Gemini 4 Argon Opens to the API at $2 per Million Tokens, and Loses 5 of 19 Rows in Google's Own Table
A.I. / news
Gemini 4 Argon Opens to the API at $2 per Million Tokens, and Loses 5 of 19 Rows in Google's Own Table
Google'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.
Google made Gemini 4 Argon available to developers on October 13 through the Gemini API, AI Studio and Gemini Platform, two weeks after a first announcement limited to cybersecurity partners. The launch post is written by Koray Kavukcuoglu, who is senior vice president at Google DeepMind and Google's chief AI architect, and calls Argon "our latest frontier model." The post is the primary source for everything below unless another publisher is named.
Price, access and context
Introductory API pricing is $2 per million input tokens and $10 per million output tokens through November 30, 2026. After that the public model costs $4 and $20. Cached input tokens are 95 percent cheaper. Context is a 1 million token limit shared between input and output.
Google AI Ultra subscribers get Argon in Antigravity, and Ultra and Pro subscribers get it in the Gemini app and AI Mode in Search. Google is also offering 10,000 scientists six months of free access to Antigravity's highest usage tier, subject to eligibility and selection. Approved cybersecurity defenders in the Fairwind Program can get access without cyber guardrails.
What Google's own table shows
The post says Argon ranks first on the Vals Index, which Google describes as measuring finance, coding, legal and tax work. Online Tech Tips reproduced the 19-row comparison Google published against GPT-6 Astra, Claude Fable 5.1 and Claude Opus 5.5. Every score below is vendor-supplied; Online Tech Tips says the non-Gemini scores mostly come from provider reports or public leaderboards and that it did not verify them.
Argon has the top score on 13 rows, ties GPT-6 Astra on CWE-bench v1 at 68.0 percent, and trails on five.
| Benchmark where Argon trails | Argon | Best score and holder |
|---|---|---|
| FrontierSWE v2 | 55.0% | 65.5% GPT-6 Astra |
| Terminal-Bench 4.0 | 57.4% | 66.4% Opus 5.5 |
| PostTrainBench | 45.3% | 49.3% Opus 5.5 |
| Terminal-Bench Science 0.1 | 57.6% | 68.1% GPT-6 Astra |
| OSWorld 2.0 (offline partial) | 69.2% | 72.6% GPT-6 Astra |
The launch post names no benchmark where Argon trails, and most comparisons say only that it outperforms other frontier models. The largest gap in Argon's favour is on Harvey's Legal Agent Benchmark at 19.6 percent against 6.7 percent for Fable 5.1.
- Gemini 4 Argon68.9 %
- Claude Opus 5.567 %
- Claude Fable 5.165.8 %
- GPT-6 Astra63.1 %
Source: Google's Gemini 4 Argon comparison table as reproduced by Online Tech Tips, accessed 2026-10-18
The lead on the ranking Google chose to headline is 1.9 points over Opus 5.5. The deficits on terminal tasks run up to 10.5 points on Terminal-Bench Science 0.1 and 9.0 on Terminal-Bench 4.0, which are the agentic coding conditions that our DeepSeek coverage also tracked.
Safety claims and the delay behind the launch
Google says Argon leads the Gray Swan indirect prompt injection benchmark, with a 0.7 percent attack success rate across 15 adaptive attempts, and that external bio red teaming unlocked less than 1 percent of the sub-tasks tested. It also describes an automated monitor that reads the agent's reasoning and actions to block unintended workarounds. These are Google's figures.
The launch followed delays. Khaleej Times reported on October 1 that Google no longer plans to release Gemini 3.5 Pro, which Chief Executive Sundar Pichai had said in June would ship, and that Demis Hassabis is stepping aside as DeepMind's chief executive in an overhaul. StarSkirmish Bench is the site's earlier look at how Google's models fare in an agentic test.
The next fixed date is November 30, when the introductory price doubles.
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.
- 03StarSkirmish 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.
- 04Statistics and Optimization Make Up 20% of arXiv Theorems but Under 1% of OpenAI's 722 Math PapersA newsletter analysis of OpenAI's October 6 release finds the fields closest to machine learning nearly absent, and OpenAI has not said why.