Gemini 3.8 Flash launches — the fourth Flash model in four months, at $0.75 per million tokens

The three lines

• Cadence — the fourth Flash model in under four months, three weeks after the last one
• Price — 0.75 dollars input and 3.75 output per million tokens, half price through end-2026
• Score — 59 on the Intelligence Index, up three points, level with far pricier models

Key questions

How much does Gemini 3.8 Flash cost?

**0.75 dollars per million input tokens and 3.75 dollars per million output tokens.** That is an **introductory rate through December 31, 2026**; the standard price is 1.50 and 7.50 dollars, so the same workload doubles in cost from 2027. Output costing five times input is standard across the industry and reflects how generation works rather than any pricing quirk. In task terms, Artificial Analysis measures a typical high-reasoning task at about **0.58 dollars and 2.5 minutes**, dropping to roughly **0.8 minutes** at low reasoning. That spread matters more in production than the headline rate does: the same model, set to a different reasoning effort, changes both latency and cost by around threefold.

How much better is it?

**59 on the Artificial Analysis Intelligence Index at high reasoning, three points above the previous version.** By setting: 59 high, 57 medium, 52 low. The number matters because of its neighbours — **GPT-5.6 Sol at extra-high reasoning and Grok 4.6 at medium also score 59**. A cheap workhorse model has reached the band of models built to be flagships. The context window is unchanged at **one million tokens**, inputs cover text, image, video and speech, and output runs at roughly **300 tokens per second**. Google's own emphasis was on **agentic work** — tool use and multi-step real tasks — rather than on raw knowledge. As always, a benchmark is one organisation's measurement under one configuration, and the same model can score very differently elsewhere.

Why release these so often?

**Because the cheap tier is where the competition actually is.** Four Flash-class releases in under four months, the latest three weeks after the last. Frontier flagship models take longer to build and cost enough that usage stays selective. The cheap tier is different: it is the layer that gets **wired into products at volume** — search summaries, document processing, code completion, customer support. At hundreds of millions of calls a day, the deciding factor is not a benchmark point but the **unit price**, which is why a three-week refresh cycle is a rational strategy here rather than a sign of haste. The broader pattern — last year's frontier performance becoming this year's commodity price — is the structural feature of this market.