Gemini 4 Argon vs. Claude Opus 5.5: Independent Tests Show a Mixed Picture

Google announced Gemini 4 Argon to a small group of testers on September 30, alongside a benchmark table that put Argon ahead of Claude Opus 5.5 on 14 of 18 tests, including DeepSWE (77.9% vs. 74.2%) and GraphWalks at 256K to 1 million tokens (84.2% vs. 66.8%). On October 13, Google opened Argon to paid Gemini API developers.

Pricing Window

Through November 30, Argon costs $2 per million input tokens and $10 per million output tokens, half the price of Opus 5.5. On December 1, the price rises to $4 and $20, matching Opus. Google also published where Opus leads, including Terminal-bench 4.0 (66.4% vs. 57.4%), which tests agentic work in a command line. The author noted that benchmarks do not always match real work, and a half price window only matters if the model matches the competitor on quality.

The author ran each task once per model through OpenRouter on the vendors' own servers, using default reasoning settings. These are single-run results, not repeated accuracy estimates.

Code Tests

On a bug hunt in a 3,500-line Python library with 14 planted logic bugs and five decoys, both models found all 14 bugs and flagged no decoys, then fixed them while keeping all 93 existing tests passing. Argon averaged about 1 cent per bug found versus 2.4 cents for Opus. On a billing spec graded by 80 hidden tests, Opus passed all 80 and all 100 after a currency follow-up, while Argon passed 11 of 80 and 20 of 100. Argon's failure traced to one line passing an extra argument to Python's datetime.combine() function; corrected manually, Argon then scored 80/80 and 100/100.

Long Context and Verdict

Using a fake wiki of 400 services, both models answered two of three graph questions perfectly at every size each actually ran. The 900K document exceeded one million tokens under Claude's tokenizer and was rejected by Anthropic's API, so Opus was not run at 900K. The nominal document sizes do not represent identical token counts across the models. The split came on the blast radius question, which asks for all 33 services that go down when one fails. Opus got all 33 at 500K tokens and 32 at 600K with no wrong entries, while Argon got 20 right with 8 wrong at 500K, 19 right with 16 wrong at 600K, and 23 right with 2 wrong at 900K. Argon also used only 243 to 425 output tokens versus over 5,000 for Opus, racing through answers without visible reasoning, while the author described Argon's reasoning errors as a big red flag and said one run could not prove a causal link with its short output.

The author said the results did not match Google's benchmarks and would not unquestioningly use Argon on any task. Artificial Analysis ranks Opus 5.5 at 58 versus Argon at 53. Across the tests each model actually ran, Argon cost $4.98 in 423.3 seconds; Opus cost $8.75 in 511.0 seconds.

Sources: The New Stack