Argon beat Opus on 14 Google benchmarks and lost a hand test

A vendor benchmark table is not the same thing as a model that works on your job. The New Stack writer Jessica Wachtel put Gemini 4 Argon and Claude Opus 5.5 through three tests of her own and found a different pattern from the one Google published, with Argon cheaper by a wide margin and wrong on a long-context question at every size she tried.

Google announced Argon on 30 September to a small group of testers with a benchmark table that put it ahead of Opus 5.5 on 14 of 18 tests, including DeepSWE at 77.9 against 74.2 percent and GraphWalks between 256K and one million tokens at 84.2 against 66.8 percent. Opus won the other four, including Terminal-bench 4.0 at 66.4 against 57.4. Argon opened to paid Gemini API developers on 13 October. Google priced it at $2 per million input tokens and $10 per million output through 30 November, half the Opus rate, then $4 and $20 from 1 December, matching Opus.

The tests

Wachtel ran each test once per model through OpenRouter, on Google's and Anthropic's own servers, at default reasoning settings, scored against a key she built and verified first.

The bug hunt was a 3,500-line Python library with 14 planted logic bugs and five decoys. Both models found all 14, flagged no decoy, and fixed every bug without breaking the 93 existing tests. Argon cost one cent per bug found against Opus's 2.4 cents.

The billing specification asked for a subscription billing engine graded by 80 hidden tests on daylight saving, month-end dates and rounding, with a follow-up adding two currencies and 20 more tests. Opus passed all 80, then all 100. Argon passed 11 of 80, then 20 of 100. Wachtel traced Argon's failures to a single line that passed an extra argument to Python's datetime.combine(), which raises an error that cascades through the module. She corrected that one line by hand and Argon then scored 80 of 80 and 100 of 100. The official scores stand as graded.

The long-context test fed each model a 400-service fake wiki with scattered dependencies, at three sizes, and asked a blast-radius question whose correct answer was 33 services. Opus got all 33 at 500K and 32 at 600K, with no wrong entries. Argon got 20 right and 8 wrong at 500K, 19 right and 16 wrong at 600K, 23 right and 2 wrong at 982K. At the two larger sizes it treated a dependency as removed after the wiki reinstated it. Argon used 243 to 425 output tokens on each run, with no shown reasoning. Opus used more than 5,000.

The totals

Across all three tests, Wachtel's run cost $4.98 for Argon and $8.75 for Opus, with Argon writing 37,588 output tokens against Opus's 64,036. Argon finished faster at 423 seconds against 511. On accuracy, Opus scored at the top of each test and Argon tied on the bug hunt only. Artificial Analysis's own intelligence index, which Wachtel also cites, puts Opus at 58 and Argon at 53.

Her conclusion, in her own words: she would not use Argon, and would not trust it on any task without checking. A reader thinking about moving a working Opus workload onto Argon for the price break would want the same side-by-side on their own code before the November window closes.

Source

• The New Stack, Jessica Wachtel