OpenAI’s new default model is aimed squarely at teams whose bills come from looking at things. GPT-6.1 Terra takes over from GPT-5.6 Terra in the API, built for visual reasoning, computer use and writing at rates that change the arithmetic for anyone pushing screens, forms and scanned documents through a pipeline at volume.
List pricing runs $1 per million input tokens and $5 per million output. That is half what GPT-6.1 Sol charges at $2 and $10, and a tenth of GPT-6 Astra at $10 and $50. Cached input drops to $0.05, the batch and flex tiers cut the standard rate in half, and a prompt above 272K input tokens bills at double the input rate and 1.5 times output. Context reaches 1,050,000 tokens, output tops out at 128,000, and the knowledge cutoff sits at June 1.
Benchmarks explain the savings. On Agents’ Last Exam v1 the model posts 57.6 percent at medium effort for $0.29 per task, where its predecessor managed 42.6 percent at $0.84 and Astra matched Terra’s score for $4.82. Pushed to maximum effort it reaches 60.0 percent at $1.01, against Claude Opus 5.5 at 63.2 percent for $8.82. OpenAI says tasks wrap up more than three times faster than either Astra or Sol.
Vision survives the downsizing at 85.5 percent on MMMU-Pro, within 1.2 points of Astra. Coding is the weak spot. Terra clears 50.3 percent of Terminal-Bench 4.0 at max effort, behind Sol at 55.9, Astra at 58.2 and Claude Sonnet 5.5 at 70.6, and OpenAI directs harder engineering work toward its pricier models.
Inside ChatGPT the model is already live for Plus, Pro, Enterprise and Edu customers in ChatGPT Work and Codex.
