Gemini 4 Carbon took three front-row spots, but don't overlook (High) and $1.80 just yet

Arena.ai posted the leaderboard on October 28: Google DeepMind's Gemini 4 Carbon (High) ranks No. 1 in Text Arena, No. 2 in Agent Arena, and No. 4 in Code Arena: WebDev. Looking at the three leaderboard positions together, Google has indeed returned to the front row this time. The leaderboard headline is exciting, but the first reaction in the comments was very engineer-like: has it been benchmaxed, where does the default setting rank, and how will costs be calculated after the promotional price ends.

Agent Arena's numbers are the most concrete. Carbon (High) ranks No. 2 with a net improvement of +16.2%. The sample consists of more than 21.9K real agentic conversations from the global community, surpassing GPT-6 Astra and Fable-5.1. Arena.ai gives a median cost per task of $1.80, 72% lower than No. 1 Claude Opus 5.5 (Max). That price gap is striking, especially since agent tasks are usually not cheap.

On cost, commenter Ricardo Dias pointed out that the $1.80 figure is based on Carbon's introductory price of $2/$10; Google says that after the promotional period it returns to $4/$20, which would make the same conversation close to $3.60, narrowing the gap with Opus from 72% to about 44%. This reminder doesn't affect the ranking, but it does affect the 'absurdly cheap' framing.

Among the three sub-signals, Steerability ranks No. 1 with an improvement of +18.6%, meaning whether the model can change direction after the user pushes back. Praise vs Complaint ranks No. 2 at +37.9%, looking at the implicit sentiment in user reactions. Confirmed Success ranks No. 3 at +18.7%, corresponding to an explicit 'Yes, this is useful.' Compared with the overall score, these three are closer to the felt experience of daily use.

In Code Arena: WebDev, Carbon (High) scores 1768 and ranks No. 4. No. 3 Claude Sonnet 5.5 (xHigh) has 1770, only 2 points ahead; No. 11 Gemini 4 Argon (High) has 1679, so Carbon is 89 points higher. On category leaderboards, it ranks No. 3 in Gaming and Simulations and No. 5 in Brand & Marketing. In Text Arena, it tops the chart with 1528 points, while Gemini 4 Argon (High) ranks No. 2 with 1520.

Google DeepMind's own description is that Carbon is a next-generation frontier model and is state-of-the-art in cybersecurity and long-horizon capabilities. Artificial Analysis provides another set of coordinates: No. 1 on Intelligence Index v5 with a score of 59; No. 2 on Coding Agent Index, with 59 points versus 62 for Claude Code plus Opus 5.5.

The disagreements in the comments are more interesting than the leaderboard. Commenter Dusan reminded that the (High) in the name is crucial; most people run the default setting all day, and he would rather see the ranking next to the default configuration. Commenter Mame, after seeing Agent Arena, said that by this measure it still hasn't beaten Opus 5.5. Commenter Alexandr Merk felt that the split between No. 1 in Text and No. 4 in WebDev is what's worth watching. prek said directly that they care most about Steerability ranking No. 1, because the most annoying thing in daily model use is when you push back and it still doesn't change.

Putting these reactions back in, Carbon's results can't be summed up with a single phrase like 'topped the chart.' What it got is a front-row position under a specific setting and a specific price window: No. 1 in Text Arena, No. 2 in Agent Arena with a cost advantage, and No. 4 in WebDev but only 2 points from No. 3. Google went from Gemini 4 Argon to Carbon within two weeks, and the leaderboard update speed is indeed fast. The next things more worth watching are the default setting, the real cost after the promotion ends, and whether the Steerability advantage can be cashed in on long tasks. The leaderboard is the result; settings and price are the context, and without context it's easy to mistake the front row for the whole picture.

Reference links: Arena.ai tweet https://x.com/arena/status/2115495002998751262 , Google DeepMind tweet https://x.com/GoogleDeepMind/status/2115489356437651654 , Artificial Analysis tweet https://x.com/ArtificialAnlys/status/2115494553780367384