Gemini 4 Argon Goes Wide at $2 and $10 Per Million Tokens Until Nov. 30, Then Doubles
A.I. / news
Gemini 4 Argon Goes Wide at $2 and $10 Per Million Tokens Until Nov. 30, Then Doubles
Google's own table has Argon ahead on seven rows it chose. Artificial Analysis, running its own index, puts the model level with GPT-6 Astra and below Claude Opus 5.5.
Google opened Gemini 4 Argon to developers, enterprises and paying Gemini subscribers on Oct. 13, 13 days after announcing it, at $2 per million input tokens and $10 per million output tokens. The price applies only until Nov. 30, 2026. After that the public-access model costs $4 and $20, according to Google's launch post, which Koray Kavukcuoglu, Google DeepMind's senior vice president and chief AI architect, wrote.
The model has been in limited hands since Sept. 30, when Google restricted it to vetted cyber defenders in its Fairwind Program. That program is expanding to hundreds more partners, Google said, and those users get the model without cyber guardrails.
Who can use Argon, and where
The Gemini API paid tier in AI Studio and Android Studio is the developer route. Google AI Ultra subscribers get Argon in Antigravity, Google's agentic coding tool. Ultra and Pro subscribers get it in the Gemini app and in AI Mode in Search.
Google also offered 10,000 scientists Antigravity's top usage tier free for six months, subject to capacity.
The API exposes a 1 million token limit that Google describes as unified across input and output. It lists no separate input context window. Earlier output was capped at 64,000 tokens, according to eesel AI's review of the launch, which found that Google published no batch, flex or free-tier prices.
| Tier | Input per 1M tokens | Output per 1M tokens |
|---|---|---|
| Introductory, through Nov. 30, 2026 | $2 | $10 |
| After Nov. 30, 2026 | $4 | $20 |
Cached input tokens are 95% cheaper than the listed input price, Google said.
What Google's table says and what Artificial Analysis found
Google ranks Argon first on the Vals Index, which covers finance, coding, legal and tax work, and reports 87% on ERQA, an embodied-reasoning test. Those are vendor-supplied figures. NeuralTrust, an AI security vendor, compiled Google's comparison table from DataCamp. No third party had reproduced it at launch, NeuralTrust said.
The table is mixed. Argon scores 77.9% on DeepSWE v1.1 against 74.2% for Claude Opus 5.5, but 57.4% on Terminal-Bench 4.0 against 66.4% for Opus 5.5, according to NeuralTrust's copy of Google's figures.
Artificial Analysis, which had pre-release access, ran its own tests. eesel AI reports an Intelligence Index of 53 for Argon, tied with GPT-6 Astra and behind Claude Opus 5.5 at 58. NeuralTrust gives the same index as 52.6, 52.7 and 57.6. Artificial Analysis tested only Argon's High setting, NeuralTrust said.
On Terminal-Bench 4.0, Artificial Analysis measured Argon at 57.1%, fourth place. The gap to Opus 5.5 is about 9 points in Google's table and about 6 in Artificial Analysis's run, according to eesel AI.
- Gemini 4 Argon1.99 USD
- GPT-6 Astra3.26 USD
- Claude Opus 5.55.98 USD
Source: NeuralTrust, citing Artificial Analysis, accessed 2026-10-13
The $1.99 figure uses the introductory price. eesel AI puts the same task at $3.98 once the standard price applies, which is above GPT-6 Astra's $3.26. Argon used 62,000 output tokens per task against 27,000 for Astra, so its cost edge comes from lower token prices, not from using fewer tokens.
The hallucination number cuts both ways
On AA-Omniscience, Artificial Analysis recorded a 15% hallucination rate for Argon against 51% for GPT-6 Astra and 54% for GPT-6.1 Sol. Accuracy was 50% for Argon against 63% for Astra, eesel AI reported. The model abstains more often, which lowers the error rate and the accuracy together.
Google's own post says attack success on Gray Swan's indirect prompt-injection benchmark fell to 0.7% across 15 adaptive attempts. That is also a vendor-supplied figure, and NeuralTrust said Google has published no score table for it. How far to trust agents with long tasks is the argument in Rich Hickey's case that LLMs are automation, not tools.
What is still unknown
Google has not said how reasoning tokens are billed. It lists no input context window. The price after Nov. 30 applies to the public-access model, and Google's post does not give a figure for Fairwind partners.
The next fixed date is Nov. 30, when the $2 and $10 rates end. At $4 and $20 Argon costs the same per token as Claude Opus 5.5, according to NeuralTrust, so the comparison will rest on independent runs rather than on Google's table. Google DeepMind's other recent release, EmbeddingGemma 2, is a separate open model.
Sources
More in A.I.
- 01Google's Gemini 4 Argon Robotics Post Reports 87% on ERQA, and Most Demos Run in SimulationGoogle's developer blog lists a dozen robot demos for Gemini 4 Argon. It does not say which hardware runs were teleoperated, and gives no per-token price for the robotics model.
- 02Google Says Penalizing Argon's Reasoning Cut Open Shortcuts 61% and Raised Hidden Ones From 11.2% to 36.4%Google's safety post for Gemini 4 Argon explains why it does not train against the model's chain of thought, and lists three ways the model gamed its own tests.
- 03OpenAI Posts 719 Math Manuscripts From an Unreleased Model and Withholds the PromptsThe Oct. 6 release arrived a week after an Institute for Advanced Study advisory group asked labs to disclose the model, prompts and compute.
- 04Paper Hides Encrypted Network in Model Weights That Fine-Tuning Could Not ElicitOxford-affiliated authors report a chemistry classifier locked behind a hash-based trigger scored 50.9 percent after 500 fine-tuning steps, while a password-locked baseline reached 92.8 percent.