OpenAI GPT-6.1 Terra: half the API price, strong computer use

What this is about

On October 26, 2026, OpenAI introduced GPT-6.1 Terra. The model sits in the middle of OpenAI's lineup: GPT-6 Astra and GPT-6.1 Sol are above it, and GPT-6 Luna is below. Terra is meant to be fast and efficient. According to OpenAI, it combines detailed image understanding, precise computer use and clearer, more natural writing. It is a clear step up from GPT-5.6 Terra and targets everyday work tasks: writing proposals, analyzing technical drawings, filling out forms or editing videos.

Terra is available now to all Plus, Pro, Enterprise and Edu users in ChatGPT Work and Codex. Developers can reach it through the OpenAI API as gpt-6.1-terra. It is a proprietary model with no downloadable weights.

Pricing

According to OpenAI, standard prices are 50% lower than those of GPT-5.6 Terra and GPT-6.1 Sol.

Cached input, meaning text that has already been processed and is reused, costs just $0.05 per million tokens, a quarter of the previous rate. The comparison figures for Sonnet and GPT-5.6 Terra come from The Decoder. Simon Willison also puts Terra at roughly ten times the price of GPT-6 Luna. In ChatGPT Work and Codex, OpenAI says the same usage allowance stretches to more tasks than with GPT-5.6 Terra, and to significantly more than with Sol or Astra.

Computer use and image understanding

OpenAI puts the emphasis on operating software. Among other things, the model was tested on Agents' Last Exam, a benchmark of professional tasks, such as turning architectural drawings into a 3D model or logging an animated film scene by scene. OpenAI's claims:

• At medium reasoning effort (how much the model is allowed to "think"), Terra is 15 percentage points ahead of GPT-5.6 Terra at about a third of the cost.
• It matches the performance of GPT-6 Astra at a tenth of the per-token price.
• At maximum effort, it is more than three times as token-efficient as Claude Opus 5.5 at a quarter of the token price, which works out to roughly nine times lower cost per task.
• It completes tasks more than three times faster than Astra and Sol.

On MMMU-Pro (visual reasoning), Terra beats its predecessor at every level and comes close to Astra and Sol. On OSWorld 2.1, which gives desktop agents multi-step tasks, it surpasses GPT-5.6 Terra's best result at medium effort for a fraction of the cost.

Fewer hallucinations, better coding

OpenAI has invested heavily in reducing hallucinations, meaning made-up statements. On its internal benchmark of error-provoking prompts, the error rate at low reasoning fell from 53.6% (GPT-5.6 Terra) to 16.7%. At maximum effort it dropped by 18 points to 4.3%, close to Astra and Sol. Artificial Analysis's measurement on AA-Omniscience points the same way: Terra now guesses in only 38% of cases when it doesn't know the answer, down from 88%. Accuracy rose by two points.

On coding, Terra solves 50.3% of tasks on Terminal-Bench 4.0 (coding agents working in the command line) at maximum reasoning. GPT-5.6 Terra managed 21.5%. On Terminal-Bench-Science 0.1, it beats Sol at lower cost and trails Opus 5.5 by less than five points at under a fifth of the estimated cost. For the hardest coding tasks, OpenAI still recommends Sol and Astra. On Artificial Analysis's Coding Agent Index v2.0, Terra scores 48 points, six fewer than Sol (54) and 17 more than Luna (31).

Independent assessment

Artificial Analysis calls Terra a new benchmark for intelligence per dollar. At maximum effort, it gains 6 points on the Intelligence Index over GPT-5.6 Terra and costs less than a third as much per task ($0.40 versus $1.40). At lower effort levels, the gains range from 7 to 13 points.

OpenAI's Ted Sanders wrote on Hacker News that Terra is probably better than Astra and Sol at technical communication, similar (but much faster) at vision and computer use, and slightly worse at most other things. For subscribers it is "not a bombshell," he said, but it helps if you run into usage limits. Simon Willison compares Terra's position to Anthropic's Sonnet: neither the cheapest nor the most expensive model, which makes it hard to place. Still, he says computer use benefits from a cheaper model, since it is often used for repetitive, automated tasks. A pelican image he generated at "max" cost 11.3 cents and took 4 minutes 11 seconds.

Safety

According to OpenAI, Terra reaches the "Critical" level for cybersecurity in the Preparedness Framework and retains the safeguards of GPT-6 Astra. It stays within the given scope more reliably and describes limitations and obstacles more honestly than GPT-5.6 Terra. According to the system card, it withstands direct attacks on the instruction hierarchy about as well as Astra and Sol (99.99%). Against indirect prompt injection, where malicious instructions are hidden in third-party content, the defense rate is 98.8%, up from 96.2% for its predecessor.

Our take

Terra noticeably lowers the price of agentic office and computer tasks, such as reading invoices and entering them into accounting software. For API developers, it is a mid-tier option: cheaper than Sol, much stronger than Luna, especially for images, computer use and communication. Most of the benchmark figures come from OpenAI itself and haven't been independently verified. Still, Artificial Analysis's measurements support the trend. The comparisons with Claude Opus 5.5 are vendor claims with chosen settings. If you need top performance, stick with Sol or Astra.