# December 1 on the Gemini API: a price rise and two forced model swaps

> On December 1, Gemini 4 Argon's list price doubles and Google reroutes Gemini 2.5 Pro and 2.5 Flash to newer models. Why we handle model changes as incidents, and how to build an agent stack that switches on your schedule.

_Youssef Boussafa, Nov 12, 2026._

## What changes on December 1

Two notices on Google's Gemini API take effect on the same day, and they hit different teams.

The first is a price change. Gemini 4 Argon went GA on October 13 at an introductory rate, and Google's pricing page now dates the end of it. Standard input moves from $2 to $4 per million tokens and output from $10 to $20 on December 1, 2026. The other tiers move with it: batch and flex go from $1 / $5 to $2 / $10, priority from $3.60 / $18 to $7.20 / $36, and Standard cached input from $0.10 to $0.20 per million tokens. We flagged the step-up in our [October guide to the autumn model wave](/insights/autumn-2026-frontier-models-builders-guide). It's now on the calendar. Gemini 4 Carbon, which Google's changelog made generally available on November 10, follows the same schedule: Standard pricing goes from $2 / $10 to $4 / $20 on December 1, and batch from $1 / $5 to $2 / $10.

The second is a model swap. On November 2, Google's changelog deprecated two widely used models. From December 1, every API request to `gemini-2.5-pro` will be routed to Gemini 3.8 Flash (`gemini-3.8-flash`) and billed at its rates, and every request to `gemini-2.5-flash` will go to Gemini 3.5 Flash-Lite (`gemini-3.5-flash-lite`). Both old models have already been removed from the AI Studio model picker. Google asks developers to change their model string before the date and to read the migration guide, which lists breaking changes: `thinking_level` replaces `thinking_budget`, sampling parameters and prefilled model turns are removed, and function calling validates `thought_signature`.

Teams that costed Argon at its launch price will see the bill double. Teams with either 2.5 model string in a service, a notebook, a scheduled job or a vendor tool running on their API key will have a different model answering on December 1, and that group worries us more. Teams without model-version monitoring or behavioral evaluations may miss the change.

## A silent reroute is a quality incident

Google's models documentation recommends a stable model string for most production apps and says that "stable models usually don't change." `gemini-2.5-pro` is a stable string. On December 1 it will point at a model from a different tier, a Flash model standing in for a Pro.

For a simple chat feature that may be acceptable. For an agent, it changes every step that uses the model: how it plans, which tools it picks, how it formats JSON, when it refuses, how long it thinks. Google bills thinking tokens as output, so the cost of a call also depends on the new model's habits. An agent whose prompts were tuned against 2.5 Pro can start failing in ways your error monitoring won't see. Behavioral regressions can occur even when requests succeed; incompatible parameters may also cause explicit errors.

So we treat the date the way we'd treat a planned outage of the model we tested. It gets an owner, a runbook, a rollback plan and a review afterwards. Google gave 29 days of notice. That's comfortable if your stack can switch models and tight if it can't.

The cost side is less obvious than it looks. Take one step on 2.5 Pro at list prices: 100,000 calls a day, each reading 8,000 tokens and writing 1,000, which is 800 million input and 100 million output tokens a day.

| Model and rate (input / output per 1M) | Input | Output | Per day |
|---|---|---|---|
| Gemini 2.5 Pro today ($1.25 / $10) | $1,000 | $1,000 | $2,000 |
| Routed to 3.8 Flash in December ($0.75 / $3.75) | $600 | $375 | $975 |
| 3.8 Flash from January 1, 2027 ($1.50 / $7.50) | $1,200 | $750 | $1,950 |

The December invoice for that step roughly halves, which will look like good news on a finance dashboard and says nothing about whether the step still works. Token counts per call will shift too, and Google's pricing page already shows 3.8 Flash doubling on January 1, 2027. The 2.5 Flash path is quieter still. For text, image and video input, both models list $0.30 per million input tokens and $2.50 per million output tokens, so the unit prices hold while the model changes. Actual spend can still move with token usage, and audio input rates differ.

Argon runs the other way. A workflow that runs 20,000 times a day at 30,000 input and 2,000 output tokens per run uses 600 million input and 40 million output tokens. That's $1,200 + $400 = $1,600 a day in November and $2,400 + $800 = $3,200 a day from December 1, or $49,600 more across December, with no change to the workload.

## Designing an agent stack that can switch

None of this is unique to Google. OpenAI's changelog notes that its `chat-latest` snapshot is regularly updated to track ChatGPT, and any vendor can reprice a model it launched at an introductory rate. The pattern we use is the same whoever sends the notice.

**One gateway for model calls.** Product code asks for a capability, such as "classify this ticket", and a gateway resolves it to a provider, a model string, a prompt version and a fallback. [Zenyt](/work/zenyt) is built this way: its specialized agents (product data, visual content, checkout, SEO and others) sit on multi-provider LLM routing, so the model behind one agent can change without touching the rest. Promptify's execution engine does the same across OpenAI, Anthropic, Cohere, Replicate and Hugging Face models. The cost is real. A gateway is code you maintain, and it tends to hide provider-specific features unless you design a way through for them.

**Assign models per step.** A planning step and a field-extraction step rarely need the same model. TaxFigure's stack lists several models behind a research coordinator that splits each question into sub-tasks, which keeps any one step easy to move. Record why each step uses its model, so a deprecation means re-deciding a couple of steps with data.

**Pin versions where the provider allows it, and know where it doesn't.** Never run a `-latest` alias in production; Google says its `latest` aliases are swapped with each new release, with two weeks' email notice. But the November 2 notice shows that a stable string can also be rerouted at the provider's end. Log the model version the response reports, where the API returns one, and alert when it differs from the model you requested.

**An evaluation set you own.** Our October guide describes how we build one: 100 to 300 real, anonymized tasks with graders. Run it on the current model and every candidate before a switch, then again on sampled production traffic after it. The before-and-after comparison is what turns "it seems fine" into a decision someone can sign.

**Shadow traffic, then a canary.** Send copies of real requests to the candidate and compare outputs offline, then route a small share of live traffic, for example 5%, adjusted to the workload's risk and volume, and widen it over a few days. For agents with side effects, such as sending emails or writing records, shadow mode needs those tool calls stubbed out, which is more work than most teams expect.

**Fallbacks that have been scored.** A fallback your eval set has never run on that step is a guess.

**Budget alerts and cost per outcome.** A per-workflow budget alert catches the Argon doubling on its first day. Cost per resolved ticket or audited page, which we covered in our piece on agent costs, tells you whether a cheaper routed model is still cheaper once retries and rework are counted.

## A pre-December 1 checklist for teams on Gemini

- Search code, config, notebooks, infrastructure templates and scheduled jobs for `gemini-2.5-pro` and `gemini-2.5-flash`, and check usage by model in Google's billing reports to find callers you don't own.
- Ask vendors whose products call Gemini on your key which models they use and what they're changing.
- Reprice every Argon and Carbon workload at $4 / $20 (and $2 / $10 for batch), and raise or reset budget alerts before the first December invoice.
- Run your eval set on `gemini-3.8-flash` and `gemini-3.5-flash-lite` for every affected step, plus at least one model from another provider. GPT-6.1 Sol lists at $2 / $10 for prompts up to 272K input tokens, which is Argon's rate until November 30. Google's chosen replacement isn't necessarily your best option.
- Migrate the parameters on the breaking-change list. The changelog doesn't say whether a routed request that still sends `thinking_budget` will fail or be ignored, so test it yourself before December 1.
- Switch to the model you chose, behind a canary, before the deadline. A switch you schedule comes with a rollback plan; the automatic one applies to all your callers at once, on Google's timetable.
- If you land on 3.8 Flash, budget for $1.50 / $7.50 from January 1, 2027.

## Vendor concentration is a procurement question

Anthropic's S-1, filed with the SEC on October 26, describes the model market from the seller's side. Anthropic tells investors that many of its customer contracts let customers reduce usage or terminate with limited or no notice, and that its top five customers accounted for about 25% of revenue in the nine months to September 30. It also lists deprecation as a risk in itself: customers who have built older models into their workflows may resist migrating, and a forced deprecation could lose it customers. And it says it has changed usage limits, rate limits and model capabilities before and may do so again.

For buyers, that cuts both ways. Your ability to leave is real, and the vendor's own risk section assumes it. Deprecations and limit changes are also routine, even at a vendor that says it may delay retiring models.

In procurement, we'd ask each model vendor for a minimum notice period before a deprecation or price change, a written commitment that a pinned model version won't be rerouted without notice, and per-request records of which model served each call. If a vendor offers a discount for a committed multiyear spend, put a price on the switching option you'd be giving up.

A second provider has costs. It doubles part of your eval work and prompt maintenance, and for a small team on one well-understood workload, a single vendor with strict version discipline can be the better call. What we'd avoid is learning on December 2 that a model you never chose has been answering your customers. If you have Gemini 2.5 in production and no eval set yet, there's still time to build a small one, and [we can help with the switch](/contact).

## Sources

- [Gemini API pricing](https://aistudio.google.com/docs/pricing), Google AI Studio docs, accessed November 12, 2026
- [Gemini API changelog](https://aistudio.google.com/docs/changelog), Google AI Studio docs, entries of October 13, November 2 and November 10, 2026
- [Gemini models](https://aistudio.google.com/docs/models), Google AI Studio docs, accessed November 12, 2026 (model version types)
- [API changelog](https://developers.openai.com/api/docs/changelog), OpenAI for Developers, entries of September 29 and October 7, 2026
- [Anthropic, PBC Form S-1 registration statement](https://www.sec.gov/Archives/edgar/data/2130357/000162828026067168/anthropicpbc-sx1.htm), U.S. Securities and Exchange Commission, October 26, 2026

## Related

- All articles: https://matious.com/insights.md
- Case studies: https://matious.com/work.md
- Start a project: https://matious.com/contact

## Canonical URL

https://matious.com/insights/llm-price-change-model-switching-2026

---

_Matious - AI engineering studio, founded 2017, Mohammedia, Morocco._
Contact: contact@matious.com | +212 661 790 436 | https://matious.com/contact
Machine-readable: [llms.txt](https://matious.com/llms.txt) - [agent instructions](https://matious.com/agents.md) - [sitemap](https://matious.com/sitemap.xml)
Every page is also available as Markdown: send `Accept: text/markdown`, or append `.md` to the URL.