Gemini 4 Argon API pricing: $2 / $10 until Nov 30, then $4 / $20

Gemini 4 Argon costs $2.00 per 1M input tokens and $10.00 per 1M output tokens on Google's standard paid tier, but only through November 30, 2026. From December 1 it's $4.00 and $20.00.

It isn't only the standard tier. Every token rate Google lists for this model doubles on the same date: input, output and cached input on the standard, batch, flex and priority tiers. And the output price includes thinking tokens, so the reasoning the model does before it answers is billed at the output rate, whether you see it or not.

All Google figures below come from Google's Gemini API pricing page, last updated October 19, 2026 (UTC) and checked on October 20. The arithmetic is ours.

What Google released#

gemini-4-argon became generally available in the Gemini API on October 13, 2026. Google calls it the “most capable model built for complex coding, agentic, math, and science tasks” (changelog). That's Google's description; we haven't tested it.

It's a paid-tier model: the pricing page says it isn't available on the free tier. Per Google's launch post, developers can build with it in the paid Gemini API, Google AI Studio and Android Studio, and Google AI Ultra subscribers get it in Google Antigravity too.

The models page lists it as stable and describes it as “Our flagship reasoning model with a unified 1M context window and long decodes up to 1M tokens”. It doesn't list formal input and output limits, supported modalities, reasoning-depth controls, tools or a knowledge cutoff, so this article doesn't cover them either. The launch post has no prices, context size or benchmarks.

Argon's price, tier by tier#

USD per 1M tokens, as on Google's pricing page on October 20, 2026.

That's prices only. The latency and availability terms for Flex and Priority aren't covered here, so check Google's page before moving a job to either tier.

Cached input costs 5% of the input rate on every tier, a 95% discount, both before and after the change. On the standard tier:

• $0.10 / $2.00 = 0.05 through November 30
• $0.20 / $4.00 = 0.05 from December 1

Three more details that matter for the math:

• The price is the same for any prompt length.
• Only implicit caching is supported; Google's page says explicit caching isn't. It lists no minimum cache size and no storage price.
• “Through November 30, 2026” is all Google says about the end date. No hour, no time zone.

What it does to a bill#

Take a hypothetical workload: 1M input tokens and 0.2M output tokens, where “output” means everything billed at the output rate, the visible answer plus any thinking. On the standard tier:

• 1 × $2.00 + 0.2 × $10.00 = $4.00 through November 30
• 1 × $4.00 + 0.2 × $20.00 = $8.00 from December 1, same tokens

Output costs more than its share of the tokens suggests. Per token it costs five times the input rate ($10.00 against $2.00), so those 0.2M output tokens already cost as much as the whole 1M of input:

• 0.2 × $10.00 = $2.00 for output
• 1 × $2.00 = $2.00 for input

Thinking is the wildcard. The Google pages we checked don't say how many thinking tokens a typical request uses, and the third-party estimates going around have no first-party source, so they're left out. What you can calculate is the price of every extra 100K thinking tokens:

• 0.1 × $10.00 = $1.00 through November 30
• 0.1 × $20.00 = $2.00 from December 1

Cache hits push the other way. If 0.8M of the 1M input tokens were billed at the cached rate:

• 0.2 × $2.00 + 0.8 × $0.10 + 0.2 × $10.00 = $2.48 through November 30
• 0.2 × $4.00 + 0.8 × $0.20 + 0.2 × $20.00 = $4.96 from December 1

That split is hypothetical, not a promise of what gets cached. The same base workload, 1M input and 0.2M output, on every tier:

Hard steps on a flagship, the rest elsewhere#

The usual fix for a bill like this doesn't depend on the vendor: split the job by step.

• Hard steps (planning, changes that span many files, anything where a wrong answer costs you an hour) go to a flagship.
• Routine, high-volume steps (extraction, classification, summaries, boilerplate, reformatting) go to a cheaper model.
• Before moving a step, run a sample of your own data through the cheaper model and read what comes back. A price table can't tell you whether it's good enough.

That's our rule of thumb, not a benchmark result.

Routine steps on CheapTokens#

If most of your tokens go to routine steps, three open-weight models on CheapTokens are worth pricing. The comparison below covers price and limits; check quality on a sample of your own data:

• DeepSeek-V4-Flash for output-heavy work such as long code or long documents: up to 32,768 output tokens per answer, 380K context. It's below OpenRouter's reference price on output.
• GLM-5.3-Flash for reading and summarizing long material: 390K context, up to 8,192 output tokens per answer. It's below OpenRouter's reference price on both input and output.
• MiniMax-M2.7 for bulk volume where price per token matters most: 200K context, up to 8,192 output tokens per answer. It's below OpenRouter's reference price on both input and output.

The OpenRouter comparisons are as of October 20, 2026. Output-heavy work first: say 0.2M input and 0.6M output tokens, a run of long generated documents. On Argon's standard tier:

• 0.2 × $2.00 + 0.6 × $10.00 = $6.40 through November 30
• 0.2 × $4.00 + 0.6 × $20.00 = $12.80 from December 1

The same tokens on MiniMax-M2.7 cost $0.10 at the current CheapTokens price. That's price per token, not price per solved task: if a cheaper model needs a second attempt or a longer prompt to get a usable result, the math changes.

Reading-heavy work looks different. Here's the 1M-input, 0.2M-output workload from above, priced live for each model, with an OpenRouter column for reference:

The table reads current prices every time the page loads, so it stays right when rates change. The Argon figures above are dated October 20, 2026 and won't update themselves.

Check limits before moving anything. Google describes Argon with a 1M context window and long decodes up to 1M tokens; a step whose prompt is longer than a model's context, or whose single answer is longer than its output cap, can't move to it. Each model's limits are in the list above.

On our side, one prepaid balance covers all three models. The API is compatible with OpenAI Chat Completions and Anthropic Messages, so a client that already speaks one of them needs a new base URL (https://api.cheaptokens.dev/v1), a CheapTokens API key and a model name. Prices are fixed per model between changes, and there are no peak hours. Your first API request with CheapTokens walks through the first call.

Before December 1#

1. Run a sample of real requests and put the output tokens from your bill next to the length of the visible answers. The difference is roughly your thinking overhead.
2. Price the workload at the December rates, not the introductory ones, if it will still be running after November 30.
3. Sort the steps by token volume, and test a cheaper model on the high-volume routine ones with your own data.

To plug in your own volumes, here are the current CheapTokens rates:

For agents, where one task means many calls, see How to estimate the cost of an AI agent. GLM vs DeepSeek vs MiniMax helps pick between the three models, and DeepSeek API pricing in 2026 looks at what you actually pay for DeepSeek models.

Current prices for every model: CheapTokens pricing.