Reviews · OCTOBER 26, 2026
GPT-6.1 Terra ships at $1/$5 per million, halving the price of OpenAI's computer-use tier
The new Terra lands at half GPT-5.6 Terra's token price, scores 50.3% on Terminal-Bench 4.0 at max effort, and cuts the low-effort error rate from 53.6% to 16.7% — reshaping the cost math for screen-agent workflows.
OpenAI released GPT-6.1 Terra on October 26, 2026, priced at $1 per million input tokens and $5 per million output, exactly half the token cost of GPT-5.6 Terra. The model is available in ChatGPT Plus, Pro, Enterprise, and Edu, in Codex, and in the API as gpt-6.1-terra.
The headline number is the price cut. The more consequential number is the error rate.
On OpenAI's internal error-inducing prompt set, GPT-5.6 Terra failed 53.6% of the time at low reasoning effort. Terra fails 16.7% of the time on the same set. At maximum effort, the error rate drops 18 points to 4.3%, which OpenAI says nearly matches GPT-6 Astra and GPT-6.1 Sol. For anyone running screen-driving agents against real CRMs and inbound forms, that gap is the difference between a workflow that needs a human babysitter and one that doesn't.
Artificial Analysis measures a roughly 20-point gain on Terminal-Bench 4.0, with Terra reaching 50.3% at maximum reasoning effort against GPT-5.6 Terra's 21.5%. On the firm's Coding Agent Index v2.0, Terra scores 48, six points behind Sol's 54 and 17 ahead of GPT-6 Luna's 31. The AA-Omniscience benchmark tells the sharper story: Terra's guessing rate collapses from 88% to 38%, a 50-point drop, while accuracy ticks up two points. The model has learned to decline more and bluff less.
Cache pricing moved too. Cache writes land at $1.25 per million, and cached input tokens drop from $0.20 to $0.05, a four-fold reduction that disproportionately rewards workloads that reuse long system prompts, exactly the pattern that lead-capture and CRM-entry agents run all day.
The practical shift is boring in a useful way. A nightly batch job that cost roughly $400 per month on GPT-5.6 Terra now costs about $200 on Terra for the same prompt volume, and it fails less often. A 5-to-30-person team that was priced out of autonomous screen agents at Sol tier can run them here. The template for this cycle was set earlier in October, when GPT-6.1 Sol landed at one-fifth of Astra's price, and reinforced days later when Claude Haiku 5.5 shipped at $0.10 per million. Each release has pulled a different category of agent workload across the cost threshold that makes it worth running unattended.
Terra isn't the frontier. It's the first OpenAI mid-tier where the cost-per-task math for computer use stops being a thought experiment.