From 1 December, the Same AI Work Costs Twice as Much. Who Planned for It?

Google is doubling the price of its newest model, Gemini 4 Argon. It is a legitimate move: the introductory price was announced in advance. For companies that have built production processes on the model, it is still a reminder that the largest variable in an AI budget is set outside the company.

What changes on 1 December

Gemini 4 Argon has been generally available since October. Google launched it at an introductory price of USD 2 per million input tokens and USD 10 per million output tokens. That price applies until 30 November 2026.

From 1 December it is USD 4 for input and USD 20 for output. The price of cached input and batch processing doubles as well. The output price also covers the tokens the model spends on thinking.

An agent that processes a monthly volume of invoices for USD 1,000 today will process it for USD 2,000 in December. Without a single change to the code, the process or the volume.

With agents the effect is stronger than with a regular chat. An agent calls the model repeatedly for a single task: it plans the steps, reads documents, checks the result and retries on error. Each of these steps is another batch of input and output tokens. The longer and more autonomous the process, the larger the share of its cost that depends on the vendor's price list.

The price per token is not the only variable

A price list change is the visible one. A bill can also grow in other ways.

In April, Anthropic released Claude Opus 4.7 at the same price as its predecessor. Its new tokenizer, however, maps the same text to roughly 1.0 to 1.35 times as many tokens, depending on the content type. At higher effort levels the model also thinks more, so it produces more output tokens.

Anthropic stated this openly and added that the net effect was favourable on its internal evaluation. It still recommends measuring the difference on real traffic. That is exactly the right advice. The result depends on what tasks the model performs, and only the company running it knows that.

The cost of a single task is therefore set by three things: the price list, the tokenizer and how much the model thinks during the task. The company sets none of them.

Almost everyone wants flexibility, few have built it in

In October, OutSystems published a survey of 858 executives (conducted by CIO Dive in June 2026, sample drawn from its databases; OutSystems sells a platform for agentic systems). 62% of respondents say they want to be able to switch models, vendors and platforms over time.

The same survey shows that 5% of companies have automated controls, audit trails and formal human review points in place. A company without an automated record of what an agent does will struggle to measure what it costs. Without that measurement it cannot judge whether moving to another model is worth it.

The ability to switch models is an operational capability. You will not find it in a vendor contract.

What to prepare before December

• Measure cost per task. Price per million tokens is the vendor's figure. What matters to the company is the cost of one processed invoice, one ticket, one contract review.
• Know which model performs which task. Document classification and complex analysis do not need the same model. Every task running on the most expensive model is a decision, not a default.
• Have a test that compares models on your own data. Before changing a model, you need to know whether the new one delivers the same quality. That test has to run on the company's real tasks.
• Set limits. A budget cap per process or per agent catches a price change before it shows up on the invoice.

None of this is new. Companies manage every other variable operating cost the same way. The only new thing is that with AI this cost can change from one month to the next.

For the 2027 budget this means one practical thing. The "AI" line should be the sum of costs of specific processes that the company can measure, compare and, if needed, move to another model.

On Elevon Platform every AI action is logged and auditable, and the model behind an agent can be replaced through configuration, without rewriting its logic.

Sources

• Google AI for Developers: Gemini Developer API pricing (ai.google.dev/gemini-api/docs/pricing)
• Anthropic: Introducing Claude Opus 4.7 (16 April 2026)
• OutSystems: The Agentic AI Era Is Here, Enterprise Readiness Is Not (15 October 2026)