On September 30, Google introduced Gemini 4 Argon to a limited group of testers, backed by benchmark data showing it ahead of Claude Opus 5.5 across 14 of 18 tests. Performance claims included DeepSWE (77.9% versus 74.2%) and GraphWalks at 256K to 1 million tokens (84.2% versus 66.8%). The model became available to paid Gemini API developers two weeks later, on October 13.

Google acknowledged where Opus 5.5 maintains an edge, winning on four benchmarks including Terminal-bench 4.0 (66.4% versus 57.4%), which evaluates agentic command-line work. Pricing favors Argon through November 30, when it costs $2 per million input tokens and $10 per million output tokens—half Opus 5.5's rate. Starting December 1, Argon pricing rises to $4 and $20 respectively, matching Anthropic's offering. This six-week window gives developers time to evaluate whether Argon delivers superior performance or merely undercuts on cost.

To assess whether benchmark claims translate to real-world capability, three demanding tasks were tested against both models through OpenRouter on each vendor's infrastructure using default reasoning settings. A pre-built answer key with expanded scoring criteria graded all responses.

Test Design

  • Bug hunt – A 3,500-line Python library containing 14 intentional logic bugs and five decoys, with each model tasked to identify and fix issues
  • Billing spec – A subscription billing engine specification evaluated against 80 hidden tests covering daylight saving, month-end dates and rounding, plus a follow-up adding two currencies and 20 additional tests
  • Long context – A synthetic wiki of 400 services with scattered dependencies at three document sizes, requiring each model to answer three graph questions containing planted traps

Bug Hunt Results

Both models performed identically on this task. Each identified all 14 bugs, including four that only emerge when tracking a value across three or more files, and neither flagged any decoys. In a follow-up test, both fixed all 14 bugs without breaking any of the 93 existing tests.

Argon delivered cost savings. Review averaged 1 cent per bug found compared to Opus's 2.4 cents, calculated from per-call totals. Even at full December 1 pricing, Argon would remain cheaper due to lower token consumption.

  • Review: Argon 53,031 input, 3,964 output, $0.15, 43.0 seconds; Opus 64,572 input, 4,053 output, $0.34, 38.9 seconds
  • Fix: Argon 54,923 input, 24,292 output, $0.35, 186.6 seconds; Opus 66,465 input, 34,263 output, $0.95, 219.8 seconds

Billing Specification Results

Accuracy diverged significantly here. Opus passed all 80 hidden tests, then all 100 after the currency follow-up. Argon passed 11 of 80, then 20 of 100. The root cause was not flawed logic but a single early mistake: Argon's code passed an extra argument to Python's datetime.combine(), which raises an error. Since most of the module depends on that helper function, cascading failures resulted. Correcting that single line manually brought Argon to 80/80 and 100/100, though official scores reflect the original performance.

Argon remained cheaper during the promotional period, though this advantage disappears after November 30. However, Argon's lower token usage would still provide cost benefits even at parity pricing.

  • Build: Argon 3,907 input, 3,727 output, $0.05, 54.1 seconds; Opus 5,236 input, 7,763 output, $0.18, 63.0 seconds
  • Follow-up: Argon 8,881 input, 4,673 output, $0.06, 41.5 seconds; Opus 11,807 input, 7,267 output, $0.19, 51.9 seconds

Long Context Results

Token counting differences between vendors became apparent. The 500K document totaled 549,483 tokens on Argon but 784,087 on Opus. The 900K document reached 1.4 million tokens on Claude's tokenizer, causing Anthropic's API to reject it. A third document of approximately 600K tokens was added as the largest both models could process.

Both models answered two of three questions perfectly across all tested sizes. Performance split on the blast radius question, which requires identifying all 33 services that fail when one service goes down. Opus retrieved all 33 at 500K and 32 at 600K with no incorrect entries. Argon retrieved 20 correct and 8 wrong at 500K; 19 correct and 16 wrong at 600K; and 23 correct and 2 wrong at 900K. At 600K and 900K, Argon treated a dependency as removed after the wiki reinstated it.

Unlike the billing bug, this wasn't a small slip. Argon got the reasoning wrong at every size. This is a big red flag.

Speed differences suggest a connection to accuracy. Argon completed the long-context test rapidly without showing step-by-step reasoning, while Opus took longer due to explicit reasoning work. Token usage reflected this: Opus consumed over 5,000 tokens, mostly for reasoning, versus Argon's 243-425. While this pattern likely connects to Argon's errors, causation cannot be established from a single run.

Cost savings on this task do not justify use. Argon's performance was substantially inaccurate and careless.

  • 500K: Argon 549,483 input, 264 output, $1.10, 24.2 seconds; Opus 784,087 input, 5,532 output, $3.25, 68.3 seconds
  • 600K: Argon 652,492 input, 425 output, $1.31, 32.4 seconds; Opus 935,332 input, 5,158 output, $3.84, 69.1 seconds
  • 900K: Argon 982,367 input, 243 output, $1.97, 41.5 seconds; Opus not run

Summary

Testing outcomes diverged from Google's published benchmarks. Argon matched Opus on code review and fixing at less than half the cost. It failed the specification test due to one problematic line of code, and it missed or invented services on every long-context blast-radius question.

Based on this evaluation, Argon does not merit adoption. While it delivered cost advantages and matched Opus on the bug hunt, experience across the other tests revealed reliability concerns. Argon cannot be trusted without reservation on any task.

I wouldn't unquestioningly trust Argon on any task.

Opus 5.5 ranks among Anthropic's flagship models, holding position 58 on Artificial Analysis's independent index, ahead of Argon's ranking of 53. Google continues working to close the gap, but has not yet achieved parity.