Flash isn't cheap anymore

Google DeepMind shipped Gemini 3.5 Flash at I/O on Tuesday, and the model card is doing something I want to read carefully.¹ The headline capability story is straightforward: 3.5 Flash beats the previous Pro tier on coding and agentic benchmarks, scores 83.6% on MCP Atlas (an agentic multi-step benchmark Google is positioning as the new measure of agent reliability), and runs roughly four times faster than what Google calls "comparable frontier models". Dynamic thinking is on by default. Context window is one million tokens. It is, by Google's account, the best small model in the world.

The pricing is where the analysis lives.

What Flash now costs. Gemini 3.5 Flash is priced at $1.50 per million input tokens and $9.00 per million output tokens, generally available on the Gemini API and AI Studio from launch.¹ Gemini 2.5 Flash, the immediate predecessor with thinking enabled, sat at $0.25 input and $3.50 output.² So this is a 6x input price increase and a roughly 2.6x output price increase against the model it directly replaces. Against Gemini 2.0 Flash, the non-thinking Flash baseline at $0.10 and $0.40, the multiples are 15x input and 22x output. Headline framing depends on the comparator the framer chooses.

Either way, this is not the Flash tier as the market has understood it. Flash was the price-compression instrument. It was the part of the menu where you could read inference economics moving in the direction labs liked to advertise — cheaper per token, generation over generation, with capability improving alongside. That trade is over, at least at Google.

What Simon Willison noticed. Willison flagged this on his blog the day before launch: "all three major labs raised effective prices in recent cycles."³ He's right, and it's the structural point. OpenAI's reasoning tiers, Anthropic's Sonnet repricing, and now Google's Flash — the cheapest tier where customers actually live is moving up across the board. Inference is not getting cheaper for the buyer. It's getting cheaper for the lab, possibly, and the spread is being kept.

Why thinking-on-by-default matters for the bill. Dynamic thinking active by default is the structural mechanism doing some of the work here. Every API call generates a reasoning trace before it generates a final response, and that trace bills at the $9 output rate. The headline token price understates the per-task cost for any workload that calls the model many times in sequence — which is to say, every agentic pipeline. Google has not published comparative trace lengths for 3.5 Flash versus 2.5 Flash, but the direction is clear: a higher per-token price multiplied by more tokens per task. The counter-argument, fairly, is that better reasoning may reduce the number of turns required to complete a multi-step task, and the net could be flat or favourable. No task-economics comparison has been published.² I would like to see one.

The frame this fits. This is inference economics, but with the polarity inverted from the version labs were selling 18 months ago. The story then was: scale compresses cost, Flash tiers prove it, developers get cheaper tokens forever. The story now is: scale compresses cost for the lab, but the lab is testing how much of that compression it has to pass through. Google appears to have decided: not much. The fact that 3.5 Flash beats 3.1 Pro on benchmarks is being used to justify Pro-tier pricing on the Flash SKU. The SKU name is doing a lot of work it didn't used to do.

The SKU name is doing a lot of work it didn't used to do.

The split that's worth naming. Google is simultaneously rolling 3.5 Flash into free consumer surfaces, the Gemini app, AI Mode in Search, while charging API customers $1.50/$9.³ This is a deliberate two-tier structure and it isn't neutral. The consumer product is being defended against agent-mediated disruption of Search, and the cost of that defence is being recovered from API customers who do not have an equivalent distribution channel to fall back on. Developers subsidising the consumer moat is a familiar pattern at Google. It's now operating at the model layer.

This is also where the agent-economics frame starts to bite. If MCP Atlas at 83.6% holds up outside the benchmark (the usual caveat), and 3.5 Flash becomes the default backbone for production agent pipelines, Google captures ongoing inference revenue at the new price tier from every agentic transaction routed through those pipelines. The benchmark leadership is not just a capability claim; it's a positioning claim about where the agent economy's inference spend lands. Anthropic and OpenAI presumably read the model card the same way I did.

What's a case of what. A few months ago I'd have called this an Anthropic-style move — lead with capability, price like you mean it, let enterprise demand absorb the increase. Google running the same playbook on the Flash tier suggests the pattern has generalised. The labs have collectively decided that the price-compression narrative was a phase, not a trajectory. The phase ended somewhere between Sonnet 3.5's repricing and this week's Flash launch. There is no longer a frontier lab visibly competing on price at the API tier. They are competing on capability, and pricing to capability.

The counterpoint, which I want to give its weight: 3.5 Flash is genuinely a more capable model than 2.5 Flash, and a model that beats the prior Pro tier has some legitimate cost basis for higher pricing.¹ This is not straightforward rent extraction. It may be that the product is better and costs more to serve, and the price reflects that. The question is whether the rate of price increase tracks the rate of underlying cost change, and on that we have no disclosure. Labs don't publish unit cost per token. They publish prices.

What to watch. Three things. First, Gemini 3.5 Pro pricing, expected next month. If Pro lands at $5/$20 or above, that confirms a systematic upward repricing across the stack and the current Flash number is a floor rather than a ceiling. Second, whether OpenAI and Anthropic reprice their equivalent tiers within the quarter — if they do, Willison's "all three labs" observation calcifies into a market structure rather than a coincidence. Third, the trace-length question: if anyone publishes per-task economics comparing 3.5 Flash to 2.5 Flash on an agentic workload, that's the number that tells you whether thinking-on-by-default is net favourable or net extractive for the customer.

I would not assume favourable.

Footnotes

¶