Introductory prices are not a cost model

Google released Gemini 4 Argon this week, and the pricing line deserves as much attention as the benchmarks. Google calls it an introductory price: $2 per million input tokens and $10 per million output tokens, with cached input tokens at 95% off.

That is a good price for a frontier model. Google's own wording also makes clear that it is temporary. For a team putting agents into production, that word matters more than the number.

What happened

Three things from the past two weeks point the same way.

• Frontier prices keep moving. Argon launched on October 13 through the API in AI Studio and the Gemini platform, at the price above. Google says it ranks first on the Vals Index, which measures performance in sectors like finance, coding, legal and tax work. Simon Willison's September digest lists a pricing war among the month's themes.
• The vendors are watching their own margins. Simon Willison summarized Financial Times reporting that Anthropic told investors its Q3 revenue was $16.8bn. The same report says Anthropic earned $1.70 in revenue for every $1 spent on computing power, with adjusted operating margins of 11% excluding stock-based compensation. Prices are a business decision, and the people setting them care about margin.
• Splitting work across models is now a platform feature. Google added local model support to the Antigravity SDK. A cloud architect agent plans the work and local builder agents carry out the narrow tasks. In one demonstration, InfoQ reports that only 95 of about 3,400 tokens went to the cloud for planning, so 97.2% ran locally. A commenter quoted in the same piece made a fair point: token share alone does not tell you whether the run was faster or cheaper on your hardware.

William also argued earlier this month for default hard budget caps on usage-based services. His proposal is to cut the service off past a set amount, not just send a warning email:

"These need to be hard limits."

Why it matters for operations teams

An agent business case usually rests on cost per completed case: a claim triaged, a benefits call finished, a submission summarized. Token price is only one input to that number. The others are how many tokens a case uses, how many retries it needs, how much of the prompt hits the cache, and how often a person has to step in.

If the business case only works at this month's launch price, it is a bet on a price list someone else controls.

The opposite risk is real too. Prices fall or a cheaper model arrives, and a team stays on the expensive one because nobody can prove the cheaper one is good enough. Both risks have the same fix: know your cost per case, and be able to measure it again quickly.

How we build for it

Measure cost per case, not per token

Every evaluation run records tokens, cached tokens, tool calls and wall time for each case, next to the pass or fail. The result is a cost per correctly resolved case for each model under test. That is the number finance should see.

Stress-test the price

Before go-live we rerun the cost model at a multiple of today's list price, and again with the cache discount removed. If the economics break at double the price, we want to know before launch. Usually the fix is in the design, not a negotiation: shorter context, fewer steps, or a smaller model for the easy cases.

Design for the cache

A 95% discount on cached input rewards prompts with a long, stable opening section. Instructions, policy text and tool definitions go first, and case-specific material goes last. We track cache hit rate as a metric and check it after every prompt change. A small edit near the top of a prompt can quietly wipe out the saving.

Route by task, and measure the whole run

The architect-builder split suits back-office work. A capable model plans the case, and smaller models do extraction, classification and checks, either in the cloud or inside the client's environment. But we judge routing on end-to-end results: cases resolved correctly, wall time and cost. A high share of cheap tokens does not prove anything by itself.

Put hard caps in the workflow

We set a limit on steps and tokens per case, plus a daily budget per workflow. When a case hits its limit, the agent stops and hands it to a person with a note of what it tried. A runaway loop then costs one capped case, not a month of budget.

Keep the model swappable

Models sit behind a single interface in the client's cloud. Every candidate model runs through the same evaluation set before it reaches production. When an introductory price ends, or a better offer appears, switching is a measured decision rather than a migration project.

The short version

Treat a launch price as useful information, not as a plan. Build the agent so that cost per case is visible, capped and easy to measure again. Then a price change becomes a routine review instead of a surprise.

Sources

1. Gemini 4 Argon is now available for everyone, Google, October 13, 2026.
2. Anthropic tells investors quarterly revenue up nearly 50% to $16.8bn ahead of IPO, Simon Willison, October 15, 2026.
3. Google Antigravity SDK Adds Local Model Support for Greater Token Efficiency and Code Privacy, InfoQ, October 14, 2026.
4. Archive for Saturday, 3rd October 2026, Simon Willison, October 3, 2026.