Gemini 4 API Pricing (2026)

Official rates for Gemini 4 Argon and Gemini 4 Carbon, the introductory price that ends November 30, and what real workloads actually cost.

Verified November 11, 2026. Gemini Developer API / Vertex AI, USD per 1M tokens. Independent guide — not affiliated with Google.

TL;DR: the numbers that matter

	Intro price (until Nov 30, 2026)	Standard (from Dec 1, 2026)
Input (text, image, video, audio)	$2.00	$4.00
Cached input (implicit)	$0.10	$0.20
Output (answers and thinking tokens)	$10.00	$20.00

Both Gemini 4 Argon (gemini-4-argon) and Gemini 4 Carbon (gemini-4-carbon) share the same rate card. Context window is 1,048,576 tokens (input and output share it); default synchronous output is capped at 65,536 tokens, extendable up to the full context via Long Decode. Non-global regions pay 10% more ($2.20 → $4.40 input, $11 → $22 output).

Act on this: the price doubles on December 1

Google's 50% introductory promotion runs through November 30, 2026. From December 1, input goes from $2 to $4 per million tokens and output from $10 to $20. If you are validating Gemini 4 for a production workload, run your volume tests now — your December bill will be exactly double at the same traffic unless you shift work to caching, batch, or a cheaper tier.

What counts as billable output

Thinking tokens are billed as output. Argon can generate tens of thousands of reasoning tokens per task even when the final answer is short. A "cheap-looking" prompt can produce a large bill if reasoning depth is set high.
Long completions are billed even on failure. From November 13, 2026, when a completion exceeds 32K tokens, you are charged for the retrievable tokens even if the request ultimately fails on a client or server timeout.
Implicit caching only. There is no explicit-cache API on Gemini 4; cached-input pricing ($0.10 intro / $0.20 standard) applies automatically when your repeated context hits the cache.

Batch and Flex: 50% off

Batch and Flex processing are typically 50% below list price. For document pipelines and other workloads where hours of latency are acceptable, batch moves the effective output rate to $5 intro / $10 standard per million tokens — cheaper than most premium alternatives.

Real workload cost examples

At introductory rates, cached-input share matters more than raw price. Three common workloads, 30 days:

Workload (per day)	Per request	Per month, no cache	Per month, 70% cached input
Support chatbot — 500 conversations (5K in / 1.4K out)	$0.024	$360	$159
Coding agent — 300 sessions (60K in / 12K out)	$0.24	$2,160	$1,071
Long-context analysis — 200 runs (250K in / 4K out)	$0.54	$3,240	$1,647

Cache math: if c is the cached share of input tokens, monthly cost ≈ uncached cost × (1 − 0.95 × c). At 70% cached input you pay about a third of the uncached bill. Re-sending the same system prompt and conversation history is exactly the pattern implicit caching rewards. Note: agents that re-read files in multi-step loops cache well; one-shot long-context runs that never repeat input do not.

Gemini 4 vs the alternatives, per million tokens (intro rates)

Model	Input	Output	Positioning
Gemini 4 Argon / Carbon (intro)	$2.00	$10.00	Google flagship; Carbon targets complex coding, research, security
Gemini 3.1 Pro	$2.00	$12.00	Previous flagship tier, still available
Gemini 3 Flash	$0.50	$3.00	High-volume, low-latency work
Gemini 3 Flash-Lite	$0.25	$1.50	Cheapest current Gemini tier

Claude Sonnet 5.5 sits within a rounding error of Argon's intro output price at the same input price — at that pairing, decide on quality measured on your own workload, not the rate card. Route the majority of traffic that never needed a flagship down to Flash or Flash-Lite and the price change on December 1 barely moves your total.

FAQ

How much does the Gemini 4 API cost?

$2 per million input tokens and $10 per million output tokens at introductory rates through November 30, 2026; $4 / $20 standard from December 1. Cached input is $0.10 intro, $0.20 standard. Both Argon and Carbon use the same rates.

Why is my Gemini 4 bill higher than the rate card suggests?

Three usual causes: thinking tokens billed as output, requests over 32K output tokens billed even when they fail (from November 13), and uncached repeated context. Check reasoning depth first — it is the biggest lever.

Does Gemini 4 have explicit prompt caching?

No. Caching is implicit only — Google caches repeated prefixes automatically and bills hits at $0.10 per million tokens intro. You cannot pre-warm or pin cache entries.

Is Gemini 4 Carbon more expensive than Argon?

No. Carbon launched with the same intro pricing as Argon ($2/$10, standard $4/$20 from December 1). The difference is capability for long-horizon reasoning tasks, not price.

Where is the official price page?

Google's Agent Platform pricing page and Gemini API docs list current rates; promotional terms state the November 30 end date directly. Prices verified November 11, 2026.

Calculate your cached-input share and monthly bill →