Gemini 4 Argon: Pricing, Benchmarks, Access
Gemini 4 Argon costs $2 and $10 per million tokens, then $4 and $20. It ties GPT-6 Astra on the AA index and is cheaper per task only until the rate doubles.
Gemini 4 Argon costs $2 per million input tokens and $10 per million output at launch, and $4 and $20 once Google’s introductory period ends. On the independent Artificial Analysis index it ties GPT-6 Astra at 53, five points behind Claude Opus 5.5. At the introductory rate it finishes an index task for 61% of what Astra costs. At the standard rate it costs 22% more than Astra. Google has not said when the switch happens, and today almost nobody can buy it at either price.
Google announced Gemini 4 Argon on September 30, 2026. It is the first model of the Gemini 4 generation, the pre-training run Google first confirmed in July.
| Gemini 4 Argon | Introductory rate | Standard rate |
|---|---|---|
| Input, per Mtok | $2.00 | $4.00 |
| Output, per Mtok | $10.00 | $20.00 |
| Cached input, per Mtok | $0.10 (95% off) | $0.20 (95% off) |
| When it applies | From launch, end date not stated | After the introductory period |
| Who can buy it today | Fairwind Program cyber defenders only | Not yet announced |
| Next in line | Paid API customers, Google AI Ultra | “As soon as possible”, no date |
| Output limit | 1M tokens, up from 64K | Same |
How much does Gemini 4 Argon cost?
Two prices, and Google publishes both on day one. Google’s announcement lists an introductory $2 per million input tokens and $10 per million output, with cached input tokens “priced at 95% off input token price”, which is $0.10. It then states that “$4 per 1M input tokens and $20 per 1M output tokens will apply” after the introductory period. It does not say when that period ends.
That standard rate is exactly the Claude Opus 5.5 rate card, $4 and $20, which Anthropic launched on September 22. The introductory rate is exactly GPT-6.1 Sol’s $2 and $10, OpenAI’s reasoning tier from DevDay on September 29. So Argon opens priced like a mid-tier model and is scheduled to become priced like a flagship.
Google has run this two-card pattern before. Gemini 3.6, 3.7 and 3.8 Flash all carry an introductory $0.75 and $3.75 that Google’s own pricing page says doubles on January 1, 2027. The difference with Argon is that the Flash cards carry a date and Argon’s does not. Any annual budget built on $2 and $10 is a budget built on a number with no stated lifespan.
When can you use Gemini 4 Argon?
Not yet, unless you defend networks for a living. Google is rolling Argon out first “to a set of trusted cyber defenders through our Fairwind Program”, the same access gate its Gemini 3.8 Flash Cyber variant ships behind, and for those defenders and Google’s own teams it releases a version “without cyber guardrails”. Google says it is “actively engaged in the U.S. government’s voluntary process for pre-release model access”.
The general release comes next, in Google’s words “as soon as possible”, starting with paid API customers and Google AI Ultra subscribers. There is no date. As of October 1, Argon does not appear on the Gemini API pricing page or the Gemini API models page, and there is no published model ID. This site’s AI model tracker lists it as a preview: released, but to a limited audience. The cost calculator prices it at the introductory $2 and $10 Google published, so treat any estimate there as the introductory-rate case.
What does a task cost on Gemini 4 Argon?
The rate card is the wrong place to start, because Argon is a verbose model. Artificial Analysis, an independent benchmarking service, ran its full Intelligence Index on Argon at high effort and recorded 110 million output tokens. GPT-6 Astra at max effort needed 60 million for the same index, and Claude Opus 5.5 at max needed 260 million. Argon generates about 1.8 times Astra’s output to reach the same score.
At the introductory rate that still comes out cheap. Artificial Analysis puts Argon at $1.99 per index task, against $3.26 for Astra and $5.98 for Opus 5.5. Argon matches Astra’s score for 61% of the cost. The saving comes from the rate, not from efficiency: Astra charges five times as much per output token, and Argon uses fewer than twice as many.
Double the rate and the order changes. Every line of Argon’s card doubles at the standard rate (input, output and cached input), so at the same token mix the cost per task doubles exactly, to $3.98. That is 22% more than Astra for the same index score. It is still 33% less than Opus 5.5, which scores five points higher.
| Item | Introductory rate | Standard rate |
|---|---|---|
| Gemini 4 Argon (high) | $1.99 | $3.98 |
| GPT-6 Astra (max) | $3.26 | $3.26 |
| Claude Opus 5.5 (max) | $5.98 | $5.98 |
| Claude Sonnet 5.5 (max) | $7.60 | $7.60 |
The Artificial Analysis cost is one fixed workload, and your workload is not that one. Its main use is the ratio. If your prompts lean on cached context, the 95% cache discount moves Argon further ahead, because Astra’s cache read is $1.00 against Argon’s $0.10. If your tasks are output-heavy agent loops, Argon’s verbosity counts against it at either rate.
Is Gemini 4 Argon better than GPT-6 Astra?
On the independent index they are level: 53 each, Argon at high effort and Astra at max. Above them sit Claude Opus 5.5 at 58 and Claude Sonnet 5.5 at 56. Below them is GPT-6.1 Sol at 52, which costs $2 and $10 and is available to anyone with an OpenAI key today.
Google’s own launch table makes the case for Argon by benchmark. As transcribed by VentureBeat (Carl Franzen, September 30, 2026), Argon leads or ties on 13 of the 18 rows Google published. All of these figures are vendor-reported.
| Benchmark (vendor-reported) | Gemini 4 Argon | GPT-6 Astra | Claude Opus 5.5 |
|---|---|---|---|
| DeepSWE v1.1 | 77.9% | 74.1% | 74.2% |
| AutomationBench | 51.3% | 41.4% | 42.5% |
| Vals Finance Agent v2 | 65.4% | 53.5% | 58.6% |
| GraphWalks (long context) | 84.2% | 71.8% | 66.8% |
| LVBench (long video) | 91.7% | 87.5% | 83.7% |
| Harvey Legal Agent | 19.6% | 5.4% | 3.8% |
| CWE-bench v1 | 68% | 68% | 67% |
| FrontierSWE v2 | 55.0% | 65.5% | not reported |
| Terminal-Bench Science 0.1 | 57.6% | 68.1% | not reported |
Two of those rows can be checked against independent boards, and they disagree. Zapier’s own AutomationBench leaderboard confirms Argon at the top: 51.29% at high effort, ahead of Claude Sonnet 5.5 at 44.75%, Opus 5.5 at 42.47% and Astra at 41.4%. The independent DeepSWE leaderboard, last updated September 22, does not list Argon at all. Its top is a three-way tie at 74% between Astra, Gemini 3.8 Flash and Claude Opus 5. The 77.9% is Google’s number, run in a harness Google does not name.
Where Astra wins, it wins by a lot: 65.5% to 55.0% on FrontierSWE v2 and 68.1% to 57.6% on Terminal-Bench Science. Those are long, open-ended engineering and research tasks, which is the work a frontier model is bought for. Google’s strongest rows are knowledge work: legal, finance and business automation.
There is one more independent signal. The Decoder (Matthias Bastian, October 1, 2026) reports Artificial Analysis’s AA-Omniscience results: Argon’s hallucination rate is 15% against 51% for Astra at max, but its accuracy is 50% against Astra’s 63%. Argon knows less and invents less. Whether that is a better model depends on whether a wrong answer or no answer costs you more.
How does Gemini 4 Argon compare with Claude Opus 5.5?
Opus 5.5 is the stronger model on the independent index, 58 against 53, and the more expensive one per task, $5.98 against Argon’s $1.99 now and $3.98 later. At the standard rate the two have identical rate cards, $4 and $20, so the whole difference in cost per task comes from token consumption: Opus 5.5 generated 260 million output tokens on the index, more than twice Argon’s 110 million.
Google’s table gives Opus 5.5 two outright wins: Terminal-Bench 4.0 (66.4% to 57.4%) and PostTrainBench (49.3% to 45.3%). Both are agent-in-a-terminal tests, the closest of Google’s rows to day-to-day coding agent work. On the independent Vals Index, which Vals scores at Argon’s $4 and $20 standard rate, Argon ranks first of 41 at 68.90% for $15.68 per test, ahead of Sonnet 5.5 (67.04%, $21.34), Opus 5.5 (66.97%, $32.14) and Fable 5.1 (65.83%, $28.71).
So the two independent boards point in different directions. Artificial Analysis, whose index is weighted toward reasoning and agentic evaluations, puts Opus 5.5 five points clear. Vals, whose index weights finance evaluations such as Finance Agent v2 at more than half the score, puts Argon first and cheaper. Neither is wrong. They measure different work, and why the Vals Index and Artificial Analysis rank models differently walks through the weighting that causes the split.
Should you plan around Gemini 4 Argon?
Not for anything shipping this quarter, because the earliest public date is “as soon as possible”. If you need a frontier model today, the choice is still between the models you can call. For general agentic work, Claude Opus 5.5 holds the top independent score. GPT-6 Astra is the cheaper way to reach 53. GPT-6.1 Sol gets within a point of Astra at a fifth of its rate.
Plan for Argon if your work is legal, finance or business automation, the rows where both Google’s table and the independent Vals and Zapier boards put it first. Budget at $4 and $20, not $2 and $10. At that rate it lands between Astra and Opus 5.5 on cost per index task, and on Vals it is still the cheapest of the top four per test. If it opens at the introductory rate, treat the gap as a discount with no published end date, not as the price.
For live rates as each model changes, the AI model tracker carries Argon as a preview and will move it to released when Google lists it on its pricing page. The day each model launched is in the AI model release timeline, and the rest of the month’s launches are in new AI models released in September 2026.
Frequently asked questions
- How much does Gemini 4 Argon cost?
- Gemini 4 Argon costs $2 per million input tokens and $10 per million output tokens at its introductory rate, with cached input 95% off at $0.10. Google states that standard pricing of $4 and $20 per million tokens applies after the introductory period, but has not said when that period ends.
- When will Gemini 4 Argon be available?
- Google announced Gemini 4 Argon on September 30, 2026 and is rolling it out first to trusted cyber defenders through its Fairwind Program. Paid Gemini API customers and Google AI Ultra subscribers come next, which Google describes as "as soon as possible" without a date. As of October 1, 2026 it is not listed on the Gemini API pricing or models pages.
- Is Gemini 4 Argon better than GPT-6 Astra?
- They tie on the Artificial Analysis Intelligence Index at 53, Argon at high effort and Astra at max. Argon is cheaper per index task at the introductory rate, $1.99 against $3.26, but costs $3.98 at the standard rate, 22% more than Astra. Google reports Argon ahead on 13 of 18 benchmarks, while Astra leads on FrontierSWE v2 and Terminal-Bench Science.
- Is Gemini 4 Argon better than Claude Opus 5.5?
- Claude Opus 5.5 scores higher on the Artificial Analysis Intelligence Index, 58 against 53, and leads on Terminal-Bench 4.0 in the table Google published. Argon ranks first on the Vals Index at 68.90% against 66.97% for Opus 5.5, and costs less per task. At the standard rate both models charge $4 and $20 per million tokens.
- What is the Gemini 4 Argon context window?
- Google states an output limit of 1 million tokens, up from 64,000 on earlier Gemini models. Artificial Analysis lists a 1 million token context window. Google has not published a model ID or a model card with the full limits.
Sources
Google (2026). Gemini 4 Argon: our next era of frontier intelligence. Google blog announcement, September 30, 2026. https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-4-argon/ Verified 2026-10-01.
Google (2026). Gemini Developer API pricing and Gemini models. Google AI for Developers documentation; Argon not listed. https://ai.google.dev/gemini-api/docs/pricing and https://ai.google.dev/gemini-api/docs/models Checked 2026-10-01.
Artificial Analysis (2026). Gemini 4 Argon (high), GPT-6 Astra (max) and Claude Opus 5.5 model pages, and the models leaderboard. Independent benchmarking service. https://artificialanalysis.ai/models/gemini-4-argon and https://artificialanalysis.ai/leaderboards/models Read 2026-10-01.
Vals AI (2026). Gemini 4 Argon benchmarks, cost and capabilities. Independent evaluation, Vals Index. https://www.vals.ai/models/google_gemini-4-argon Read 2026-10-01.
Zapier (2026). AutomationBench leaderboard, v1.0.6. Benchmark owner’s leaderboard. https://zapier.com/benchmarks Read 2026-10-01.
Datacurve (2026). DeepSWE leaderboard, updated September 22, 2026. Independent benchmark leaderboard. https://deepswe.datacurve.ai/ Read 2026-10-01.
Franzen, C. (2026). Google unveils Gemini 4 Argon, retaking benchmark lead over OpenAI and Anthropic, but in limited release. VentureBeat, secondary transcription of Google’s benchmark charts. https://venturebeat.com/technology/google-unveils-gemini-4-argon-retaking-benchmark-lead-over-openai-and-anthropic-but-in-limited-release
Bastian, M. (2026). Google Gemini 4 Argon closes the gap with OpenAI and Anthropic but doesn’t take a clear lead. The Decoder, secondary reporting of Artificial Analysis results. https://the-decoder.com/google-gemini-4-argon-closes-the-gap-with-openai-and-anthropic-but-doesnt-take-a-clear-lead/