News · 2026-09-02

Google's new Flash model scores higher and costs more to finish a job

Google released Gemini 3.8 Flash on September 2, 2026, and the headline price per token did not change, but the cost of actually finishing a task did. Google says the model deliberately spends more effort on hard problems, and the independent measurement service Artificial Analysis found it burns roughly 30% more output tokens per task than the model it replaces, pushing the cost of a completed job from about $0.40 to about $0.58. It is a smarter model that is also a more expensive one, and the two facts live in different columns of the invoice.

Key facts

• Gemini 3.8 Flash and a restricted variant called Gemini 3.8 Flash Cyber launched on September 2, 2026.
• Introductory API pricing is $0.75 per million input tokens and $3.75 per million output tokens through December 31, 2026, then $1.50 and $7.50.
• Artificial Analysis measured about 48,000 output tokens per task, roughly 30% more than Gemini 3.7 Flash, lifting cost per task from about $0.40 to $0.58.
• Primary source: Google's launch post.

For two years the Flash tier has meant one thing: the cheap, fast model you reach for when you have a lot of small jobs and not much patience. Google is now bending that definition. It calls 3.8 Flash "our most intelligent workhorse model" and aims it at long-horizon software engineering, autonomous agents, and multi-step enterprise work rather than at bulk summarization.

The mechanism behind the improvement is unusually candid. Google writes that on complex tasks the model "executes extra reasoning steps, calls tools iteratively" and "might use more tokens to maximize performance." In other words, the gain is not a free architectural win. It is a decision to let the model think longer, check itself, and call tools again when it is unsure. Think of it as the difference between a contractor who quotes a fixed hourly rate and then takes more hours on a difficult job. The rate on the invoice is the same. The invoice is not.

The model card says 3.8 Flash is built on Gemini 3.7 Flash, supports a one-million-token context window and 64,000 tokens of output, and ships through the Gemini app, Google AI Studio, the Gemini API, Google's AI Mode in search, the Gemini Enterprise Agent Platform, and the Antigravity development environment. Alongside it, Google released Gemini 3.8 Flash Cyber, the same core model packaged for defensive security work and gated behind a vetting program.

The independent numbers are the interesting part. Artificial Analysis scores 3.8 Flash at 59 on its Intelligence Index, three points above 3.7 Flash, and clocks it at roughly 300 output tokens per second, which is genuinely fast for a model at this capability level. But the same analysis records the token inflation. The model's score went up; so did its appetite. Anyone who has run an agent loop overnight knows which of those two numbers shows up on the bill first.

Why this matters is a question of accounting rather than benchmarks. Most teams still budget model spend in dollars per million tokens, a unit that made sense when a request was one prompt and one answer. Once a model runs a multi-step agent loop, verifies its own output, and retries, the meaningful unit is dollars per completed task, and a per-token price cut can coexist with a per-task price increase. This is the same distinction that makes output tokens cost more than input tokens, and it is now the distinction that separates a headline price from a real one.

The Hacker News thread on the launch reached 859 points and more than 500 comments, and the split there was concrete on both sides. Practitioners praised the speed, the quality of generated HTML and JavaScript, and the fact that cheap, fast models are excellent when a task is verifiable and can simply be retried until it passes. That is exactly the workload Google is targeting. The pushback was equally specific: some users reported that coding reliability was still uneven, that the model handled current-information queries poorly, and that the improvement looked like the product of extra spend rather than a genuine efficiency gain.

That last objection is the strongest counter-argument, and it is hard to dismiss, because Google essentially concedes the premise. The honest caveat cuts the other way too, though. Spending more compute at inference time to get better answers is a legitimate engineering choice, not a trick, and it is the same idea behind every test-time compute result of the past two years. The question is not whether the trade is real. It is whether your workload wants it. If you are running verifiable, retryable jobs at volume, a faster model that occasionally thinks harder is a good deal. If you are paying per completed agent run, read the second number, not the first.

Primary source, verified: read the paper →

Ground Truth. (2026, September 2). Google's new Flash model scores higher and costs more to finish a job. Ground Truth. https://groundtruth.day/news/gemini-3-8-flash-scores-higher-and-costs-more-per-task.html

@misc{groundtruth:gemini-3-8-flash-scores-higher-and-costs-more-per-task,
  title  = {Google's new Flash model scores higher and costs more to finish a job},
  author = {{Ground Truth}},
  year   = {2026},
  month  = {sep},
  url    = {https://groundtruth.day/news/gemini-3-8-flash-scores-higher-and-costs-more-per-task.html}
}

Topics: models · google · gemini · agents · inference-cost · coding