Gemini API price increase: what doubles on December 1 and January 1

The Gemini API price increase is already printed on Google's own pricing page, with dates attached: Gemini 4 Argon goes from $2.00 to $4.00 per million input tokens on December 1, 2026, and the 3.8 and 3.6 Flash line doubles on January 1, 2027. Below are the new rates model by model, what they look like next to Claude and OpenAI list prices, and where the cheaper token is now.

This matters because Argon's launch price was explicitly temporary. When Google made the model broadly available on October 13, VentureBeat reported the $2/$10 rate as limited-time introductory pricing that expires on November 30. If you benchmarked Argon in October and picked it on cost, the number you based that decision on stops existing in a few weeks.

When the Gemini API price increase takes effect

Two dates, both published as notices inside the price cells on Google's Gemini API pricing page (checked October 23, 2026). The Argon row reads "$2.00 through November 30, 2026. $4.00 starting December 1, 2026." The Flash rows carry the same construction with a January 1, 2027 switch.

All figures are per million tokens on the paid tier, and the doubling is consistent: standard, batch, cache read and cache storage all move together. Google's pricing page shows no transition tier and no grandfathering notice, so we would plan on a request sent on December 1 being billed at the new Argon rate even if your integration has been running since October.

After December 1, Argon costs what Claude Opus 5.5 costs

The clean way to read a price change is against the alternatives you could actually deploy this quarter. Here are current list prices for text models at the flagship end, each from the vendor's own pricing page as of October 23, 2026.

Anthropic's pricing page lists Opus 5.5 at $4 and $20 per million tokens, the exact pair Argon moves to, while Sonnet 5.5 sits at $2 and $10. OpenAI's API pricing puts gpt-6.1-sol at the same $2 and $10 for inputs up to 272K tokens. In other words, December 1 does not make Argon expensive in absolute terms - it moves Argon from undercutting Anthropic's top model to matching it, and to twice the list price of two mid-tier models that many teams already run.

That is the decision, and it is a quality question rather than a price question. On the independent side, VentureBeat's launch coverage notes that Artificial Analysis scored Argon 53 on its Intelligence Index against 58 for Claude Opus 5.5, while Vals AI ranked Argon first on its own index; the same piece reports Google's own comparison showing Argon behind Opus 5.5 on Terminal-Bench 4.0, 57.4% to 66.4%, and flags those Google figures as company-reported. In our view, paying Opus money for a model that trails Opus on the vendor's own coding benchmark needs a reason beyond habit - a context limit, a region, a Google Cloud commitment - and if you cannot name that reason, the honest answer is that you were buying the introductory price.

The January 1 Flash increase will cost high-volume apps more

For most products, the Flash line is the bill. A support assistant, a classifier, a summarizer, an extraction pipeline: these run on cheap models at volume, and that is exactly where a doubling compounds.

Take a workload of 50 million input and 10 million output tokens a month, which is a mid-sized production feature rather than a lab experiment. Our arithmetic on the list prices above:

• Gemini 3.8 Flash today: $75.00 a month. From January 1, 2027: $150.00.
• Gemini 3.5 Flash-Lite, which shows no scheduled increase: $40.00.
• gpt-6-luna at $0.10 and $0.50 per million: $10.00.
• Claude Haiku 5.5 at $0.10 and $0.50 per million for prompts up to 100K tokens: $10.00.
• deepseek-flash at peak cache-miss rates of $0.30 and $1.20: $27.00, and roughly half that off-peak.

The ratios are what matter, not our example workload. After January 1, Flash will cost 15 times gpt-6-luna and Claude Haiku 5.5 per input token at list price. In our view that gap is wide enough to survive a generous allowance for quality differences on simple tasks, and it is the single strongest argument for moving cheap, high-volume calls off Flash before the new year.

Which Gemini models are not going up in price

Not everything moves, and the page is specific about what does. As of October 23, 2026, these rows carry no dated increase:

• Gemini 3.5 Flash-Lite at $0.30 input and $2.50 output, with a free tier.
• Gemini 3.1 Flash-Lite at $0.25 for text, image and video, $0.50 for audio, and $1.50 output.
• The Gemini 2.5 line, including 2.5 Flash at $0.30 and $2.50 and 2.5 Pro at $1.25 and $10.00 under 200K tokens.
• Gemini Embedding 2 at $0.20 per million text tokens.
• The live and transcription models, including 3.5 Transcribe at $2.00 audio in and $12.00 text out.

If your workload runs on Flash mainly because it was the default in a tutorial, the cheapest migration we can see is one line: drop to Flash-Lite within the same API, same SDK, same region. Test it on your own evaluation set rather than on benchmark tables - in our view the gap between a Flash and a Flash-Lite tier is task-shaped rather than uniform.

Cut the bill before you switch providers

Switching providers costs engineering time and re-evaluation, so exhaust the levers inside the API first. Google's pricing page documents three:

1. The Batch API, listed as a "50% cost reduction" on the paid tier. Argon batch is $1.00 and $5.00 today and $2.00 and $10.00 from December 1 - still half the standard rate after the increase. Anything that does not need a live response, including nightly enrichment and backfills, belongs here.
2. Context caching, though read the small print: Argon supports implicit caching only, with explicit caching not supported, and the cache prices double on the same dates as the models.
3. The free tier and Google AI Studio. The page offers "Free input & output tokens" on the free tier for Flash and Flash-Lite models, says "Google AI Studio usage is free of charge in all available regions", and shows no free tier for Argon. Prototyping is free; production is not.

A fourth lever sits outside the vendor. A gateway lets you move traffic between models without rewriting your client, which is what makes a price change survivable. OpenRouter states in its FAQ that it passes through provider pricing with no markup on inference, charging instead on credit purchases and a 5% fee on bring-your-own-key usage above a $25,000 monthly allowance. liteLLM offers the same shape as a self-hosted proxy with spend tracking, and LLM Stats is useful for comparing price and benchmark scores side by side before you commit. Upfynd lists 157 products in LLMs & AI Infrastructure as of October 2026, and gateways and routers are a large part of that shelf.

Cheaper alternatives to the Gemini API, by what you are running

DeepSeek's API pricing page shows deepseek-flash at $0.15 input and $0.60 output off-peak against $0.30 and $1.20 at peak, with peak defined as 01:00-04:00 and 06:00-10:00 UTC on weekdays, so a batch job scheduled outside those windows pays half. Mistral's pricing page lists Mistral Large at "$0.5 /M tokens in and $1.5 /M tokens out".

We'd skip the switch entirely in two cases. If your spend is small - a few hundred dollars a month - a doubling will likely cost less than the engineering week it takes to migrate and re-evaluate, and we'd spend that time on caching and batching instead. And if you are locked to Gemini for a reason that is not price, such as a Google Cloud commitment, a region requirement, or Antigravity in your coding workflow, then the increase is a budget line to update, not a migration to run. The same trap caught teams last month when model access moved behind extra charges in a coding tool, which is why we argued then that you should do the math before you renew.

What we'd do before December 1

1. Pull your last 30 days of token usage by model and multiply it by the new rates in the table above. The figure, not the percentage, decides whether this is worth any engineering time.
2. Move every non-interactive job to the Batch API this week. It is a flag, not a migration, and it halves the affected spend.
3. Run your evaluation set against one cheaper model per workload - Flash-Lite inside Gemini, plus one outside it - and record the quality delta before you need the decision.
4. Put the integration behind a gateway or router so the next price change is a config edit. Upfynd's 121 AI coding tools and the infrastructure category are a reasonable place to start that shortlist.

The dates are fixed and published. The useful work is knowing, before November 30, which of your calls you would move and which you would pay more for.