FoxTPNL / Wikimedia Commons (CC BY 4.0) The model, previously referenced as an internal judge for coding evaluations, is now open for users to test on Arena.ai's agentic and coding platforms
OpenAI has added GPT-6.1 Terra to Agent Arena and Code Arena, giving users a chance to test the model on their own tasks.
Both platforms are run by Arena.ai. They exist to let the community stress-test AI systems on practical agentic and coding work, rather than relying on the companies that built the models to grade themselves.
From referee to contestant #
Terra’s earlier role makes this move interesting. According to the research findings, Microsoft had referenced GPT-6.1 Terra as an internal model, with no public release date specified.
In that setting, Terra served as a judge model for coding evaluations across 18 configurations.
Agent Arena handles agentic tasks, meaning jobs where a model has to plan and carry out multiple steps rather than spit out a single answer. Code Arena focuses on programming work.
The scale behind those measurements is substantial. Agent Arena tracks over 2.5 million sessions and uses causal tracing methodologies to evaluate AI performance.
AI, tech, and the markets they move—in one daily briefing.
Daily. Free. Join 34,000+ readers across crypto, finance, and policy.
The bar Terra has to clear #
Terra is not arriving at an empty table. Its sibling, GPT-6.1 Sol, launched for community evaluation on September 29, 2026, on the same two Arena.ai platforms.
Sol’s pitch was efficiency. Per the research findings, it delivers performance near the top-tier GPT-6 Astra variant at one-fifth of the token cost.
The numbers on Sol were specific. Its median task cost came in at approximately $0.56 to $0.82, well below what flagship competitors charge.
Within a few days of launch, Sol climbed into the top five for efficiency. It posted an 11.23% improvement in cost-performance metrics and settled onto the Pareto frontier for cost and efficacy.
What this means for developers and the evaluation game #
There is also a structural question worth watching. Terra’s known history is as an evaluator, scoring coding output across 18 configurations. Seeing how a judge model performs when it is the one being judged could offer a useful window into how well such grading systems reflect real-world quality.
The key things to track from here are where Terra settles in Agent Arena and Code Arena rankings, how its per-task costs compare with Sol’s approximately $0.56 to $0.82 range, and whether the arena debut leads to broader availability. Microsoft’s earlier reference listed no public release date, so the arena may be the clearest signal yet of where Terra is headed.
Disclosure: This article was edited by Diego Almada Lopez. For more information on how we create and review content, see our