TL;DR. Since October 13, Google's new frontier model Gemini 4 Argon has been open to every paying developer and business. In the independent Vals Index, which tests tasks from finance, law, tax and software development, it leads by a narrow margin and costs half as much per test run as Claude Opus 5.5. A second independent comparison, however, puts it only in third place. An introductory price applies until November 30; after that, Argon costs as much as Opus 5.5. A good moment to test it, but no reason to switch right away.
What Google has released
On September 30, Google first opened Argon to a small circle of security specialists. Since October 13, the model has been available through the paid Gemini API, Gemini for business and the Antigravity development environment, and in the Gemini app for Google AI Pro and Ultra subscribers. According to the model card, it processes text, images, audio and video, handles up to one million tokens of input and output combined, and its knowledge runs until March 2026.
The most important number does not come from Google. Vals AI measures models on work tasks from finance, law, tax and software development, weighted by their share of the economy. In the Vals Index of October 13, Argon ranks first. It is not only slightly more accurate there, but also clearly cheaper and faster.
Vals Index, as of October 13, 2026
Argon needed 46 minutes for the full run, Sonnet 5.5 and Opus 5.5 about 78 minutes each. Google's own table also highlights a field that matters to many companies: long documents. On search tasks across 256,000 to one million tokens, Argon reaches 84.2%, Opus 5.5 66.8%. In Google's measurements it also leads on finance and legal tasks.
The fine print
- The lead is narrow. Argon is 1.86 points ahead of Sonnet 5.5. Vals itself states an uncertainty of ±0.97 points for Argon's score.
- A different test, a different order. In the Artificial Analysis Intelligence Index, a second independent comparison with different tasks, Argon scores 53 points, Opus 5.5 58 and Sonnet 5.5 56.
- In the terminal, Argon falls behind. On agentic programming at the command line (Terminal-Bench 4.0), it reaches 57.4% according to Google's own model card, Opus 5.5 66.4%.
- Much of it is self-reported. The figures from Google's own operations, such as 35% more code submitted by agents and almost a third fewer rollbacks, have not been checked by anyone outside.
Two independent comparisons, two rankings
The price: half until November, then on par with Opus
According to the pricing page, Argon costs $2 per million input tokens and $10 per million output tokens until November 30. That is the price of Claude Sonnet 5.5. From December 1, both double to $4 and $20, the level of Opus 5.5. Cached input costs 95% less, batch jobs half. There is no free tier.
Output price per 1M tokens
The price per token alone says little. What matters is the cost of a completed task. At the same token price as Opus 5.5, a complete Vals run with Argon cost $15.68 instead of $32.14, roughly half. Whether that holds for your tasks only a test with your own cases can show.
What this means for a mid-sized company
- Three providers on a par. Google, Anthropic and OpenAI are a few points apart in the independent comparisons, and the order changes every few weeks. Anyone who ties themselves firmly to one provider gives away the price pressure.
- The strengths fit typical tasks. Long contracts, manuals and tenders, finance and legal questions: this is where Argon leads in the measurements. For programming agents at the command line, Opus 5.5 remains stronger according to Google's own table.
- The data path decides. Through Google Cloud (Vertex AI), Argon can run in the EU region; processing within Germany only is not yet listed for Argon there (as of October 13). According to the pricing page, Google does not use content from paid usage to improve its products.
What you should do now
- Test before November 30. Take 20 to 50 real cases from your daily work and run Argon next to your current model. Count hits, time and cost. At the introductory price, the test is cheap.
- Calculate with the December price. Base every calculation on $4 and $20 per million tokens, not on the introductory price.
- Clarify the data path first. Confidential or personal data belongs in the EU region via Vertex AI, not on a global endpoint.
- Keep models interchangeable. If prompts, test cases and integration are built so that switching models is a setting, you benefit from every price move without rebuilding.
Every few weeks a different model is on top. Which one is right for your tasks is not answered by a leaderboard, but by a test with your own cases. If you want to set up such a comparison properly once, we can help you put the test cases together.
