Artificial Analysis's indices give Terra max an Intelligence Index score of 55 and a Coding Agent Index score of 77, placing it between Sol and Luna on cost; however, Terra is not the optimal point on every cost/intelligence Pareto frontier.
Indices: Artificial Analysis Intelligence Index v4.1 and Coding Agent Index.
Coding agent harness: The article says the index combines agent harnesses including Codex, Claude Code, and Grok Build, and covers DeepSWE, Terminal-Bench v2, and SWE-Atlas-QnA.
Reasoning level: The article primarily reports the max configuration for each GPT-5.6 model and compares the cost/intelligence frontier across different reasoning efforts.
Data source: Artificial Analysis's unified indices and task-cost statistics; the article notes that it supported evaluating OpenAI's Sol, Terra, and Luna during the pre-release phase.
The article discloses the index names, model configurations, per-task costs, and some component evaluations, but does not provide per-question prompts, complete sampling parameters, or downloadable Terra execution traces. Full reproduction requires Artificial Analysis's evaluation service or a public harness at the same version.
Intelligence Index: Terra max scores 55; Sol max scores 59; Luna max scores 51.
Intelligence Index cost per task: approximately $0.55 for Terra, $1.04 for Sol, and $0.21 for Luna; the article summarizes Terra's and Luna's cost differences relative to Sol as approximately 50% and 80%, respectively.
Coding Agent Index: Terra max scores 77, Sol max scores 80, and Luna max scores 75.
Coding Agent cost: The article says that Terra max and Luna max are approximately 60% and 80% cheaper than Sol, respectively.
AA-Briefcase: The article mainly publishes Sol's comparison results and warns that the GPT-5.6 family still needs to be assessed separately on knowledge work using rubric, Elo, and Presentation Elo; Coding Agent Index should not be treated as a measure of professional document quality.
Pareto observation: The article says Luna and Sol continue to lead Terra on the Intelligence and cost frontiers; Terra max's “mid-range price” does not automatically make it the globally optimal value choice.
Terra is a clear mid-range cost/intelligence option: it is cheaper than Sol and has a higher intelligence score than Luna, but the optimal budget point cannot be determined from the model configuration alone.
For Codex- and terminal-based agents, Terra max's score of 77 is enough to place it in frontier coding comparisons; the practical choice still depends on the cost of task failures, the tool harness, and output length.
For batch workloads, first compare “per-task cost × pass rate” on your own task distribution rather than comparing only the price per million tokens.
The article explicitly says that Artificial Analysis supported OpenAI during the pre-release phase, creating potential selection and configuration bias; it remains an external index, but should not be regarded as a completely conflict-free blind test.
The Coding Agent Index is tied to different agent harnesses; model scores and costs cannot be compared directly outside the context of the harness.
Per-task cost is measured for a specific date, reasoning level, and input distribution; caching, retries, tool fees, and account pricing will change the actual bill.
The article does not publish Terra's input, output, and failure cases for each question, so the overall score cannot be independently recalculated.
Record the Artificial Analysis article version, model aliases, max configurations, and collection date.
In the same coding agent harness, fix the task set, timeouts, tool version, and concurrency, then run Terra, Sol, and Luna separately.
For each task, save the pass status, number of model turns, input/output tokens, tool calls, elapsed time, and dollar cost.
Separate the results into Intelligence, Coding Agent, and Briefcase/document quality; do not apply cross-index weighting.
Calculate each model's pass rate, cost per task, and cost per success, then compare them with the directional conclusions in the article.
The article reports “GPT-5.6 Terra (max) and Luna (max) score 55 and 51 respectively”. These figures are index scores, not percentages.
GPT-5.6 Terra
