Gemini 4 Argon faces the tests that really matter

Google presents Gemini 4 Argon as faster and cheaper than Claude Opus 5.5. But practical tests show a more nuanced picture, especially on long code and reasoning tasks.

Gemini 4 Argon, Google's model open to paying developers since October 13, 2026, promised better value for money compared to Claude Opus 5.5. However, independent tests published on October 18 show that it does not perform well on all tasks, even if it can compete on some coding tests. This topic matters because the gaps between benchmarks and real-world use directly impact product teams' choices. And on the bill, the difference is not trivial.

When the price drops, trust does not always follow

Google already highlighted this commercial argument at launch: Gemini 4 Argon costs 2 dollars per million input tokens and 10 dollars per million output tokens until November 30, then 4 dollars and 20 dollars from December 1, the same level as Claude Opus 5.5. According to the test published by Jessica Wachtel, Argon outperforms Opus on 14 out of 18 benchmarks announced by Google, but loses on four others, including Terminal-bench 4.0, which evaluates agentic work in command line. In short, the price gap is not enough to overshadow the behavioral differences.

Tests conducted via OpenRouter on the servers of both providers covered three exercises: bug hunting in a 3,500-line Python library, a billing specification with hidden tests, and a fake wiki of 400 services to measure reasoning over long context. Result: Argon and Opus were equal on bug hunting, but the comparison quickly became tense when more stable reasoning was required. Google's model sometimes had the speed of a very confident intern. That is not always reassuring.

A good code review is not enough

On bug hunting, both models found the 14 errors introduced in the code, with no false positives, then fixed them without breaking the 93 existing tests. Argon was even less costly: $0.15 for the review versus $0.34 for Opus, then $0.35 for the fix versus $0.95. Jessica Wachtel also notes that Argon used fewer tokens, helping it remain cheaper even at the full rate planned for December.

But the picture changes on the billing specification. There, Opus passed 80 out of 80 hidden tests, then 100 out of 100 after adding a second currency. Argon only passed 11 out of 80, then 20 out of 100. The initial cause was a single error in the call to datetime.combine(), but the domino effect crippled the entire module. The model was therefore not "bad" everywhere; it mainly broke the mechanism in the wrong place.

The long context, the real test?

It is on the fake wiki of 400 services that Argon falls most clearly. Opus identified the 33 services affected by an outage on datasets of 500K and 600K tokens, with zero errors on the first test and only one on the second. Argon found only 20 correct services and 8 false positives on the first set, then 19 correct and 16 false positives on the second. On a third corpus of about 900K tokens, it still responded incorrectly, while Anthropic refused to run theirs beyond one million tokens.

Jessica Wachtel also points out that Google's model responded very quickly, with little visible reasoning, while Opus took more time to structure its response. The numbers support this: on the 500K token test, Argon consumed 549,483 input tokens and 264 output tokens for $1.10, versus 784,087 input tokens and 5,532 output tokens for $3.25 with Opus. At 600K tokens, Argon remained cheaper, but its answers stayed inaccurate. The speed gain thus has a hidden cost: reliability.

In the end, the debate is not only about raw power but about how a model handles everyday technical tasks: code review, business logic, long memory, robustness. Argon can be cheap and convincing on a targeted exercise without becoming a generally trusted option for teams wanting to avoid surprises in production. That is where the nice benchmark curve stops ruling.

Key points

  • Google opened Gemini 4 Argon to paying developers on October 13, 2026.
  • Argon costs half as much as Opus until November 30.
  • Both models are equal on bug hunting, with no false positives.
  • Opus passes 80/80 hidden tests, Argon 11/80 on billing.
  • On long context, Argon produces many false positives.
  • Jessica Wachtel judges she would not use it without reservations.

By the numbers

  • 14 out of 18 — benchmarks won by Gemini 4 Argon according to Google at launch.
  • $2 / $10 — price per million input/output tokens until November 30, 2026.
  • 605 deaths — agents cited in the other standalone Minecraft demonstration, according to the source provided.
  • 33 services — correct blast radius on the long context test, achieved by Claude Opus 5.5.
  • 2,783 JavaScript scripts — total written in the autonomous Minecraft session mentioned in another source.

To read