Gemini's API Price Doubles on December 1. Check Your Exposure.
Google's Gemini API price list, last updated on 22 October 2026, carries two dated price changes. Gemini 4 Argon goes from $2 to $4 per million input tokens and from $10 to $20 per million output tokens on 1 December 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 1 January 2027. Every rate doubles.
Google's Vertex AI price list shows the same dates and the same standard rates, and is explicit about what is ending: the current Argon rate is promotional pricing, delivered as 50% credits back on net spend, and the Flash rates are introductory. Nobody is raising a price in the usual sense. A discount is expiring, which lands on the invoice the same way.
This post sets out what changes, who is exposed, how to find the usage before the first December invoice, and which levers are still worth pulling.
What changes, and when
Gemini Developer API, standard paid tier, US dollars per million tokens, list prices accessed 23 October 2026:
| Model and rate | Until the change | After the change | Effective |
|---|---|---|---|
| Gemini 4 Argon, input | $2 | $4 | 1 December 2026 |
| Gemini 4 Argon, output (including thinking tokens) | $10 | $20 | 1 December 2026 |
| Gemini 4 Argon, cached input | $0.10 | $0.20 | 1 December 2026 |
| Gemini 4 Argon, Priority input / output | $3.60 / $18 | $7.20 / $36 | 1 December 2026 |
| Gemini 3.8 Flash and 3.6 Flash, input | $0.75 | $1.50 | 1 January 2027 |
| Gemini 3.8 Flash and 3.6 Flash, output | $3.75 | $7.50 | 1 January 2027 |
| Gemini 3.8 Flash and 3.6 Flash, cache storage per million tokens per hour | $0.50 | $1.00 | 1 January 2027 |
Batch work stays at half the standard rate, so Argon batch input moves from $1 to $2. Gemini 3.5 Flash-Lite, at $0.30 input and $2.50 output, shows no dated change. Argon is not on the free tier. On Vertex AI, requests served outside the global endpoint cost 10% more at both the old and the new rates.
Vertex AI also lists a billing rule that is easy to miss: from 13 November 2026, a Gemini 4 completion longer than 32,000 tokens is charged for the tokens generated even if the request then fails on a client or server timeout. Long agent runs with tight timeouts can produce charges with no usable result.
What the doubling looks like on a real workload
An illustration, not a forecast: a document-processing service on Gemini 4 Argon that sends 500 million input tokens and receives 50 million output tokens a month.
| Input | Output | Monthly total | |
|---|---|---|---|
| Until 30 November 2026 | 500 × $2 = $1,000 | 50 × $10 = $500 | $1,500 |
| From 1 December 2026 | 500 × $4 = $2,000 | 50 × $20 = $1,000 | $3,000 |
Nothing about the workload changed. The same service costs $18,000 a year more. Multiply by the number of teams that picked Argon during the promotional period because it was the best model at a mid-tier price, and the variance is large enough to reach a budget review.
Who is exposed
Three groups, and only the first knows it today:
- Direct API users. Teams that call the Gemini API with their own keys, billed to a Google Cloud billing account. Their usage is in Google's billing data, but often in a project nobody in SAM or procurement has looked at.
- Vertex AI users. Teams that run Gemini through Google Cloud's managed platform, usually inside a wider cloud commitment. The increase draws down that commitment faster, which can bring a renewal conversation forward.
- Customers of SaaS products built on Gemini. Many applications embed a model and bill their customers through credits, AI add-ons or usage tiers. Whether the December change reaches you depends on the contract: some vendors absorb model costs, some reprice their credits, some pass them through. Ask the vendor which model powers each AI feature and whether your price for it is fixed for the term.
How to find the usage before December
Start with the billing data, not a survey. In Google Cloud, Cloud Billing reports break cost down by project and SKU, so every project with Gemini API or Vertex AI charges is visible to anyone with billing access. List the projects, the models they call, the monthly token volumes and the owner of each.
Then follow the money that does not go through Google Cloud. Search expense reports and card statements for AI services, and ask the owners of your largest SaaS applications which of their AI features run on Gemini. For each one, record the contract term and any clause that lets the vendor change the price of AI usage mid-term.
The output is a short register: who uses Argon or the two Flash models, through which channel, at what volume, and who can change it.
Levers that still work
Batch what can wait. Batch processing is half price before and after the change. Any job that does not need an answer in seconds, such as classification, extraction, overnight summaries or evaluation runs, belongs there.
Cache the repeated part. Long system prompts and reference documents that every request repeats cost a twentieth of the input rate on Argon when served from the cache, $0.20 per million from December. Argon supports implicit caching only, so the gain depends on how prompts are structured, which is an engineering change rather than a setting.
Route by task. Not every request needs the flagship. Gemini 3.5 Flash-Lite has no dated change, and the Flash models stay at the introductory rate until the end of December. A routing rule that sends simple requests to a smaller model is often the largest single saving.
Cap the output. Output costs five times input, and thinking tokens are billed as output. Setting sensible output and reasoning limits per use case reduces cost without changing the model.
Negotiate the commitment. If Gemini spend runs through a Google Cloud commitment, the December change is a reason to revisit its size and timing before it is consumed early, not after.
Why "AI prices only fall" is the wrong assumption
The average price paid per million tokens has been falling. OpenRouter's traffic, described by Peter Walker in OnlyCFO, shows it, though that is one router's view rather than the whole market. The averages fall because buyers move to cheaper models and because new models launch at lower prices, as Anthropic's Opus 5.5 did at $4 per million input tokens against $5 for Opus 4.8 (list prices, accessed 23 October 2026).
The price of a specific model you depend on can still go up. Introductory and promotional rates end. A newer model can count more tokens for the same text: Anthropic states that its tokenizer in Claude 4.7 and later produces about 30% more. A budget built on the assumption that the unit price only falls has no line for the month it doubles. AI Vendor Lock-In Turns Usage Into a License Metric covers why the cost of moving away from a model grows with every workload built on it.
The question this leaves open
The Gemini change is dated and public, and still most organizations will learn about it from the December invoice. The reason is not the price. It is that nobody has one list of where Gemini is used: in which projects, through which keys, inside which SaaS contracts. Where do the seats, the meters and the keys sit side by side, so that a vendor's price change becomes a list of owners rather than a surprise?
Software Inventory Manager (SIM) inventories your software. For the wider price picture, read Four Vendors, Four Meters: AI Pricing Compared and If You Cannot Read the AI Bill, You Have No Position.
