Gemini API Price Increase: What Doubling Token Prices Does to Your AI COGS and Gross Margin

Google's Gemini API pricing page lists two dated price doubles. Gemini 4 Argon goes from $2 to $4 per million input tokens and from $10 to $20 per million output tokens on December 1, 2026. Gemini 3.8 Flash and Gemini 3.6 Flash go from $0.75 to $1.50 input and from $3.75 to $7.50 output on January 1, 2027. Cached input, Batch, Flex and Priority prices double on the same dates. For a SaaS product that runs its AI feature on these models, AI cost of goods sold doubles too, and gross margin falls by the AI cost as a share of revenue.

This guide covers the exact prices, what the change does to AI COGS and gross margin in a worked example, and the options a founder has before the dates hit. For the general model of AI cost and margin, read what AI COGS does to SaaS gross margin, which also keeps a dated price history across vendors.

What changes, and when

All figures below are from Google's Gemini Developer API pricing page, paid tier, US dollars per million tokens, read on October 25, 2026 (the page says it was last updated October 22, 2026). "Output" includes thinking tokens, which Google bills as output.

Four details matter for a forecast:

• Every line doubles, not just the headline rate. Cached input, Batch, Flex, Priority and cache storage all move by the same factor. So whatever mix of standard, cached and batch traffic you run today, your bill on these models doubles. Savings you already made do not offset the change. Only new savings do.
• Argon has one price band. Unlike Gemini 3.1 Pro Preview, which charges more for prompts over 200,000 tokens, the Argon table shows no context-length tier. Argon also supports implicit caching only; the pricing page says explicit caching is not supported.
• Older model strings already point at the new models. Google's release notes say that since October 8, 2026, requests to gemini-3.7-flash are routed to gemini-3.8-flash and requests to gemini-3.5-flash are routed to gemini-3.6-flash. If your code still names one of those, you are on a model that doubles on January 1.
• Some models carry no dated change. Gemini 3.5 Flash-Lite ($0.30 / $2.50), Gemini 3.1 Flash-Lite ($0.25 / $1.50 for text) and Gemini 3.1 Pro Preview ($2 / $12 up to 200,000 tokens) show a single price with no date on the same page.

Why the price is rising

Google does not frame this as a price increase. It frames the current price as an introduction. The Gemini 4 Argon guide says the model is "available through Nov 30, 2026 at an introductory price of $2/1M input tokens and $10/1M output tokens ($4/1M input and $20/1M output thereafter)." The release notes used the same wording when Gemini 3.7 Flash launched on August 13, 2026: "available at an introductory price through December 31, 2026." We found no Google statement giving any other reason, so we do not guess at one.

The finance lesson is general. Introductory and promotional prices are now common. OpenAI's pricing page notes that one model's "promotional pricing is available at least through November 21, 2026." If your forecast uses a price with an end date, model the price after that date, not the one on the invoice today.

How a price double flows into gross margin

Inference is cost of revenue, on the same line as hosting and support (see what counts as COGS). The chain is short:

• Cost per request = uncached input tokens x input price + cached input tokens x cached price + output tokens x output price, all divided by one million.
• AI COGS per customer = cost per request x requests per customer per month.
• Gross margin = (revenue - non-AI COGS - AI COGS) / revenue.

When the token price doubles, AI COGS per customer doubles, and gross margin falls by exactly the AI cost as a share of revenue before the change. A feature that costs 5% of revenue takes 5 points off margin. A feature that costs 15% of revenue takes 15 points. That is why the first number to know is AI COGS as a percentage of revenue, by plan.

Worked example: a $49 AI plan on Gemini

The numbers below are illustrative, not data from any real company. A SaaS company has 1,500 customers on a $49 per month plan. Each customer makes 300 AI requests a month. The product routes 80% of requests to Gemini 3.8 Flash and 20% of the hard ones to Gemini 4 Argon. Each request sends 12,000 input tokens, half of them cache hits. A Flash request returns 1,500 output tokens and an Argon request 4,000, thinking included. Non-AI COGS (hosting, support, payment fees) is 12% of revenue, so gross margin before AI is 88%.

Without any change on the company's side, gross margin drops about 11.6 points in two steps, and AI COGS rises by $8,541 a month, or about $102,500 a year at the same volume. The Argon slice is only 20% of requests but carries more than half of today's AI cost ($3.16 of $5.69 per customer), so it does most of the damage in December.

The heavy-user problem gets worse too. A customer who uses 2.5 times the average costs $14.24 a month in AI today (58.9% margin on that account) and $28.47 from January (29.9% margin). The AI COGS guide explains why the tail, not the average, is where flat pricing breaks.

The inference efficiency ratio (plan revenue divided by AI cost, a metric from Ben Murray of The SaaS CFO) falls from 8.6 to 4.3.

What it takes to hold margin

The same example, at January 2027 prices, under each option. Every line is re-derived in our notes.

No single lever gets back to 76.4%. The combination gets within two points. A price rise is the other route:

• Hold gross profit dollars: the plan has to go from $49 to $55.47 (+13.2%). Gross margin at that price is still only 67.5%, because the AI cost is now a bigger share.
• Hold gross margin percent: with non-AI COGS a fixed share of revenue, the price has to rise in step with AI cost, from $49 to $98. That is not a realistic plan, which is why cost work comes first.
• Hit a 70% target: $63.27. A 75% target needs $87.60.

A 13.2% rise breaks even on MRR if fewer than about 11.7% of affected customers cancel because of it (break-even share = rise / (1 + rise)). Our price increase calculator runs that over 12 months with grandfathering and renewal timing.

Your options before December 1

1. Measure first

Pull one month of usage by model, plan and customer: input, cached and output tokens (thinking included) and request count. Without AI COGS per customer by plan you cannot tell whether the change costs you 3 points or 15.

2. Switch model or tier, with evals

Current list prices from each vendor's own page, read October 25, 2026, per million tokens, standard tier:

After December 1, Argon's list price matches Claude Opus 5.5, and its current introductory price matches Claude Sonnet 5.5 and gpt-6.1-sol. Two cautions before you treat these as swaps. First, models are not interchangeable: run your own evals on your own prompts before moving production traffic. Second, equal list prices do not mean equal cost, because vendors count tokens differently. Anthropic's pricing page says its Claude 4.7 and later models use a tokenizer that "produces approximately 30% more tokens for the same text." In our example, moving the hard 20% to a $2 / $10 model gives 71.2% margin at equal token counts and 69.3% if that model counts 30% more tokens. Measure tokens on your own traffic.

Staying on Google is also an option. Gemini 3.5 Flash-Lite has no dated change. Going back to Gemini 2.5 models is limited: since September 18, 2026, Google restricts 2.5 access to users who have used them before.

3. Route by difficulty

Routing is usually the biggest lever, because the expensive model carries a share of cost far above its share of requests. Send simple calls (classification, extraction, short answers) to a small model and keep the large one for the cases that fail on the small one. In the example, halving Argon's share of requests saves more than any caching change.

4. Cache and batch

Cache reads cost a tenth of the input price on Gemini 3.8 Flash and a twentieth on Argon (Google's Argon guide says implicit caching "discounts cached input by 95%"). A stable system prompt and stable context at the start of each request raise the hit rate. Batch and Flex cost half the standard rate, and from the new dates the Batch price equals today's standard price on both Argon and Flash. Anything that does not need an instant answer (nightly summaries, enrichment, report drafts) belongs there. Note that cache storage on Flash is billed per hour, which our example leaves out.

5. Pass the cost through

If AI cost moves with usage, the price can too. Options include a usage component on top of the seat, prepaid credits, or pricing per completed outcome. Each moves part of the token risk to the customer. See usage-based pricing, moving from per-seat to hybrid pricing and outcome-based pricing. If you sell credits, set the credits-per-action rate so it can change, and read how prepaid credits are recognized before you change the rate on credits already sold.

6. Contracts, on both sides

With customers: an annual plan sold at a fixed price before January 1 keeps that price until renewal, while the token cost rises on the date. The margin hit lands on every month left in the term. New contracts can include usage limits, fair-use terms or a clause that ties the AI allowance to model cost. With Google: the pricing page lists volume-based discounts and provisioned throughput under the Enterprise tier, through Gemini Enterprise Agent Platform. Whether an annual commitment gets you a better rate is a sales conversation, and we have no published figure for it.

This guide is general information, not accounting, tax or legal advice. How you classify AI costs, recognize credit revenue or word a pass-through clause depends on your facts; check with your accountant or counsel.

Model the dates, not an average

A forecast that uses one blended token price will miss this. Put the price change in the month it happens: December 2026 carries Argon at the new rate for the whole month (the change starts on the first), and January 2027 adds Flash. Then run a scenario for each mitigation with its own start date, since routing and caching changes take engineering time. Carry the result through to gross margin, burn and the unit economics investors look at, including CAC payback, which uses gross margin. If you are drafting next year's operating plan now, this belongs in the COGS lines from month one.

For a quick check on one model, the AI cost calculator (linked here with Gemini 3.8 Flash's January 2027 prices) takes your token counts and plan price and shows margin before and after AI, plus the price rise or cost cut that holds a target. It also sits at the end of this page.

Frequently asked questions

How much does the Gemini API cost?

It is billed per million tokens and varies by model. As of October 25, 2026, Gemini 4 Argon costs $2 input and $10 output until November 30, then $4 and $20. Gemini 3.8 Flash and 3.6 Flash cost $0.75 and $3.75 until December 31, then $1.50 and $7.50. Gemini 3.5 Flash-Lite costs $0.30 and $2.50. Batch requests cost half.

Is the Gemini API price going up?

For some models, yes, on dates Google has already published. Argon doubles on December 1, 2026. Gemini 3.8 Flash, 3.6 Flash, the 3.8 Flash TTS models and Robotics ER 2 Preview double on January 1, 2027. Google calls the current Argon price introductory.

Does the change affect the free tier?

Argon is not available on the free tier at all. Gemini 3.8 Flash and 3.6 Flash list input and output as free of charge on the free tier, which has lower rate limits, and Google uses free-tier content to improve its products. A production SaaS feature normally runs on the paid tier.

Is Gemini cheaper than OpenAI or Claude?

It depends on the tier. From December 1, Gemini 4 Argon lists at the same $4 / $20 as Claude Opus 5.5. From January 1, Gemini 3.8 Flash at $1.50 / $7.50 costs more per token than the smallest models from Anthropic and OpenAI at $0.10 / $0.50, but those are different classes of model. Compare on your own prompts and token counts, not list price alone.

How do I lower my Gemini API costs?

In order of typical impact: route easy requests to a smaller model, raise the cache hit rate with stable prompt prefixes, move non-urgent work to Batch or Flex for half the price, trim output and thinking length, and check that legacy model names are not silently routed to a model that doubles.

Model it in Adlega

In Adlega you build a driver-based 36-month SaaS model, so AI cost can sit in COGS as requests per customer times cost per request, with the price change dated to the month it applies. You can keep one scenario at current prices and one per mitigation, and see the effect on gross margin, burn and runway next to MRR, churn and CAC. The AI CFO explains how each number was calculated if you want to check the chain from tokens to margin.