Google cut the price of its agent-building AI model in half. That's the number that decides whether an always-on submittal bot pencils out
Gemini 3.7 Flash launched August 13 at half the token price of its predecessor, with the biggest score jump on Google's own benchmark for enterprise workflow automation. The model itself isn't a construction tool — but the cost curve it represents is what determines whether an AI agent runs on every submittal or only the ones someone remembers to flag.
Google released Gemini 3.7 Flash on August 13, its model built for coding and AI agents, at half the per-token price of the version it replaced. The bigger number, buried in Google's own benchmark table, is a 79% relative jump on the test Google uses for enterprise workflow automation. Neither fact touches a jobsite directly. Both determine whether the AI agent inside your project-management software runs on every document or just the ones somebody remembers to feed it.
What did Google actually release?
Gemini 3.7 Flash is a fast, lower-cost model — not Google's flagship "Pro" tier — that the company is pitching as its best model yet for coding and for building AI agents. It shipped three weeks after Gemini 3.6 Flash, at an introductory price of $0.75 per million input tokens and $3.75 per million output tokens, half of 3.6 Flash's rate; the price is set to roughly double, to $1.50 and $7.50, once the introductory window ends after 2026. On Google's own coding benchmarks, it scored 43.6% on FrontierCode (up from 34.4%) and 65.3% on DeepSWE v1.1 (up from 49.0%). It's live now in Google AI Studio, Android Studio, Google Antigravity, the Gemini Enterprise Agent Platform, and the Gemini app through Google's Spark assistant.
Why does a coding-model price cut matter to a construction back office?
Because the model isn't marketed as a coding tool alone — Google specifically benchmarked it on AutomationBench, its internal test for enterprise workflow automation, where Gemini 3.7 Flash scored 30.4%, nearly double the 17.0% of its predecessor and well ahead of the 23.6% Google recorded for GPT-5.6 Terra. That's the category of task an RFI-triage bot, a submittal-logging agent, or a daily-log summarizer actually runs: reading a document, extracting what matters, taking an action, doing it again a hundred times a day without a person in the loop each time.
The economics of that loop are what decide whether a vendor runs an agent on every document that comes in, or rations it to a sample because the per-call cost adds up. A submittal-review agent processing a spec section of roughly 20,000 tokens and returning a 2,000-token compliance summary would cost about 2.4 cents per document at Gemini 3.7 Flash's introductory rate — half of what the identical task cost on the model it replaced. That's a back-of-envelope figure for illustration, not a vendor's actual bill, but it's the kind of math a software company runs before deciding whether "AI reviews every submittal" is a feature they can afford to ship by default instead of gating behind a premium tier.
Who does this actually reach?
| Role | What runs on this kind of model today | What a lower cost curve changes |
|---|---|---|
| Submittal coordinator | AI-assisted first-pass review, gated to flagged items | Wider coverage — more submittals get an automated first pass instead of only the ones a human queues up |
| PM tracking RFIs | AI drafting or routing suggestions on a subset of RFIs | Cheaper per-call cost makes it more affordable for a vendor to run triage on the full RFI log, not a sample |
| Estimator | Spec extraction tools priced per seat or per document batch | Vendor margin on those tools improves, or the vendor passes the savings through as a lower price |
None of this is guaranteed to reach a contractor's software bill. Google sets the wholesale price of the underlying model; what a construction-tech vendor charges for the finished feature is a separate decision, and there's no requirement they pass a price cut through. What did happen this week is the input cost of doing this at scale, for the model class vendors reach for by default, moved by half in one release.
What this doesn't mean
Gemini 3.7 Flash still trails Claude Sonnet 5 on some of the harder agent tasks — Google's own table shows Claude ahead on Agent's Last Exam, a benchmark for multimodal desktop and operating-system tasks (33.3% versus 26.3%). This is one model from one lab in a market where Anthropic, OpenAI, and Google are all racing on price and capability at once. The story isn't "this model is now the best choice for a construction agent." It's that the price of the category dropped again, three weeks after the last drop, which is the actual trend line worth watching.
Related: OpenAI's Jalapeño chip made the same inference-cost curve visible from the hardware side — this week's story is the same curve moving again, this time on the software-pricing side.
Forward this to whoever owns the AI-agent line item on your Procore, Autodesk, or Trimble contract — ask what model it runs on and whether this week's price move changes anything.
Construction AI Brief publishes three times a week. Subscribe at constructionaibrief.com.
- What is Gemini 3.7 Flash?
- Gemini 3.7 Flash is a Google AI model released August 13, 2026, positioned as the company's workhorse model for coding and AI agents. It arrived just three weeks after its predecessor, Gemini 3.6 Flash, with higher coding-benchmark scores and a 50% introductory price cut.
- How much does Gemini 3.7 Flash cost?
- $0.75 per million input tokens and $3.75 per million output tokens through the end of 2026 — half of Gemini 3.6 Flash's rate. The price is scheduled to roughly double to $1.50 and $7.50 per million tokens after the introductory period ends.
- Does this affect Procore, Autodesk, or other construction software directly?
- No. Google hasn't announced a construction product built on Gemini 3.7 Flash, and no AEC vendor has said it's switching to it. The relevance is upstream: it's a price and capability drop in the general-purpose infrastructure that vendors use to build document- and workflow-automation agents, the kind already showing up in tools like Procore's Datagrid-based agents.
- What is AutomationBench and why does it matter?
- AutomationBench is one of Google's internal evaluations for enterprise workflow automation tasks. Gemini 3.7 Flash scored 30.4% on it, up from 17.0% for the prior model — the largest relative jump of any benchmark Google published for this release, ahead of gains on pure coding tests.
- Should a GC or sub act on this now?
- Not directly — there's nothing to buy here. The useful move is watching whether the construction software you already use passes this kind of cost and capability improvement through to what an AI agent feature costs or how much of your document backlog it can cover, rather than assuming the sticker price on your software subscription reflects what's happening underneath it.