Gemini 3.7 Flash's Price Doubles in 5 Months. Plan Now

On August 13, Google shipped Gemini 3.7 Flash and called it its most intelligent workhorse model yet for coding and agents. The benchmark gains are real. The price is $0.75 per million input tokens and $3.75 per million output tokens, half what Gemini 3.6 Flash cost at launch.

That price expires on December 31, 2026.

On January 1, 2027, Gemini 3.7 Flash bills at $1.50 input and $7.50 output per million tokens. Exactly double. Same model, same weights, same API call, twice the invoice. Google published this in the launch announcement. Most of the coverage led with FrontierCode scores instead.

You have 138 days to build a budget that survives the rollover.

Quick Verdict

Question	The Answer
What launched?	Gemini 3.7 Flash, August 13, 2026. Google’s workhorse model for coding and agent workflows.
Introductory price	$0.75 input / $3.75 output per million tokens.
How long does it last?	Through December 31, 2026.
Standard price	$1.50 input / $7.50 output per million tokens, starting January 1, 2027.
Size of the increase	100%. Exactly double on both input and output.
Blended cost at 80/20 input-output	$1.35 per million now. $2.70 per million in January.
Is the model actually better?	Yes. 43.6% vs 34.4% on FrontierCode 1.1 Main. 65.3% vs 49.0% on DeepSWE v1.1.
Are those numbers independent?	No. Google’s own published benchmarks, not yet third-party verified.
Still cheap after the increase?	Yes. Standard rate still undercuts Claude Sonnet 5 and GPT-5.6 Terra.
Who gets hurt?	Teams that scale agent volume now and forecast Q1 off the August rate.
Does any code change trigger the increase?	No. The bill moves on a calendar date.
What to do this week	Build the Q1 forecast at $1.50/$7.50 before you ship anything on it.

What Google Actually Improved

Credit where it’s earned. This is a real capability jump, not a rebadge.

Google’s published numbers show 43.6% on FrontierCode 1.1 Main versus 34.4% for Gemini 3.6 Flash, and 65.3% on DeepSWE v1.1 versus 49.0%. The agentic benchmarks moved harder: AutomationBench went from 17.0% to 30.4%, and document comprehension on GDP.pdf climbed from 22.0% to 34.0%. Per SiliconANGLE’s launch coverage, the model ships across the Gemini API, Google AI Studio, Android Studio, Google Antigravity, and the Gemini Enterprise Agent Platform.

Two caveats before you treat those as decision-grade. These are Google’s own benchmarks published at launch, and the buildfastwithai review is right to call them directional until independent testing lands. This is also a value-tier model. It is not competing for the hardest reasoning work, and Google is not claiming it does.

For the workload most small businesses actually run, meaning multi-step agents, code generation, document extraction, ticket triage, that capability tier is fine. Better than fine at 75 cents per million input tokens.

Which is exactly the problem.

What is an AI pricing cliff?

An AI pricing cliff is a scheduled increase to a model’s per-token rate that takes effect on a fixed calendar date, with no change to the model, the API, or your code. Introductory pricing creates the cliff by setting a rate the vendor never intended to keep. Your bill changes because the promotional window closed, not because your usage moved.

That distinction matters more than it sounds. Every cost-control instinct your finance team has is built around usage. Cliffs are invisible to usage monitoring, and they land in full on the first invoice after the date.

The Math Nobody Ran

Here’s the version that shows up in a Q1 board deck.

Say your team runs an agentic coding pipeline burning 400 million input tokens and 80 million output tokens a month. That’s a modest number for a team of six engineers running agents through a normal sprint cycle.

Line item	Now (through Dec 31)	Jan 1, 2027
400M input tokens	$300	$600
80M output tokens	$300	$600
Monthly total	$600	$1,200
Annualized	$7,200	$14,400

Doubling a $600 line item is not a crisis. Nobody escalates over $600.

Now add the second variable, which is the one that actually bites. Agent workloads do not hold flat. Rippling’s internal token spend was climbing 80% month over month before it audited, a number I broke down in Rippling’s $50K-a-Month Engineer Is Your Warning. Take a much tamer 20% monthly growth from August through December. That’s roughly 2.5x the volume by January.

2.5x the volume times 2x the rate is five times the August bill. Your $600 becomes $3,000. Nobody changed a line of code.

That’s the shape of the problem. The price increase is knowable and fixed. The usage curve is the multiplier, and teams that adopt a cheap model tend to route more work to it precisely because it’s cheap. Google knows this. It’s the entire reason the introductory rate exists.

Even Doubled, It’s Still the Cheapest Option

This is the part that makes the cliff easy to shrug off, and the reason I want the numbers on the page.

Model	Input / 1M	Output / 1M	Blended at 80/20
Gemini 3.7 Flash (intro, through Dec 31)	$0.75	$3.75	$1.35
Gemini 3.7 Flash (standard, Jan 1)	$1.50	$7.50	$2.70
GPT-5.6 Terra	$2.00	$12.00	$4.00
Claude Sonnet 5 (permanent, as of Aug 10)	$2.00	$10.00	$3.60

Competitor rates per MarkTechPost’s launch comparison and, for Claude Sonnet 5’s now-permanent rate, Anthropic’s official pricing page.

At the standard January rate, Gemini 3.7 Flash blends to $2.70, about 25% below Claude Sonnet 5’s rate and roughly a third below GPT-5.6 Terra. The model stays the value play after the increase. Google is not pulling a bait and switch on relative position.

But look at the third row again. Claude Sonnet 5’s introductory rate was scheduled to expire August 31, 2026, with a 50% hike to $3/$15 set for September 1. Anthropic canceled that increase on August 10, five days before this post published, and made the $2/$10 rate permanent instead.

Right now, Google is the only one of the three major vendors selling its mid-tier coding model at a promotional price with a published expiration date. Anthropic had a cliff lined up and chose to walk away from it instead. Worth noting: the vendor with the cheaper long-term story isn’t always the one with the lower headline rate today.

Why Vendors Do This

The switching cost on an agent pipeline is not the model. It’s everything wrapped around the model.

Once you’ve tuned prompts against a specific model’s instruction-following, built tool definitions its function calling handles well, set retry and timeout budgets against its latency profile, and shipped evals that pass at its capability level, moving to a different model means redoing that work. Google is buying that migration window with a 50% discount. Anthropic tried the same play on a shorter clock, then canceled the hike on August 10 rather than find out whether switching costs would hold at the higher rate.

The bet is straightforward and mostly correct: teams that build on the cheap rate in Q4 will not rip out working pipelines in Q1 over a price change, because the engineering cost of switching exceeds the delta. Especially when the doubled price is still the cheapest option on the table.

I watched the same play run at GitHub. Copilot’s promotional credit allowance gave Enterprise seats 7,000 credits instead of the standard 3,900, then cliffed on September 1, which I covered in GitHub Copilot’s New Bill Will Shock Your Dev Team. Same structure. Generous window, hard date, no code change required to trigger it.

And the direction of travel isn’t always down. OpenAI doubled GPT-5.5’s list price at launch while its own serving costs fell 35x. Cheaper compute does not automatically become cheaper tokens.

What should you do before January 1, 2027?

Six steps, in order, and five of them are calendar work rather than engineering work.

1. Forecast Q1 at $1.50/$7.50 today. Not the introductory rate with an asterisk. Put the standard number in the model that goes to finance, and mark December 31 on the budget calendar.
2. Instrument token volume per workflow now. You need an August baseline to project against. If you cannot break spend down by pipeline and by user, you cannot tell a price problem from a volume problem in January.
3. Set a spend cap that assumes the doubled rate. Configure the ceiling in dollars against the January price, so the alert fires before the invoice does rather than after.
4. Keep the model behind an abstraction layer. One config value, not a hardcoded model string scattered across your codebase. This is the model-agnostic workflow argument applied to a specific date.
5. Run a routing audit in October. Find every call going to Flash that a lighter model handles fine. Do this while the cheap rate is masking the waste, because the waste doubles too.
6. Ask Google Cloud for committed-use pricing in November. If your volume is real, the standard rate is a starting position. Vendors negotiate hardest against a competitor’s number, and you’ll have two of them from the table above.

Steps 1 through 3 take a finance analyst about two days. Do them before you ship anything else on this model.

The Anti-Hype Read

Three cautions before this becomes a routing decision.

The benchmarks are vendor-published. A 16-point jump on DeepSWE is a large claim. Google’s models have generally delivered on the Flash tier, and the pattern from Gemini 3.5 Flash’s launch in May held up in practice. Run your own evals on your own workload anyway. Benchmark gains on FrontierCode do not predict how a model handles your tool schemas.

“Still cheapest” is a moving claim. Anthropic canceled its own pricing cliff on August 10, five days before this posted. Google’s cliff still kicks in January 1. OpenAI cut Terra 20% in July. The relative ordering in that table has a shelf life measured in weeks, which is the actual argument for keeping your model selection in configuration rather than in code.

Cheap tokens change behavior, and the behavior outlasts the price. Teams route more work to a model when it’s cheap: longer context windows, more retries, more speculative agent runs. Those habits calcify into architecture. When the price doubles, you’re not paying double for the same usage pattern. You’re paying double for the more expensive usage pattern the cheap price encouraged.

My Read

Google priced this one honestly. The expiration date is in the launch announcement, the standard rate is published, and there’s no fine print games. That’s better disclosure than most of this market offers.

The failure here will be entirely on the buyer side, and it will be a forecasting failure rather than a procurement one. Somebody builds an agent pipeline in September, watches it run at $600 a month through Q4, and puts $7,200 in the 2027 budget. In February they’re explaining a number three to five times that, and the honest answer is that nobody read the pricing page past the headline.

The broader pattern is the thing to internalize. Introductory pricing with a published cliff is now standard practice at the mid-tier, even when a vendor blinks and cancels the hike, the way Anthropic did on August 10. That means every token price in your financial model needs an expiration date attached to it, and your renewal calendar needs to track model pricing the way it tracks SaaS contracts. I made a version of this argument during the June price war, when the risk was locking into volume commitments ahead of cuts. The risk has inverted. Now it’s forecasting off a rate the vendor already told you it’s going to change.

The teams that come out of Q1 clean are the ones who wrote $1.50/$7.50 into the spreadsheet in August and then enjoyed the discount as upside. The teams that get hurt are the ones who treat 138 days as a long time.

Your Next Step: Open your 2027 AI budget line today and reprice every Gemini Flash workload at $1.50 input and $7.50 output per million tokens. Then pull your August token volume by workflow, apply your actual growth rate through December, and multiply. If that number would require an approval conversation, have it now while it’s a forecast instead of in February when it’s an invoice. Then set one calendar reminder for December 1 to run the routing audit, because the cheapest thing you can do about a price increase is stop paying for calls you never needed.

---

Related Reading:

• Rippling’s $50K-a-Month Engineer Is Your Warning
• GitHub Copilot’s New Bill Will Shock Your Dev Team
• Google Cut the Price. Now Rebuild Your AI Stack.
• OpenAI Doubles the Price, Cuts Own Costs 35x
• The AI Price War Starts Now. Don’t Lock In Wrong.
• Your AI Stack Has an Expiration Date
• What Running AI Agents Actually Costs in 2026