Gemini 4 Argon’s score on the Bug Hunt Bench changed substantially depending on the CLI harness used, according to benchmark results posted by Paweł Huryn on October 14, 2026. Through Antigravity CLI, Argon scored 40.7 out of 105 planted bugs—just 0.7 points behind Opus 5.5 in its max configuration. The same model scored 33 through Gemini CLI.
Harness choice changed the ranking
Bug Hunt Bench tests models against 105 bugs planted across two repositories. The benchmark focuses on issues that frontier models reportedly missed during the first quarter of 2026. Unplanted bugs are tracked separately and do not affect the main score.
The leaderboard, updated October 14, lists the following results:
| Model | Configuration | Score |
|---|---|---|
| GPT-6 Astra | max | 45/105 |
| Fable 5.1 | max | 43/105 |
| Gemini 4 Argon | Antigravity CLI | 40.7/105 |
| Opus 5.5 | max | 41.7/105 |
| Opus 5.5 | xhigh | 36/105 |
| Muse Spark 1.3 | max | 32.2/105 |
| Gemini 4 Argon | Gemini CLI | 33/105 |
| Grok 4.7 | xhigh | 29/105 |
| Gemini 3.8 Flash | Antigravity CLI | 18/105 |
The leaderboard’s ordering places Opus 5.5 slightly ahead of Argon when both use their highest listed configurations. Huryn described the Antigravity result as moving Argon from “good” to a frontier model, while also noting that the benchmark was run three times to verify the outcome.
Different costs and runtimes
The benchmark records an estimated runtime of 56.5 minutes and a listed cost of $7.16 for Argon through Antigravity CLI. Through Gemini CLI, the estimated runtime rises to 85.1 minutes, with a listed cost of $32.08.
Huryn separately reported costs of $5.74, $6.43 and $9.30 for three Antigravity CLI runs. The benchmark’s displayed prices are list-rate estimates unless otherwise noted, so those figures do not necessarily match every individual run.
According to Huryn, the lower Antigravity costs came from fewer tokens and more cache hits. The test runs used isolated folders, empty sessions and simple prompts. Depending on the harness implementation, runs could share only the system prompt.
Argon was initially unavailable in Antigravity CLI, Huryn mentions, so the first attempt used a workaround. Gemini 3.8 Flash was tested through Antigravity CLI during that initial attempt.
The results suggest that the interface and harness can materially affect model benchmark outcomes. They do not establish that Argon’s underlying capabilities changed between tests, and the benchmark’s own historical uncertainty was described by Huryn as roughly two to three points.
Benchmark data is available on GitHub through the benchmark website.
Source: Paweł Huryn on X
