Google's standard rate for Gemini 3.7 Flash is $1.50 and $7.50 per million input and output tokens — which is exactly what Gemini 3.6 Flash already listed at. The 50% headline is a discount that expires on 31 December 2026, so what you are actually being offered is a better model at last generation's price, plus a dated 2× step in your unit costs that lands on whichever trajectories you tuned while tokens were cheap.
What actually shipped
Released 13 August 2026, about three weeks after 3.6 Flash, and positioned squarely at coding and agentic workflows rather than at chat.
| Attribute | Gemini 3.7 Flash |
|---|---|
| Input price, to 31 Dec 2026 | $0.75 per million tokens |
| Output price, to 31 Dec 2026 | $3.75 per million tokens |
| Standard rate, from 1 Jan 2027 | $1.50 input / $7.50 output |
| Context and output ceiling | 1M-token context, up to 64K output tokens |
| Modalities and control | Text, image, audio, video; configurable thinking budgets |
Read the chart as a product decision rather than a price cut and it becomes clear what Google did: it held the list price flat across a generation and passed the improvement through as capability, then attached a four-and-a-half-month promotion to drive adoption during the window when switching costs are lowest. That is a completely ordinary commercial move. It is also one that interacts badly with how agent workloads accumulate cost.
Why an agent workload feels this differently from a chatbot
For an assistant answering one-shot questions, a doubling on 1 January is an annoyance in a spreadsheet. Two properties of agent workloads turn the same event into something you have to plan for now.
Agent spend is superlinear in the thing the discount encourages
Models are stateless, so every step of a loop re-sends the transcript so far. Input tokens per step grow with the step count, which means total input across a task grows with roughly the square of it — the arithmetic in agent cost control. A discount does not change that curve; it just moves you down it temporarily. What matters is where on the curve you settle while the discount is on, because that is the configuration that gets repriced.
The configuration is sticky in a way the price is not
Thinking budgets, step ceilings, retry counts, how much context you carry, whether you route the easy steps elsewhere — these get tuned once, encoded into prompts and evals, and then defended. Prices change by announcement; harness configurations change by project. Four and a half months is comfortably long enough for a cheap-window configuration to become the thing your quality numbers assume.
Put the two together and the failure mode is specific rather than vague. A team adopts 3.7 Flash in September, finds that raising the thinking budget closes the remaining quality gap on their hardest tasks, ships that, and books the savings against the introductory rate. In January the rate doubles on a token volume that also grew — because thinking tokens bill as output, at the higher of the two rates. The bill does not go up 2×.
The benchmark numbers carry a harness with them
The published results deserve the same scepticism as the price. Google's own evaluation material records the coding results as self-computed using a mini SWE-agent harness with high thinking — meaning the scaffold and the effort setting are part of the measurement. That is normal and unavoidable for agentic evaluation, but it has a direct cost consequence: the configuration that produced the headline number is the expensive configuration, because high thinking means more output tokens, billed at $3.75 today and $7.50 in January.
The spread between suites makes the same point from the other direction. The release reports roughly 85.8% on Terminal-Bench 2.1 and about 14.9% on Terminal-Bench 3.0 — same weights, same week, different suite version. Neither number is wrong; they are answers to different questions, and quoting either one without its version and scaffold is close to meaningless. As of writing there is no independent SWE-bench Verified comparison against rival models, so the honest position is that the model looks strong on the axes Google chose to publish and nobody outside has confirmed the ordering yet. The general form of this problem is the agent harness; we made the same argument about a different vendor in DeepSeek's harness scores include the scaffold.
What to actually do before January
Budget at the standard rate from day one
Run every model-selection and unit-economics calculation at $1.50 / $7.50 and treat the discount as margin rather than headroom. This is one line in a spreadsheet and it removes the entire failure mode. If the workload only pencils out at the introductory rate, you have not found a cheap model — you have found a deadline.
Measure cost per completed task, not per token
Token price tells you nothing about whether the work got done, and a model that finishes in eight steps instead of twenty-five is frequently cheaper at a higher list price. Track spend divided by tasks that passed your eval, watch the p95 and p99 for the runs that cost 100×, and you will be able to answer the January question — is this still the right model — with a number instead of an argument.
Make the thinking budget an explicit, reviewable setting
If effort is configurable, it is a cost dial that quietly moved while nobody was watching. Record it per run, chart output tokens per task over time, and set a threshold that triggers a review. The teams that get hurt in January will mostly be the ones who cannot say what their effort setting was in September.
Keep a second model qualified
Not because you expect to switch, but because a dated 2× step is exactly the event that makes switching worth costing out, and a migration you have never rehearsed takes a quarter you will not have. Keep one alternative running against your eval set on a cadence — the warm-candidate discipline in model deprecation and migration applies unchanged to a repricing.
The general principle survives this particular model: an introductory rate is a term sheet, not a price. When a vendor publishes a discount with an end date, the number that belongs in your planning is the one after the date, and the interesting question is not how much you save between now and then — it is what you will have built that assumes the cheap number.
FAQ
Might Google just extend the introductory pricing?
It might, and promotional windows are extended often enough that it is a reasonable hope. It is not a plan. Extension is a decision made by someone whose incentives you do not control, announced on a schedule you do not know, and the cost of assuming it wrongly is a doubled bill on a workload you have already committed to.
Is Gemini 3.7 Flash actually cheap compared with the alternatives?
At the introductory rate it is priced well below the frontier tier, and even at the standard rate it sits in the workhorse band rather than the premium one. The question that matters is not the per-token comparison but cost per completed task on your workload, which depends on how many steps each model needs and how often it has to retry. That ordering is measurable and it is frequently not what the price list suggests.
Does the 1M-token context change the cost picture?
Only if you use it, and using it is usually the expensive mistake rather than the feature. A large context ceiling is permission to carry more, and everything you carry is re-sent and re-billed on every subsequent step. Treat the window as a limit you are staying well inside, not a budget to spend.
Should we wait for independent benchmarks before adopting?
Waiting for third-party numbers is reasonable, but the more useful move is to stop treating anyone else's benchmark as decisive and run your own eval set against your own harness. That takes days rather than months, answers the question you actually have, and gives you the baseline you will need in January anyway.
What is the single number to write down today?
Cost per completed task at $1.50 / $7.50 with your current effort setting. If that number is acceptable, the January step is a non-event. If it is not, you have four months to fix it, which is a great deal more time than the teams who only look at the invoice.
Further reading
On this wiki:
- Agent cost control — why agent spend grows with roughly the square of the step count.
- The agent harness — why a published agentic score is a model-plus-scaffold measurement.
- Adaptive thinking & effort budgets — how to set the dial this post says to watch.
- Unit economics — the model for pricing a workload rather than a token.
- Model deprecation & migration — keeping a second candidate warm.