Evidence at a glance
The Positioning Shows Up in the Cost per Task
OpenAI released GPT-6.1 Terra as a reasoning model for visual reasoning, computer use, and writing, replacing GPT-5.6 Terra. It is available through the API, ChatGPT Work, and Codex. The model accepts text and images and returns text, with support for tools including computer use, hosted shell, code interpreter, web search, file search, and MCP. Tool calling requires the Responses API. For teams automating screens, forms, and documents, the notable change is not just a model upgrade: Terra brings comparable benchmark results at a much lower listed task cost.
Agents’ Last Exam v1 offers the clearest comparison. At medium effort, Terra scores 57.6% at an estimated $0.29 per task; GPT-6 Astra also scores 57.6%, at $4.82 per task. GPT-5.6 Terra scored 42.6% at $0.84. This shows a substantial gain on that test, not that Terra and Astra behave identically across workloads, or that production costs will fall in direct proportion to the benchmark estimates.
Lower Costs Come with a Choice of Reasoning Effort
Terra offers reasoning effort levels from none to max, with medium as the default. The results show a familiar trade-off: higher effort can improve scores, but it also raises the estimated cost per task. On Agents’ Last Exam, Terra reaches 60.0% at max effort for an estimated $1.01 per task. Claude Opus 5.5 scores 63.2% at $8.82. The practical choice is not simply between cheap and accurate; teams need to decide which tasks merit a larger reasoning budget and which can run at the default setting.
OpenAI also says Terra completes tasks more than three times faster than GPT-6 Astra and GPT-6.1 Sol. The supplied material does not break down latency, retries, or tool execution time enough to independently assess that claim. In production, model speed is not the same as end-to-end throughput. Page loading, tool calls, recovery from failures, and human review can all change the time and cost of a successful task, so benchmark task prices are a point of comparison, not a budget guarantee.
Strong Vision, Better Computer Use, but No Overall Lead
Vision is one of Terra’s clearest strengths. On MMMU-Pro Vision, it scores 85.5% at max effort, below GPT-6 Astra at 86.7% and GPT-6.1 Sol at 86.3%, but above GPT-5.6 Terra’s 81.7%. The supplied figures put Terra’s cost at $0.012 per task, compared with $0.065 for Astra. For use cases OpenAI lists, such as matching product photos to catalog entries or interpreting technical drawings, these results give a concrete reference for balancing capability and cost. They do not replace validation on a team’s own image distribution.
OSWorld 2.1 gives a more mixed picture of computer use. Terra scores 66.7% at medium and 74.2% at max; GPT-6.1 Sol reaches 78.0%, Astra 80.9%, and Claude Sonnet 5.5 83.9%. Terra improves considerably on the previous version’s 57.1%, but still trails these competitors. Matching Astra at medium on Agents’ Last Exam should not be generalized into a claim of parity across computer-use tasks: test sets cover different work, and a benchmark score cannot guarantee the same outcome on a team’s interfaces, applications, and failure modes.
Fewer Errors Do Not Make It a Coding Leader
On an internal set of error-inducing prompts published by OpenAI, Terra’s error rate falls from the previous model’s 53.6% to 16.7% at low effort, and to 4.3% at max effort, an 18-point reduction from the prior version. The set is deliberately adversarial, and OpenAI notes that production error rates will differ. The results support a narrower conclusion: Terra performs better on this error-focused test. They do not establish the probability of errors on ordinary prompts or in a specific business workflow.
Coding results draw a clearer boundary. On Terminal-Bench 4.0, Terra scores 50.3% at max, a large improvement on GPT-5.6 Terra’s 21.5%, but below GPT-6.1 Sol at 55.9%, GPT-6 Astra at 58.2%, and Claude Sonnet 5.5 at 70.6%. OpenAI recommends Sol or Astra for its most complex coding tasks. Yet on Terminal-Bench-Science 0.1, Terra scores 58.3% at an estimated $4.15 per task, ahead of Sol’s 57.0% at $5.47. Coding is not one uniform category. Teams should evaluate against their actual task mix rather than rely on a single overall ranking.
Evaluate the Price Advantage in the Full Deployment
Standard API pricing is $1 per million input tokens and $5 per million output tokens, with cached input at $0.05. Batch and Flex cost half the standard rate, while Fast mode costs twice as much. Prompts above 272K input tokens are billed at twice the input rate and 1.5 times the output rate. For comparison, GPT-6.1 Sol is listed at $2/$10 per million input/output tokens and GPT-6 Astra at $10/$50. Terra’s price does create room for high-volume use, but long prompts, Fast mode, and output volume can materially change the bill.
Deployment constraints matter too. Terra has closed weights and is available only through OpenAI-hosted services, with the API, ChatGPT Work, and Codex listed as entry points. It has a 1.05-million-token context window and a maximum output of 128,000 tokens. A sound evaluation is not to make Terra the default for every task at once, but to select representative screens, forms, images, and documents, then compare success rates, retries, tool costs, and human intervention before deciding where medium or max effort belongs. If workloads are dominated by complex coding, require self-hosting, or demand reliable performance in a particular interface environment, the available evidence does not establish Terra as the better choice.