Google shipped Gemini 3.7 Flash on August 13, and the headline number everyone repeated was the price: $0.75 per million input tokens, $3.75 per million output. That's cheap for a model Google calls its most intelligent workhorse for coding and agents.
Read the pricing page footnote and the number has an expiry date. On January 1, 2027, input goes to $1.50 and output goes to $7.50. Both exactly double. You have 140 days.
Key Takeaways
Gemini 3.7 Flash costs $0.75 in and $3.75 out per million tokens through December 31, 2026. On January 1, 2027 it becomes $1.50 and $7.50, a clean 2x on both.
Context caching doubles too, from $0.075 to $0.15, and cache storage goes from $0.50 to $1.00 per million tokens per hour. That's the lever most teams use to control agent costs, so it hits twice.
Output pricing includes thinking tokens. A reasoning-heavy agent pays the doubled rate on tokens the user never sees.
The same day, OpenAI previewed Ultrafast: GPT-5.6 Sol at up to 750 output tokens per second, which it calls 14x standard speed. I could not find a price for it in any of the coverage.
Anthropic's comparable feature, fast mode, is 2.5x and its own docs say it runs "at premium pricing." Speed is a paid tier now at all three labs.
If you re-platform onto 3.7 Flash during the discount, budget for the January number, not the August one.
Google's own announcement calls 3.7 Flash a workhorse for coding and agents. The word workhorse is doing a lot of work: this is the model they expect you to run in a loop, all day.
What did Google actually ship on August 13?
Gemini 3.7 Flash, a replacement for 3.6 Flash, which shipped three weeks earlier. Google frames it as the product of core optimizations and developer feedback, with better coding and agentic performance. It's live in the Gemini API, AI Studio, Android Studio, Antigravity, and the Gemini Enterprise Agent Platform.
What it isn't: Gemini 3.5 Pro. Google promised that flagship at I/O in May, said June, and June came and went. Ars Technica's Ryan Whitwam made the point sharply, that a Flash release three weeks after the last Flash release may be more about maintaining the appearance of constant improvement than about the improvement itself.
There's a tell that supports him. The regular Gemini chatbot still runs 3.6 Flash. Individual access to 3.7 goes through Spark, the agent feature, and only if you pay for AI Pro or Ultra. Google shipped its newest workhorse to developers and to its own agent product, and left the consumer chat surface alone. That's not the rollout of a model you're confident about across the board. That's the rollout of a model you want in agent loops, where the token volume is.
Why does Gemini 3.7 Flash pricing double on January 1, 2027?
Because $0.75 and $3.75 are introductory rates with a hard stop. Google's own developer pricing page spells it out per line item: input is "$0.75 through December 31, 2026. $1.50 starting January 1, 2027." Output reads the same way, $3.75 becoming $7.50. Nothing about the model changes on that date. Only the invoice does.
The doubling is written into every row, not just the headline two. Note the output line: it says "including thinking tokens", which is where reasoning-heavy agents get expensive.
I read the headline rate first and almost filed this one as good news. I was wrong, and the footnote is the whole story.
Here's the part I nearly missed, and it's the one that actually stings. Context caching doubles as well, from $0.075 to $0.15 per million tokens, and cache storage goes from $0.50 to $1.00 per million tokens per hour.
Caching is the first thing I reach for when an agent's bill gets ugly. Long system prompt, big tool schema, a retrieved document set that barely changes between turns: cache it once, pay a fraction on every subsequent call. On the chatty agents I've deployed it's the single biggest lever there is, worth more than model choice. So when the base rate doubles and the cost-control lever doubles alongside it, a well-optimized agent gets no protection from the increase. The teams who did the careful engineering land in exactly the same place as the teams who didn't.
One more line worth reading before you commit. Grounding with Google Search gives you 5,000 free search requests a month, shared across all Gemini 3.x models, and then it's $14 per 1,000 requests. Shared across all models is the phrase to notice. If you're running three Gemini-backed agents, they're all drinking from the same 5,000.
What does the increase cost a real agent workload?
Depends entirely on your output-to-input ratio, and output is where the pain lives at a 5:1 price difference. Below are three workloads I keep seeing, priced at both rates. These are monthly token volumes, not requests.
| Monthly agent workload | Input / output tokens | Bill today | Bill Jan 1 | Increase |
|---|---|---|---|---|
| Small: support triage agent | 20M / 4M | $30.00 | $60.00 | +$30.00 |
| Mid: document pipeline | 200M / 40M | $300.00 | $600.00 | +$300.00 |
| Large: coding agent fleet | 800M / 250M | $1,537.50 | $3,075.00 | +$1,537.50 |
The support agent going from $30 to $60 is noise. Nobody reforecasts over thirty dollars. The coding fleet going from $1,537.50 to $3,075 is $18,450 a year that appeared on your P&L without a meeting, on a date you didn't pick.
And that middle row is the one that catches people, because $300 a month is precisely the size of bill nobody watches. It's under the threshold where finance asks questions and over the threshold where it matters by year end. I wrote about a version of this failure when the Army promised its people unlimited AI tokens and ran dry in weeks, and again when Rippling found itself about to spend 40% of its engineering payroll on tokens. The pattern is always the same. Token spend doesn't spike. It creeps, and then a rate change lands on top of the creep.
Is 3.7 Flash actually better than 3.6 Flash?
On Google's own benchmarks, yes, and by more than a rounding error. The agentic and document numbers moved the most, which is consistent with a model tuned for tool loops rather than for chat.
| Benchmark | 3.6 Flash | 3.7 Flash | Change |
|---|---|---|---|
| FrontierCode 1.1 Main | 34.4 | 43.6 | +9.2 pts (+26.74% relative) |
| DeepSWE v1.1 | 49 | 65.3 | +16.3 pts (+33.27% relative) |
| WebDev Arena (Elo) | 1538 | 1588 | +50 Elo |
| GDP.pdf | 22 | 34 | +12 pts (+54.55% relative) |
| AutomationBench | 17 | 30.4 | +13.4 pts (+78.82% relative) |
AutomationBench measures how well a model executes common business workflows, and it went from 17 to 30.4. That's the biggest relative jump on the board and it's the benchmark closest to what most of my clients actually pay a model to do. But read the absolute number, not the delta. Thirty percent. A model that completes fewer than a third of standard business workflows end to end is not something you point at a customer-facing process without a human in the loop and a rollback path.
That's my honest read of the whole table. These are real gains on a weak base. Vendors quote you the relative jump because +78.82% sounds like a different category of product. The absolute score tells you what you're actually deploying.
Ars filed this under the kicker "Inching closer to 4.0". Shipping Flash point releases three weeks apart while the 3.5 Pro promised for June still hasn't arrived is a version-number strategy, not a capability strategy.
What is OpenAI's Ultrafast mode, and what does it cost?
Ultrafast is a preview mode that runs GPT-5.6 Sol at up to 750 output tokens per second, which OpenAI describes as 14x standard speed. The company pitched it at enterprise workflows, naming incident response specifically. On price, I came up empty. TechCrunch's write up doesn't carry one, and I couldn't find a rate anywhere else in the coverage.
Do the division on their own claim and you get the baseline they're comparing against: 750 divided by 14 is about 53.6 tokens per second. Take a single agent step that emits 2,000 tokens. At 53.6 per second that's 37.3 seconds of wall clock. At 750 it's 2.7. Across a twelve-step agent loop, you save roughly 416 seconds, near enough seven minutes per run.
Seven minutes is not a nice-to-have if a human is sitting there during an incident. That's the difference between an agent that participates in the response and one that files a report after the fact. OpenAI's own framing was that until now, real-time speed meant picking a smaller or more specialized model, and that Ultrafast points toward more useful work per second. On the engineering claim, I think they're right, and it's the more interesting of the two announcements.
On the commercial claim, nobody should assume it's free. Anthropic says the quiet part out loud in its docs.
Anthropic's fast mode is 2.5x, not 14x, and the docs say plainly that it bills at premium pricing. It's also gated behind a waitlist and unavailable on Bedrock, Google Cloud, or Microsoft Foundry.
That 2.5x versus 14x gap is worth sitting with, and TechCrunch answers it: Ultrafast runs on OpenAI's partnership with the chipmaker Cerebras, and the preview is going to a small group of customers, with access expanding as capacity grows. So the gap is silicon, not a measurement trick. It's also borrowed silicon on a limited allocation, which makes the missing price tag more interesting rather than less. Anthropic's docs are specific enough to check: fast mode covers Claude Opus 5 and Opus 4.8, you opt in with speed: "fast" and a beta header, and if you send that flag to Opus 4.6 it quietly runs at standard speed and bills at standard rates. That's the level of detail I want before I put a speed tier in a production path. Ultrafast doesn't have it yet, which is fine for a preview and not fine for a commitment. My rule after OpenAI shipped a voice engineering post with zero latency numbers is simple: an unpriced, unbenchmarked speed claim is marketing until the docs catch up.
The 140 days you have, and what I'd do with them
Between now and January 1 there are four things worth doing, and none of them take a week.
Price your current workload at the January rate today. Not the August rate. Pull last month's token counts, double them in a spreadsheet, and look at the number. If that number would require a conversation with someone, have the conversation now, while it's a forecast instead of an invoice.
Pin your model version and instrument the swap. Three Flash releases in a quarter means whatever you pin is stale in weeks, and whatever you don't pin drifts under you in production. I pin, and I keep a small eval set of maybe forty real requests from the actual workload so that upgrading is an afternoon of measurement rather than a leap. Google moving its consumer chat surface and its agent surface onto different Flash versions on the same day should tell you how stable they consider this.
Separate the speed decision from the model decision. Ultrafast, fast mode, and Flash are three answers to three different questions. Flash is cheap capability for high-volume loops. Fast mode and Ultrafast are latency purchases for the narrow set of paths where a human is waiting. Most agent steps have no human waiting and should never touch a premium speed tier. In my experience the flag ends up set globally in a config file nobody re-reads after launch, and then a nightly batch job quietly pays a latency premium for months.
Don't re-platform for the discount. This is the one I'd argue about. A 140-day introductory rate is not a reason to move a working agent onto a new model, because the migration cost, the eval rebuild, and the prompt retuning will outlast the discount. Move for the AutomationBench and DeepSWE gains if those map to your workload. Move for the price only if you've priced it at $1.50 and $7.50 and it still wins.
If you're weighing this across providers rather than inside Google's lineup, I've written up how I actually choose between OpenAI and Claude after shipping 126 systems on both, and there's a longer piece on custom code versus frameworks versus no-code if you're earlier than that. The pricing logic here is the same one I worked through when Claude Opus 5 halved frontier pricing and most agent bills didn't move: headline rates change more often than your actual unit economics do, and the gap between those two facts is where budgets die.
Not sure which of your processes should be running on an agent in the first place? The AI readiness assessment takes a few minutes and tells you where the volume actually is.
Frequently asked questions
When exactly does Gemini 3.7 Flash pricing increase?
January 1, 2027. The introductory rate of $0.75 per million input tokens and $3.75 per million output tokens runs through December 31, 2026. From January 1 the rates are $1.50 and $7.50, published in Google's own developer pricing table.
Does the price increase apply to context caching too?
Yes. Context caching goes from $0.075 to $0.15 per million tokens, and cache storage goes from $0.50 to $1.00 per million tokens per hour. Both double on the same date, so caching your way out of the increase doesn't work.
Are thinking tokens billed at the output rate?
They are. Google's pricing page labels the row "Output price (including thinking tokens)". For reasoning-heavy agents this matters more than the headline rate, because a large share of what you pay for at $3.75, soon $7.50, is never shown to the user.
Is Gemini 3.7 Flash worth upgrading to from 3.6?
If your workload is coding or multi-step tool use, probably yes. DeepSWE v1.1 went from 49 to 65.3 and AutomationBench from 17 to 30.4. If your workload is chat or simple extraction, the gains are smaller and 3.6 Flash still runs the consumer Gemini app, so you're not on an abandoned model.
How much faster is OpenAI's Ultrafast mode in practice?
OpenAI claims up to 750 output tokens per second, which it calls 14x standard. On a 2,000-token agent step that's roughly 2.7 seconds instead of about 37. Over a twelve-step loop it saves close to seven minutes. No price appears anywhere in the coverage, and the preview is limited to a small group of customers at launch.
How does Anthropic's fast mode compare to Ultrafast?
Anthropic's fast mode delivers up to 2.5x higher output tokens per second on Claude Opus 5 and Opus 4.8, and the docs state it bills at premium pricing. It's a research preview behind a waitlist and isn't available on Amazon Bedrock, Google Cloud, or Microsoft Foundry.
Should I switch my agents to Gemini 3.7 Flash to save money before January?
Only if it wins at the post-January rate. Price your workload at $1.50 and $7.50 first. Migration, eval rebuilds, and prompt retuning typically cost more than 140 days of a 50% discount returns.
What happens to the free tier?
Gemini 3.7 Flash input, output, and context caching are all listed as free of charge on the free tier, with limited access and lower rate limits. The doubling applies to the paid tier only. Note that content on the free tier is used to improve Google's products, which rules it out for most production work.
Citation Capsule: Gemini 3.7 Flash introductory pricing of $0.75/$3.75 per million tokens rising to $1.50/$7.50 on January 1, 2027, plus context caching at $0.075 rising to $0.15, per Google's developer pricing table. Benchmark gains (FrontierCode 1.1 Main 34.4 to 43.6, DeepSWE v1.1 49.0 to 65.3, AutomationBench 17 to 30.4) per Google. OpenAI Ultrafast at up to 750 output tokens per second and 14x standard speed per TechCrunch. Anthropic fast mode at up to 2.5x at premium pricing per Anthropic's platform docs. Gemini Developer API Pricing (Aug 2026) · Google, Introducing Gemini 3.7 Flash (Aug 13, 2026) · Ars Technica (Aug 13, 2026) · TechCrunch, OpenAI Ultrafast (Aug 13, 2026) · Anthropic, Fast mode docs.
Top comments (0)