Code Arena WebDev Gets Terra Max and Carbon High in 48 Hours as Arena Tracks Compute Tiers Separately
Arena.ai updated its Code Arena WebDev leaderboard twice in 48 hours, adding OpenAI’s GPT-6.1 Terra (Max) on October 27 and Google’s Gemini 4 Carbon (High) on October 28. Both variants sit alongside their standard-tier counterparts on the same leaderboard, a pattern Arena has been formalising as frontier labs converge on tiered compute offerings.
What Changed
GPT-6.1 Terra (Max) is the high-compute tier of Terra, OpenAI’s October 26 release targeted at coding and browser-use tasks. Terra launched with a Terminal-Bench 4.0 score of 50.3% at max reasoning effort, more than double the 21.5% posted by its predecessor GPT-5.6 Terra. It runs at $1/$5 per million input/output tokens and is reported to be 3x faster than GPT-6 Astra. The (Max) variant adds additional compute per inference on top of that baseline.
Gemini 4 Carbon (High) is the elevated compute tier of Carbon, which Google launched on October 28. Carbon had already entered Agent Arena on its launch day. Adding Carbon (High) to Agent Arena, Text Arena, and Code Arena WebDev simultaneously signals that Google is making both tiers available for human preference evaluation from day one of launch. On the Artificial Analysis Intelligence Index, Carbon (High) now co-leads with Claude Opus 5.5 (max with fallback), the first Google model to reach that position since the index was introduced.
The Pattern Behind the Additions
Arena now tracks at least eight model variants across multiple compute tiers from three labs: OpenAI (Sol, Terra, Astra each with standard and Max), Anthropic (Opus, Sonnet with Max and xHigh compute options), and Google (Carbon in standard and High). The Code Arena WebDev leaderboard has become the fastest-growing in terms of new entrant density: three new entries in the past seven days.
For the leaderboard itself, separate compute-tier entries reveal data that a single aggregated model row would hide. A Max-tier Terra may rank substantially above standard Terra on a WebDev task that rewards extended reasoning chains, while performing similarly on tasks where additional compute provides diminishing returns. Arena does not publish the separate ELO scores until a model accumulates sufficient battles, so the current rankings for these two new entrants are still settling.
Context: Terra’s Benchmark Profile
Terra’s agentic numbers at standard tier: Terminal-Bench 4.0 at 50.3%, Coding Agent Index v2.0 at 48 (versus GPT-6.1 Sol’s 54 and GPT-6 Luna’s 31), and a LiveBench 2.0 Browser Use score of 93.1%—tied with Sol for the highest Browser Use score among frontier models. Its LiveBench 2.0 overall is 77.8, placing it sixth among models currently on that benchmark.
Carbon’s lab-reported agentic metrics: Terminal-Bench 4.0 at 65.2% (Artificial Analysis evaluation), FrontierSWE 2 at 61.2%, Agents Last Exam v1.1 at 77.4%, and AutomationBench at 53.9%. These numbers come from Google’s launch materials; independent verification via Arena battles will take several days to accumulate.
The two new Code Arena WebDev entries will provide the first direct head-to-head comparison of Terra (Max) versus Carbon (High) under live human-preference conditions—a data point neither lab’s benchmark reports can replicate.