Thinking Levels · Google
Gemini 4 Carbon: Writing Quality by Reasoning Effort
Last updated , from board data generated at 17:52 UTC.
This page compares Gemini 4 Carbon at 3 reasoning-effort settings, each tested on the same 10 scripts: the Writing Score (the 0 to 100 consensus of the judges and code), the Writing Elo (a rating from pairwise comparisons of those scores), and the cost and time of each setting. As a model, Gemini 4 Carbon ranks #27 of 87 models through its best setting. The other ranks on this page are configuration ranks: one setting's place among all 151 configurations.
What More Effort Changes for Gemini 4 Carbon
Thinking lifts Gemini 4 Carbon a long way, and the biggest step is the first one. At low, the model recorded no thinking tokens at all and lands at #76. Medium climbs 184 Elo to #63, with ranges that don't overlap. High adds another 75 and reaches #54, the best Google configuration on the board. Medium's range tops out right where high's starts, at the edge both share, though comparing the two ranges doesn't test the gap between them. From low to high, that's 259 Elo, enough for high to beat low in about 82% of script comparisons. The scores show where the thinking goes. Tone & Voice Match and writing craft barely move across the three levels, and the hook even slips a little. What climbs is length: at low, the scripts run well past their target, medium gets much closer, and high lands almost right on it, closer than any other Google row. The slop, substance and visual cue scores tick up a little at each step too. High thinks about three times as long as medium, and the bill follows: a script costs $0.061 at low, $0.218 at medium and $0.586 at high, and the wait grows from about 0.8 minutes to 3.1 and 8.7.
Recommended Setting
I'd run Gemini 4 Carbon at high when the script is going out. Its scripts come back right at their target length, so there's almost nothing to trim, but at $0.586 and about 8.7 minutes a script, it's also the priciest and slowest Google row. For everyday drafts, I'd take medium: #63 for $0.218 and about 3.1 minutes, 75 Elo behind high on the point estimates, while high costs 2.7 times as much. Low is for bulk drafts you'll fix by hand. At $0.061 a script, it still ranks above Gemini 4 Argon (high thinking) on the point estimates, with overlapping ranges, and Argon at high costs 3.7 times as much. Just expect to cut its scripts down.
Set the level on every call. Google documents medium as what a request gets when it sets no level, and a run here with no level set thought about as much as the medium row did, so the board counts it as medium and doesn't list it twice. That's the opposite of Gemini 4 Argon, where leaving the level out got no thinking at all, so I wouldn't leave it to the API.
One thing to know if you call it yourself at medium or high. One non-streaming Gemini call caps its output, thinking included, at just under thirty-three thousand tokens. Carbon at high usually thinks past that, and at medium it sometimes does. Its reply then comes back with the finish reason CONTINUATION instead of STOP, with almost no script in it yet, and you get the rest by sending the same request again with the continuation token Google returns, then joining the parts. That's how Google serves long replies, and most high drafts here took two calls. Google bills every call, and the costs on this page include all of them. So check the finish reason: code that stops at the first reply can hand you a lot of thinking and barely any script.
Keep the price in perspective, though. GPT-6.1 Terra (high) ranks a few places above Carbon at high, at #49, with a range that overlaps Carbon's, for $0.027 a script and about 0.5 minutes. Overlap alone doesn't say which one writes better, but Carbon at high costs 21.6 times as much. And GPT-6 Sol (medium) ranks #21, with a range clear of Carbon's, for $0.056. So I'd run Carbon at high when the work has to stay on Google, and go to another lab for the best script per dollar.
Every Setting on the Chart
Each point is one effort setting of Gemini 4 Carbon, joined from the lowest setting to the highest. Higher and further left is better: a higher score for less money or less time.
How to Use the Chart
- A hollow marker is the provider default: the level the provider applies when the request sets no effort level.
- The tabs choose the score on the vertical axis, X axis switches between cost and time, and Table lists the same numbers as rows.
- Select a point for its numbers and a link to its scorecard.
On a small screen, Table lists the same numbers as rows.
All Effort Settings, Step by Step
| Setting | Writing Elo | Step | Writing Score | Cost / Script | Time / Script | Reasoning Tokens | Configuration Rank |
|---|---|---|---|---|---|---|---|
| low | 1590 ±40 | - | 77.2 | $0.061 | 0.8 min | - | #76 |
| medium | 1774 ±40 | +184 | 79.2 | $0.218 | 3.1 min | 14,771 | #63 |
| high | 1849 ±33 | +75 | 80.1 | $0.586 | 8.7 min | 43,844 | #54 |
Other Google Models by Effort
Writing Elo comes from comparing every pair of models on every script, which puts these settings on the same scale as the leaderboard. Step is the Writing Elo change from the previous explicit setting. Cost is one full script at API list prices, and time is the average for one accepted script, retries included. The thinking-levels page covers the other models, and the Google family page lists every Google model side by side.