Gemini 4 Argon vs. Claude Opus 5.5: Half the price and about half as accurate
The New Stack Jessica Wachtel
Google’s Gemini 4 Argon is out for paid API developers, and it’s half the price of Claude Opus 5.5 for now. But a head-to-head test found it can be much cheaper without being reliably as good.
Based on reporting by The New Stack, Jessica Wachtel — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Google opened Gemini 4 Argon to paid Gemini API developers on October 13, after first showing it to a small tester group on September 30. The company’s own benchmark table painted a strong picture: Argon came out ahead of Claude Opus 5.5 on 14 of 18 tests, including DeepSWE and GraphWalks at 256K to 1 million tokens. Google also showed where Anthropic’s model still wins. Opus led on the other four benchmarks, including Terminal-bench 4.0, which checks agentic command-line work.
For the next few weeks, the price is the most obvious difference. Until November 30, Argon costs $2 per million input tokens and $10 per million output tokens, which is half of Opus 5.5’s price. On December 1, Argon moves up to $4 and $20, matching Opus. So developers have about six weeks to decide whether they’re getting a bargain or just a temporary discount.
I ran both models through three hard tests and kept the setup simple: one run per model, through OpenRouter, on Google’s and Anthropic’s own servers, using default reasoning settings. The first test, a bug hunt in a 3,500-line Python library with 14 planted logic bugs and five decoys, ended in a tie. Both models found all 14 bugs, avoided the decoys, and fixed the code without breaking the 93 existing tests. Argon did it for less money, though, using fewer tokens in both review and repair.
The billing spec test is where the gap widened. Opus passed all 80 hidden tests, then all 100 after a currency follow-up. Argon passed 11 of 80, then 20 of 100. That failure turned out to come from one bad call to Python’s datetime.combine(), which broke much of the module. When that line was fixed manually, Argon scored 80/80 and 100/100 too, but the graded result still says something about how brittle the run was.
The long-context test was worse. Both models handled two of the three questions perfectly at every size, but the blast-radius question exposed Argon’s weakness. Opus got all 33 services right at 500K tokens and 32 at 600K, with no wrong entries. Argon missed or invented services at every size, and it did so while using far fewer reasoning tokens. That made it fast. It also made it sloppy. In the author’s own words, this was a big red flag.
My take — AI-written commentary, not fact-checked reporting
Cheap is not the same thing as good, and this test set makes that painfully clear. Google can win benchmark charts all day; if the model falls apart on long-context reasoning, the bargain price is just a nicer wrapper on uncertainty. The AI market keeps trying to sell “almost there” as a product, and that joke is getting old.
Read more about this at: The New Stack