Gemini 3.7 Flash lands three weeks after 3.6, at the same price

Google released Gemini 3.7 Flash on 13 August, calling it "our most intelligent workhorse model yet for coding and agents." The release comes, by Google's own count, just three weeks after Gemini 3.6 Flash — an unusually short gap for a model line that most teams treat as their default cheap workhorse. The capability gains Google reports are substantial. The pricing story attached to them is more complicated than the announcement's framing suggests.

What Google measured

All of the figures below come from Google's own testing, published in the launch post. No independent evaluation was available when this article was written, so treat them as vendor-reported until third-party numbers land. Compared with 3.6 Flash, Google reports:

• FrontierCode 1.1 Main: 43.6%, up from 34.4%
• DeepSWE v1.1 (debugging and issue resolution): 65.3%, up from 49.0%
• WebDev Arena Elo: 1588, up from 1538
• GDP.pdf (complex document comprehension): 34.0%, up from 22.0%
• AutomationBench (real-world business workflows): 30.4%, up from 17.0%

The DeepSWE and AutomationBench jumps are the ones worth pausing on. A model that resolves roughly two-thirds of issues rather than half changes how much supervision an agent loop needs, and AutomationBench nearly doubling suggests Google is targeting the failure mode where a cheap model handles the easy steps and stalls on the rest. Google also claims softer improvements that benchmarks capture poorly: better instruction fidelity, more diligent multi-step planning, and clarifying intent instead of guessing. Those are the qualities that decide whether an agent finishes a task or burns tokens retrying, and they are exactly the qualities you cannot verify from a scoreboard.

The price question

Google's post says 3.7 Flash arrives "with an introductory price of half the original 3.6 Flash cost per million tokens." That sentence is doing careful work. The comparison is against 3.6 Flash's original list price — not against what developers actually pay for 3.6 Flash today.

As of publication, Google's own pricing page lists Gemini 3.6 Flash and Gemini 3.7 Flash at identical rates, across every serving tier — standard, batch, flex and priority. Both are $0.75 per million input tokens and $3.75 per million output tokens through 31 December 2026. Both show $0.075 per million for context caching. Both list batch at $0.375 and $1.875.

So the practical position for anyone choosing between them today is not that 3.7 Flash is half the price of 3.6 Flash. It is that 3.7 Flash is the same price as 3.6 Flash, with better reported benchmarks. That is still a straightforward reason to migrate — it is simply a different claim than the one the announcement leads with.

The more consequential detail is in the footnote. Introductory pricing expires on 31 December 2026, and from 1 January 2027 the rate becomes $1.50 per million input tokens and $7.50 per million output tokens — double. Context caching doubles to $0.15, and batch to $0.75 and $3.75. If you are modelling the unit economics of an agent product on today's Flash pricing, that model has a hard expiry date about four and a half months out, and the correction is not marginal.

Introductory pricing expires on December 31, 2026. Starting January 1, 2027, $1.50/1M input tokens and $7.50/1M output tokens will apply.

What changes in the API

The model ID is `gemini-3.7-flash`. Per Google's developer documentation, it carries a 1M-token context window with a 64k maximum output, and exposes tunable thinking levels: low for latency-critical work, medium as the default, and high for maximum reasoning. It supports the same built-in tools as its predecessor, which means for most codebases the migration is a string change rather than a rewrite.

The thinking-level control is the lever that matters for cost. Output pricing includes thinking tokens, so on a model whose output rate is five times its input rate — and which doubles in January — the difference between running everything at high and reserving high for the steps that need it is a real line item, not a tuning detail. If you are adopting 3.7 Flash, setting thinking levels per call rather than globally is the single change most likely to pay for itself.

Where it runs

• Developers: the Gemini API via Google AI Studio and Android Studio, plus Google Antigravity for agent-first workflows
• Enterprises: Gemini Enterprise Agent Platform and the Gemini Enterprise app
• Individuals: Gemini Spark, Google's always-on personal agent, for AI Pro and Ultra subscribers in 160+ countries, switched to 3.7 Flash as of the announcement

Google also says 3.7 Flash ships with updated safeguards against misuse in chemical, biological, radiological and nuclear domains and in cyber offense, and points to the model card for detail. As with the benchmarks, that is Google describing its own safety work; the claim has not been independently assessed.

The honest summary

A materially better workhorse model at an unchanged price is good news, and a three-week iteration cycle on the Flash line tells you something about how fast the cheap tier is now moving. But two things in this release deserve to be read plainly rather than as announced: the price is not a cut relative to what 3.6 Flash costs today, and the rate everyone is budgeting against doubles on 1 January.