Gemini 3.8 Flash still tops Gemini 4 Argon, Carbon and Hydrogen in our tests
Published 2026-11-13 · Data captured 13 Nov 2026, 00:46 UTC
At high effort, Gemini 3.8 Flash scored 9.91/10 and ranked #2 on the leaderboard. Gemini 4 Argon high came closest at 9.55 and answered twice as fast (19.0 s vs 40.6 s) for about the same cost. Gemini 4 Hydrogen did best at medium (9.30) and low (9.22 in 8.3 s for $0.669), not at its default high setting (8.86). Every answer Gemini 4 Carbon returned at high was correct, but 5 of 69 attempts failed with Google API errors or timeouts. That left it at 8.82, with the slowest (115.3 s) and most expensive ($5.218) run of the four.
See current Gemini 4 Hydrogen results →
- Tested through OpenRouter. Prices per million tokens: Hydrogen $1.50 input and $7.50 output; Argon and Carbon $2 and $10 (Google's introductory price, $4 and $20 later); 3.8 Flash $0.75 and $3.75.
- Costs are recorded spend for one full run (23 tests, 3 repeats each). Failed API calls stay in the results and count as unsolved attempts.
Key results
Gemini 3.8 Flash still leads. 9.91 at high, #2 on the leaderboard, 68 of 69 attempts correct, $1.927 per run.
Argon is the faster near-equal. 9.55 at high for $1.802, with 19.0 s average responses vs 40.6 s for 3.8 Flash.
Hydrogen's default is its weakest setting. High scored 8.86. Medium scored 9.30 and low 9.22, both above Argon at the same effort.
Carbon was held back by the API, not its answers. At high, every answer it returned was correct, but 5 of 69 attempts ended in Google API errors or timeouts. Its run cost 2.7 times as much as 3.8 Flash and took 2.8 times as long per answer.
Google's claims, checked. Google reports Hydrogen above Argon at its default high setting and 2.5 times faster than 3.8 Flash. In our tests Hydrogen high trailed Argon high by 0.69 points, and answered 1.25 times faster than 3.8 Flash high (32.6 s vs 40.6 s).
| Model | Effort | Rank | Score | Solved tests | Correct attempts | Run cost | Cents per solved test | Avg response |
|---|---|---|---|---|---|---|---|---|
| Gemini 3.8 Flash | high | #2 | 9.91 | 22/23 | 68/69 | $1.927 | 8.756 | 40.6 s |
| Gemini 4 Argon | high | #19 | 9.55 | 21/23 | 67/69 | $1.802 | 8.578 | 19.0 s |
| Gemini 4 Carbon | high | #77 | 8.82 | 19/23 | 64/69 | $5.218 | 27.463 | 115.3 s |
| Gemini 4 Hydrogen | high | #74 | 8.86 | 18/23 | 64/69 | $2.028 | 11.263 | 32.6 s |
Score, cost and speed
3.8 Flash and Argon cost about the same per run; Argon is twice as fast. Carbon's high run was the most expensive and by far the slowest.
Score (out of 10, higher is better)
Full-run cost in USD (lower is better)
Average response time in seconds (lower is better)
Which effort to use
Argon and 3.8 Flash improve with effort: Argon goes from 8.25 at low to 9.55 at high, 3.8 Flash from 7.21 to 9.91.
Hydrogen stays between 8.86 and 9.30 at every effort. Xhigh cost 13 times as much as low and took 10.8 times as long per answer, for 0.06 points less. Use low or medium.
Carbon rose from 8.36 at low to 8.82 at high, and lost 5 to 7 attempts to API failures at every effort.
| Model | Effort | Rank | Score | Solved tests | Correct attempts | Run cost | Cents per solved test | Avg response |
|---|---|---|---|---|---|---|---|---|
| Gemini 3.8 Flash | low | #210 | 7.21 | 16/23 | 56/69 | $0.323 | 2.014 | 7.5 s |
| Gemini 3.8 Flash | medium | #149 | 7.97 | 19/23 | 62/69 | $0.769 | 4.046 | 16.6 s |
| Gemini 3.8 Flash | high | #2 | 9.91 | 22/23 | 68/69 | $1.927 | 8.756 | 40.6 s |
| Gemini 4 Argon | low | #133 | 8.25 | 15/23 | 55/69 | $0.438 | 2.916 | 5.7 s |
| Gemini 4 Argon | medium | #63 | 8.96 | 19/23 | 58/69 | $0.674 | 3.545 | 6.6 s |
| Gemini 4 Argon | high | #19 | 9.55 | 21/23 | 67/69 | $1.802 | 8.578 | 19.0 s |
| Gemini 4 Carbon | low | #122 | 8.36 | 17/23 | 58/69 | $0.315 | 1.851 | 16.9 s |
| Gemini 4 Carbon | medium | #89 | 8.68 | 19/23 | 61/69 | $1.578 | 8.302 | 42.4 s |
| Gemini 4 Carbon | high | #77 | 8.82 | 19/23 | 64/69 | $5.218 | 27.463 | 115.3 s |
| Gemini 4 Hydrogen | low | #38 | 9.22 | 19/23 | 63/69 | $0.669 | 3.518 | 8.3 s |
| Gemini 4 Hydrogen | medium | #35 | 9.30 | 20/23 | 62/69 | $1.310 | 6.550 | 20.8 s |
| Gemini 4 Hydrogen | high | #74 | 8.86 | 18/23 | 64/69 | $2.028 | 11.263 | 32.6 s |
| Gemini 4 Hydrogen | xhigh | #46 | 9.16 | 19/23 | 62/69 | $8.734 | 45.967 | 89.9 s |
Strengths and weak points
All four scored 10 in agentic, coding, tool-calling, instruction-following and puzzle tests. Only 3.8 Flash high also scored 10 in combined tests.
Carbon's low combined (3.75) and trivia (4.67) scores come from attempts that failed with API errors, not from wrong answers. Hydrogen high also lost combined and general-intelligence attempts to API failures, and answered two domain-specific questions wrong.
Gemini 3.8 Flash high
Gemini 4 Argon high
Gemini 4 Carbon high
Gemini 4 Hydrogen high
Who should choose which
Gemini 3.8 Flash high: the most accurate here, if 40-second answers are fine.
Gemini 4 Argon high: close behind at about half the response time and a similar cost.
Gemini 4 Hydrogen low or medium: a strong everyday choice. Low scored 9.22 in 8.3 s for $0.669 per run. Skip xhigh.
Gemini 4 Carbon: wait until Google's API settles. Its completed answers were all correct, but its runs were slow and expensive.
Availability and prices
Gemini 4 Hydrogen launched on November 12, 2026 in the Gemini API, AI Studio, the Gemini app and Google Search, with four thinking levels and high as the default.
Google announced Gemini 4 Argon on September 30, 2026 and Gemini 4 Carbon on October 28, 2026, first for a limited group of cyber defenders. Both cost $2 input and $10 output per million tokens at the introductory price, rising to $4 and $20 later.
Gemini 3.8 Flash launched on September 2, 2026 at $0.75 input and $3.75 output per million tokens.
Method and limits
Every model ran the same 23 tests three times through OpenRouter. Scores are out of 10, ranks are leaderboard positions at capture time, and comparisons use the 422 complete, non-archived chat results.
Gemini 3.8 Flash was tested in September and October 2026, Argon in October, and Carbon and Hydrogen on November 12, in their launch week. Tests 14, 15 and 22 kept the same prompts, answers and pass scores; only their recorded evaluator version differs.
Failed API calls count as unsolved attempts, so a later retest may score Carbon higher. Response times include provider and workload effects. Costs are recorded spend, not a price guarantee.
Sources
- Google: Introducing Gemini 4 Hydrogen (accessed 2026-11-13)
- Google: Introducing Gemini 4 Carbon (accessed 2026-11-13)
- Google: Gemini 4 Argon (accessed 2026-11-13)
- 9to5Google: Gemini 3.8 Flash launch (accessed 2026-11-13)
Linked model pages show current results, which can differ from these figures.