Gemini API Pricing (2026): Cost per 1M Tokens for Every Model

Gemini API pricing starts at $1.25 per 1M input tokens on Gemini 2.5 Pro and reaches
$2 on Gemini 3.1 Pro, with output between $7.50 and $12. The headline rates are
competitive, but Gemini’s pricing has a feature none of its rivals share to the same degree:
prompts above roughly 200K tokens are billed at a higher rate on several models. For a family
sold on its million-token context, that is the number that decides your real bill.

Which Gemini model should you use?

The interesting thing about Google’s current line is that the cheaper tier is often the
better one.

Gemini 3.5 Flash outperforms Gemini 3.1 Pro on coding and agentic
benchmarks while running roughly four times faster and about 25% cheaper, with the same
1M-token context. When the faster, cheaper model also scores higher, the Pro tier stops being
the default and becomes a specialist choice. For coding assistants, agent loops and anything
where a user is waiting on a response, Flash wins on every axis at once — which is rare in
model selection, where speed, cost and quality normally trade against each other.

Gemini 3.6 Flash, released 21 July 2026, is the newest in the line. Its
input rate is unchanged at $1.50 while output drops from $9 to $7.50 — a 17% cut on the side of
the bill that usually dominates. For generation-heavy workloads that is a straightforward
saving with no capability trade-off implied by the pricing, which makes it the default choice
within the Flash tier for new work. On prompt-heavy retrieval workloads the two are
interchangeable on cost, since input is identical. To clear up a common search: there is no
Gemini 4 as of July 2026.

Gemini 3.1 Pro is worth choosing for multimodal depth rather than by
default, given that Flash beats it on coding at a lower price.

Gemini 2.5 Pro remains widely deployed and well supported. It reached
general availability on 17 June 2025 and was a benchmark leader in coding, mathematics and
long-context reasoning, with a native 1,048,576-token context and fully multimodal input — text,
images, audio, video and PDFs. Newer 3.x models supersede it, but at $1.25 input it is still
one of the strongest value picks among frontier reasoning models.

The 200K threshold is the thing to plan around

Gemini’s long-context billing is where budgets break. Gemini 3.1 Pro bills prompts above
200K tokens at roughly double the input rate. Gemini 2.5 Pro is explicitly tiered: $1.25 per
1M input up to 200K, rising to $2.50 above it, with output moving from $10 to $15.

The practical consequence is that a feature costing one amount in testing can cost twice
that in production, because the threshold is crossed per request and depends entirely on how
much a user pasted in. Budgeting a long-context feature on the headline figure will understate
the bill substantially once real prompts start arriving.

There are two sensible responses. Design around the threshold with aggressive retrieval and
reranking so prompts stay under it — which usually improves answer quality anyway, since
attention across very long contexts is not uniform. Or compare against a model that charges one
flat rate across its whole window; Claude Opus
4.8 has no long-context premium at all, which for a genuinely retrieval-heavy workload can
work out cheaper despite a higher sticker price.

Google’s open models are priced separately

Gemma 3 is Google’s open-weight family and sits outside this table because it is a different
proposition: 4B, 12B and 27B models under an open licence, multimodal across text and images,
128K context, and cheap enough on hosted APIs ($0.05 to $0.16 per 1M) that the real decision is
whether to run them yourself. Gemma 3 27B needs about 16 GB of VRAM at 4-bit, which puts it on
a single consumer card. Size it in the
VRAM calculator, or work out whether
self-hosting beats the API at your volume with the
self-hosting vs API
calculator.