A 71.9 out of 100 bug-hunting score and a 7.19 averaged column are the headline numbers from one reviewer's Gemini 4 Argon coding benchmarks. Those results placed the model outside the top five on that leaderboard, below older models the same tester had already scored.
What the Gemini 4 Argon coding benchmarks actually measured
The Gemini 4 Argon coding benchmarks in this review covered 23 prompts across several project types, each run five times, plus separate bug-finding and code-generation tasks. The tester scored implementation work on a leaderboard built for short, well-planned tasks of roughly 2 to 10 minutes rather than long agentic runs.
The review's author runs a coding-model leaderboard and published this evaluation shortly after the model became publicly available in October 2026. Each prompt was attempted five times so that repeated failures on the same test would stand out from one-off errors.
The benchmark separates three kinds of work: implementing features in existing projects, finding bugs in existing code, and generating new code whose quality a separate model judged. That split matters because a model can look fine on feature work and weak on edge cases.
One scope limit is explicit in the review. All tasks are short implementation jobs, so a low or average score here says little about long-running orchestration or multi-hour agent sessions, which the reviewer says needs a separate benchmark.
How do you get access to Gemini 4 Argon?
Gemini 4 Argon launched with restricted access, initially limited to Ultra plan subscribers inside Google's Antigravity environment rather than all Gemini subscribers. The reviewer reports that Google AI Pro documentation appeared to include the model while Antigravity access for Pro users did not, a mismatch that frustrated commenters.
The reviewer kept a Pro subscription specifically to test the model and could not run it there. Access then widened through third-party providers, including OpenRouter and the Cognition coding agent, which the reviewer used because API pricing was too high to sustain.
Google's own Gemini API documentation is the place to check current model availability and quota terms. Availability changed within days during the review period, and plan limits are a moving target.
The transcript's account of a named Google executive announcing the restriction on social media could not be verified against a primary source, so it is left out here.
Gemini 4 Argon cost: API pricing and plan quotas
Gemini 4 Argon API costs dominated the review, with single high-effort prompts billed at roughly 80 to 93 cents on OpenRouter according to the tester's own terminal logs. That per-prompt figure made a full multi-run benchmark unaffordable on metered API pricing.
The tester reports three effort levels for the model, low, medium and high, and benchmarked the high setting. Higher effort means more reasoning tokens, which is the likely driver of the per-prompt cost, though the reviewer does not break down token counts.
Subscription quotas brought their own ceiling. The reviewer used nearly 100 percent of a daily quota within 24 hours on the Cognition agent and hit the limit mid-run, forcing some prompts to restart. Google's published Gemini pricing page lists current rates and is the correct place to confirm numbers before budgeting.
Rumors of Ultra subscribers exhausting five-hour limits in 30 minutes circulated on social media during launch week. The reviewer quotes one such post but had no way to verify it.
The 23-prompt results in detail
Across the main prompt set, some projects passed cleanly and others failed consistently, and speed was slow enough that the tester raised the per-prompt timeout from 15 to 30 minutes. The pattern of identical failures suggests a missing training case rather than random error.
The Flutter and Dart project passed all tests on every attempt. A CSV import task in PHP failed exactly one test on all five runs: the model never handled the case of a user uploading a file that is not a CSV. The reviewer says this edge-case gap is what separates average models from the ones that score perfectly on the same evaluation.
Bug finding produced 16 of a possible 21 bugs, judged by a separate model, but only 11 of the 16 most important ones. That translated to 71.9 out of 100 on the bug-finding test, or 7.19 on the reviewer's ten-point column.
Code generation in Laravel, React and TypeScript was judged by the same model the reviewer uses across all leaderboard entries, to keep judging consistent. The reviewer notes the judge model is older but argues consistency matters more than using the newest judge for one entry.
The table below summarises the reported outcomes.
| Task type | Result reported | Approach |
|---|---|---|
| Feature implementation, 23 prompts | Mid-table | Five runs per prompt |
| Flutter and Dart project | All tests passed | Repeated runs |
| PHP CSV import project | One test failed every run | Missing non-CSV upload case |
| Bug finding | 16 of 21 found, 11 critical | Judged by separate model |
| Bug-finding score | 71.9 out of 100 | Converted to 7.19 |
| Code generation | Average quality | Judged by GPT 5.6 Soul |
Where Gemini 4 Argon landed on the leaderboard
Gemini 4 Argon at high effort placed outside the top five on the reviewer's leaderboard, with an older mid-tier model scoring higher on the same set. The reviewer calls that result ironic given the launch attention, and places the model as average for coding.
The reviewer frames the outcome as Google rejoining the race as an average coding model rather than a leader. He suggests the model may be aimed at tasks other than coding, which would explain a mid-table coding score without implying the model is weak overall.
A second reviewer's bug-fixing test drew a similar conclusion, and the creator of a known codebase analysis benchmark reported the model as disappointing. Those are separate evaluations by different people with different setups, and none of them is an independent controlled study of the model.
The leaderboard's own caveat limits how far the result travels. It measures short implementation tasks, so a strong score would not prove good long-running behavior either.
Should you switch to Gemini 4 Argon for coding?
For short implementation tasks at high effort, the reported evidence does not support switching from established top models, and per-prompt API pricing makes casual use expensive. Subscription access through a coding agent is the practical route if you want to test it yourself.
The reviewer's own conclusion is that many people who tried the model will return to whichever model they already use. He names older flagship models from Anthropic OpenAI as the ones people are likely to keep, and calls the new model worth waiting on.
Google's official material reportedly ends with a line saying the model is just the beginning of the Gemini 4 family. That signals more releases in the family without committing to dates, capability levels or pricing, so treat future models as unannounced until Google publishes them.
A fair test on your own codebase matters more than any single leaderboard. If your work is short feature tasks, run a handful of your own prompts at more than one effort level and compare cost per finished task, not cost per token.
What Gemini 4 Argon coding benchmarks leave out
These benchmarks measure short, planned implementation work and say little about long-running agent sessions, tool orchestration, or repository-scale refactors. The reviewer is explicit that orchestration needs its own test design and that this set does not cover it.
Cost comparisons are also limited. Per-prompt figures depend on effort level, output length and retries, and the reviewer's numbers come from one account and one provider's rates. Treat them as an illustration of magnitude rather than a price list.
Model access changed within days during the review window, so any plan restriction quoted here may have shifted since October 2026. Check the provider's current documentation before assuming a model is included in a given subscription.
FAQ
- How did Gemini 4 Argon score on coding benchmarks? In one reviewer's October 2026 evaluation of 23 prompts run five times each, the model placed outside the top five and scored 71.9 out of 100 on a separate bug-finding task. The reviewer described the overall result as average for coding.
- Is Gemini 4 Argon expensive to use through the API? The tester logged roughly 80 to 93 cents per high-effort prompt on OpenRouter, which made a full benchmark run impractical on metered pricing. Rates vary by provider, effort level and output length, so check current published pricing.
- Can Pro subscribers use Gemini 4 Argon? At launch the reviewer reports it was limited to Ultra plan subscribers inside Antigravity, despite Pro documentation appearing to include it. Access widened to third-party providers within days, so check current plan terms.
- Does a mid-table score mean the model is bad at all coding? No. The benchmark covers short, well-planned implementation tasks of roughly 2 to 10 minutes, and the reviewer says it does not test long-running agent orchestration. A low score here says little about other workloads.
- Will there be more Gemini 4 models? Google's launch material reportedly says the model is just the beginning of the Gemini 4 family. No dates, names or capability levels for later models have been published.
Fork this article
Start a new branch from the same video, shaped your way. You keep the credit; the original keeps the attribution.
A fork in another language is filed as a translation of this article, so the two pages point at each other. You can unlink it later from the editor.
0/240
You are creating
- Format
- For
- Language
- Source
- Your angle
You will be asked to sign in before it is generated.
Buy credits