all against 3.6 Flash on Google's own runs.
• The $0.75 price expires. Introductory rates of $0.75 input and $3.75 output run through December 31, 2026, then double to $1.50 and $7.50. Model your costs on the January number.
• Four API parameters are gone. temperature, top_p, top_k and candidate_count are no longer supported on this model, and thinking_budget is now thinking_level with three values.
• CharXiv regressed. 85.2% to 84.5% without tools, 89.4% to 88.7% with tools. Chart-heavy pipelines deserve a real eval before upgrading.
• It bought price, not the crown. Intelligence Index 56 against 57 for GPT-5.6 Terra and Muse Spark 1.2, 55 for Sonnet 5, at a fraction of the input cost.
• Anthropic went the other way. Sonnet 5's introductory $2/$10 became permanent and a hike to $3/$15 was cancelled. Google put a hard expiry on its discount instead.

Sources: Google blog: Introducing Gemini 3.7 Flash, Google AI for Developers: What's new in Gemini 3.7 Flash, Artificial Analysis: Gemini 3.7 Flash, Anthropic: Claude pricing, OfficeChai: Gemini 3.7 Flash benchmarks, Kingy AI: Gemini 3.7 Flash review, MarkTechPost: Google AI releases Gemini 3.7 Flash