Gemini 4 Argon and Claude Opus 5.5 split across the reported coding tests: Opus led The New Stack’s billing and long-context trials, while both models tied on a bug hunt. In a separate 20-run experiment, Argon passed more coding runs but cost more overall. NeoTeo’s earlier comparison covers the broader matchup; this follow-up looks at what the practical tests found.
What the practical tests found
The New Stack’s October 18, 2026 comparison reports one run per model on each task. Its bug hunt used a 3,500-line Python library containing 14 planted bugs and five decoys. Both models found all 14 bugs, flagged none of the decoys, and fixed the bugs while preserving all 93 existing tests. The New Stack’s account also describes the results on billing and long-context tasks:
| Task | Gemini 4 Argon | Claude Opus 5.5 | Test conditions |
| Bug hunt | Found 14 of 14 bugs; flagged zero decoys; preserved 93 of 93 existing tests after fixes | Found 14 of 14 bugs; flagged zero decoys; preserved 93 of 93 existing tests after fixes | One run per model on a 3,500-line Python library with 14 planted bugs and five decoys |
| Billing engine | 11/80 hidden tests; then 20/100 after a currency follow-up | 80/80 hidden tests; then 100/100 after the currency follow-up | Tests included daylight-saving changes, month-end dates, and rounding; the follow-up added two currencies and 20 tests |
| Blast-radius question, 500K document | 20/33 services correct; 8 wrong | 33/33 services correct | Asked which of 33 services would be affected by a failure; the tokenizers counted the same nominal document as 549,483 Argon tokens and 784,087 Opus tokens |
| Blast-radius question, 600K document | 19/33 services correct; 16 wrong | 32/33 services correct | Asked which of the same 33 services would be affected by a failure |
| Six-change coding test | All tests passed in 16 of 20 runs | All tests passed in 13 of 20 runs | Paddo.dev’s October 14, 2026 test covered six changes from one production monorepo, with 20 runs per model |
The billing scores need one distinction: The New Stack reports that a manual one-line correction to Argon’s code brought it to 80/80 and then 100/100. Those were corrected results, not the original scores shown in the table. At 600K, The New Stack says Argon treated a dependency as removed after the fictional wiki reinstated it.
Why the results do not name an overall winner
The New Stack’s findings point in different directions even within one comparison: the models tied on bug finding and fixing, while Opus scored higher on the original billing tests and the 500K and 600K blast-radius questions. But each task ran once per model, so those outcomes describe those particular executions.
Paddo.dev’s separate test adds a different kind of evidence: 20 runs across six changes in one codebase. Argon had more runs in which every test passed, but the experiment does not repeat The New Stack’s tasks. The results answer different questions, not a shared ranking scale.
What the reported costs mean
In Paddo.dev’s 20-run experiment, the reported totals were $57.68 for Argon and $45.81 for Opus. The same account reports 1.37 million output tokens for Argon versus 495,000 for Opus. Argon’s introductory per-token rates were lower, yet its total for this workload was higher—a reminder that output volume affects the bill alongside the rate per token.
Google announced introductory Gemini 4 Argon API rates of $2 per million input tokens and $10 per million output tokens, followed by $4 and $20 after the introductory period. Google’s announcement states the rates; The New Stack reported that the introductory prices ran through November 30, 2026, with the higher rates starting December 1. These per-token prices do not determine the cost of a completed coding task on their own.
