Gemini 4 Argon Reaches Paying Developers at $2 Per Million Tokens, Trailing Rivals on Four Benchmarks
A.I. / news
Gemini 4 Argon Reaches Paying Developers at $2 Per Million Tokens, Trailing Rivals on Four Benchmarks
Figures Google released put Argon behind GPT-6 Astra or Claude Opus 5.5 on four of eleven benchmark rows, all vendor-chosen.
Google opened Gemini 4 Argon to paying developers and subscribers on Oct. 13, two weeks after announcing it, at an introductory $2 per million input tokens and $10 per million output tokens. The price doubles after Nov. 30.
Google's own benchmark figures, as reproduced by one outlet, show Argon leading on six of eleven rows, tied on one and behind on four. The four are terminal, software-engineering and post-training tests. Google chose the tests, and no independent run has been published.
What opened on Oct. 13
Sundar Pichai, Google's chief executive, wrote that Argon is "available globally today for developers and enterprises in the paid tier of the Gemini API," according to OfficeChai, which also cites Google's release notes. Enterprises can reach it through Gemini for business.
Google AI Ultra subscribers get it in Antigravity, AI Studio and Android Studio. AI Pro and Ultra subscribers get it in the Gemini app and AI Mode in Search. Android Headlines put the start of the app rollout on Oct. 14 and listed AI Pro at $20 a month and AI Ultra from $99. The two outlets differ by a day on timing, and Google's launch post of Sept. 30 said only "rolling out soon."
The API limit is 1 million tokens, and the launch post says the output ceiling rose from 64,000. It does not state a context window.
| API price per million tokens | Input | Output |
|---|---|---|
| Through Nov. 30, 2026 | $2 | $10 |
| After Nov. 30 | $4 | $20 |
Cached input is 95 percent cheaper, according to Google's launch post, written by Koray Kavukcuoglu, senior vice president of Google DeepMind and Chief AI Architect.
Where Argon trails
The table below is from Let's Data Science, which reproduced Google's released figures for Argon, GPT-6 Astra and Claude Opus 5.5. Google says Argon leads or ties on 13 of 18 categories. The outlet printed 11 of them.
| Benchmark | Argon | Best rival shown |
|---|---|---|
| DeepSWE v1.1 | 77.9% | 74.2% (Opus 5.5) |
| AutomationBench | 51.3% | 42.5% (Opus 5.5) |
| FrontierSWE v2 | 55.0% | 65.5% (Astra) |
| Terminal-Bench Science 0.1 | 57.6% | 68.1% (Astra) |
| Terminal-bench 4.0 | 57.4% | 66.4% (Opus 5.5) |
| PostTrainBench | 45.3% | 49.3% (Opus 5.5) |
- Claude Opus 5.566.4 %
- Gemini 4 Argon57.4 %
Source: Google-released figures as reproduced by Let's Data Science, accessed 2026-10-16
The widest gap is 10.5 points on Terminal-Bench Science 0.1. Let's Data Science said the comparison set was selected and run by the vendor, did not state test settings, and did not say whether competitor scores were rerun or taken from published sources.
Google's post names outside benchmark owners, including Vals AI for its Vals Index and Zapier for AutomationBench. It does not say who ran the rows.
The cyber version stays restricted
Google said it is releasing Argon "without cyber guardrails" to trusted defenders and its own teams, through the Fairwind Program. The launch post gives no eligibility terms. Security company Wiz uses Argon in its "Scan for Good" initiative, which fixes high-risk exposures in public infrastructure for free.
Google also said Argon monitors its own reasoning and actions and can halt execution. Defenders can still apply to Fairwind, and Google is taking applications from 10,000 scientists for six months of free access to Antigravity's top usage tier.
Introductory pricing ends Nov. 30. None of the sources reviewed carries an independent evaluation of Argon.
For how other labs handle outside access to their models, see Anthropic's arrangement with METR and Meta's Muse Spark Open.
Sources
More in A.I.
- 01Vals Test: Three of Six AI Agents Threatened Court Over a £350 Train RefundTold to try everything to recover a refund the rules no longer allowed, Grok 4.7 sent 187 messages in 15 simulated days while Opus 5.5 sent eight in twelve weeks.
- 02MiniCPM-V 4.7 Is Apache 2.0 With 3B Active Parameters and No Benchmark TablesOpenBMB's first mixture-of-experts vision model claims a win over Qwen3.6 35B-A3B in prose and charts, and 13 tokens a second on an iPhone, but the model card prints no scores.
- 03Gemini 4 Argon Opens to the API at $2 per Million Tokens, and Loses 5 of 19 Rows in Google's Own TableGoogle's October 13 launch post leads with a Vals Index ranking; the comparison table behind it shows Argon behind GPT-6 Astra or Opus 5.5 on both Terminal-Bench tests and on OSWorld 2.0.
- 04StarSkirmish Has Nine Models Write StarCraft Bots in One Hour, and the Best Wins 7% Against Top Human BotsGemini 4 Argon lands in a three-way tie at 37 to 39, but it ran at a lower reasoning setting than most rivals and its runs vary the most.