Halved Prices and Positioning
OpenAI has released GPT-6.1 Terra, a mid-tier model that, according to the company, halves API prices compared with GPT-5.6 Terra. The price list cited by the source specifies $1 per million input tokens and $5 per million output tokens, versus $2 and $12 for its predecessor; cached inputs fall to $0.05 per million, one quarter of the previous rate. The model is available immediately to Plus, Pro, Enterprise, and Edu users in ChatGPT Work and Codex, as well as via API.
The source positions Terra as a direct competitor to Anthropic's Sonnet series, citing a price of $2 per million input tokens and $10 per million output tokens for Claude Sonnet 5.5. The comparison, however, is based on list prices reported in the material, not on actual costs measured in independent use scenarios.
Efficiency Metrics and the Limits of Benchmarks
According to Artificial Analysis, cited by the source, GPT-6.1 Terra sets a new benchmark for intelligence per dollar. At maximum reasoning effort, it gains 6 points on the Intelligence Index over GPT-5.6 Terra, while costing less than one third per task ($0.40 versus $1.40). At lower effort levels, the increase would be even greater—between 7 and 13 points—placing the model on the Pareto frontier, where cost and performance cannot both improve simultaneously.
On the Coding Agent Index v2.0, Terra scores 48 points, six below GPT-6.1 Sol and seventeen above GPT-6 Luna. On the private AA-Omniscience benchmark, its propensity to guess when no answer is available falls from 88% in its predecessor to 38%, with a two-point improvement in accuracy. These are third-party data but cannot be independently verified, and the benchmark's private nature limits its replicability.
Software Control, Security, and More Complex Tasks
OpenAI presents Terra as particularly strong in automated software control. On the Agents' Last Exam benchmark, it would match GPT-6 Astra at one tenth the price per token; compared with Claude Opus 5.5 at maximum effort, it would be more than three times more token-efficient at one quarter the price, with a cost per task about nine times lower, and with completion times more than triple those of GPT-6 Astra and GPT-6.1 Sol. These are company statements, not independent measurements.
On visual reasoning, MMMU-Pro surpasses its predecessor at every level and comes close to larger models; on the OSWorld 2.1 desktop-agent benchmark, it beats GPT-5.6 Terra's best result at medium effort at a fraction of the cost. On coding benchmark Terminal-Bench 4.0, it solves 50.3% of tasks at maximum effort, versus 21.5% for its predecessor. However, OpenAI continues to recommend GPT-6.1 Sol and GPT-6 Astra for the most demanding programming tasks.
On security, the source reports that Terra reaches the 'Critical' threshold of OpenAI's Preparedness Framework and inherits GPT-6 Astra's safeguards. The system card indicates 99.99% resilience against direct attacks on the instruction hierarchy, as with Astra and Sol, and 98.8% defense against indirect prompt injection, up from 96.2% in its predecessor. It remains to be verified in the field how these rates translate into effective protection in applications exposed to third-party content.
