GPT-6.1 Sol or Terra? Which to pick, and what each one costs

OpenAI sorts its models into four classes, from cheapest and least capable to most powerful: Luna, Terra, Sol and Astra. Terra is the middle, everyday model. Sol sits one step below Astra, the company’s top model.

• Price (API): Terra costs $1 per million input tokens and $5 per million output tokens. Sol costs $2 and $10. Tokens are the chunks of text a model reads and writes.
• Speed: OpenAI says Terra completes tasks over three times faster than both Sol and GPT-6 Astra.
• Best at: Terra handles computer use, visual reasoning and routine office tasks. Sol targets complex programming, debugging and multi-step business workflows.
• Where to use them: both are in ChatGPT Work and Codex for paid plans, and in the API as separate models.

Terra is built for the repetitive work that fills a normal day. OpenAI’s examples include editing videos, filling out forms, reading technical drawings and drafting proposals. A big part of that is computer use, meaning an AI that looks at your screen and clicks and types the way a person would. OpenAI’s own example of high-volume work is reading incoming invoices and typing the details into accounting software.

It’s also a big jump over the Terra it replaces. On Agents’ Last Exam, a test of professional tasks, the new Terra beat its GPT-5.6 predecessor by 15 percentage points when set to medium effort, while costing about a third as much. OpenAI says it matches GPT-6 Astra on that test at one-tenth Astra’s price per token.

Accuracy improved too. On OpenAI’s internal set of prompts designed to trip models up, Terra’s error rate at low reasoning effort fell from 53.6% to 16.7%, about one prompt in six. At maximum effort it dropped to 4.3%, close to Sol and Astra. All of these figures are OpenAI’s own.

Sol is the model for hard problems. OpenAI claims GPT-6.1 Sol nearly catches Astra in science, office work, programming and operating a computer, at roughly one-fifth of Astra’s standard API cost.

• Coding: on DeepSWE, a software engineering test, Sol matches GPT-6 Astra and beats the previous GPT-6 Sol by 6.4 percentage points.
• Computer use: on the offline set of OSWorld 2.0, Sol comes within 2.1 points of Astra.
• Science: it more than doubles GPT-6 Sol’s score on Terminal-Bench Science 0.1, though Astra still leads at 68.1%.
• Safety: its failure rate in computer-use safety stress tests fell from 17.4% on GPT-6 Sol to 4.3%. Astra is lower still, at 2.4%.

Terra’s limits show on tough coding. On Terminal-Bench 4.0, which tests coding agents working in a computer’s command line, Terra solved 50.3% of tasks at maximum effort, more than double GPT-5.6 Terra’s 21.5%, but that still leaves about half unsolved. OpenAI’s advice is to use Sol or Astra for the most complex coding.

There’s one surprise in Terra’s favor. On Terminal-Bench Science 0.1, OpenAI says Terra at maximum reasoning beats GPT-6.1 Sol at a lower cost. So for some research-style coding, the cheaper model can win.

Yes, if you pay, but not everywhere. Both models live in ChatGPT Work, OpenAI’s agent mode that works across apps and websites for you, and in Codex, its coding tool. Neither has arrived in regular ChatGPT conversations yet.

• GPT-6.1 Sol: available to Plus, Pro, Business, Enterprise and Edu users.
• GPT-6.1 Terra: OpenAI lists Plus, Pro, Enterprise and Edu users.
• Free users: get the lower-tier GPT-5.6 Luna.

Developers can call Terra in the API as gpt-6.1-terra. Cached input, text the model has already seen in a conversation, is cheaper on both: $0.05 per million tokens for Terra and $0.10 for Sol.

For most people, the choice comes down to how hard the task is and how often you run it.

• Routine office work: Terra. Filling forms, sorting invoices, summarizing documents and clicking through apps are exactly what it’s tuned for, and speed matters when you run them all day.
• Building an app at scale: Terra, unless quality falls short. Half the price on every token is a big saving across millions of requests.
• Hard coding or debugging: Sol. OpenAI itself steers complex programming away from Terra.
• Long multi-step business workflows: Sol, which OpenAI pitches for exactly that.
• Scientific or research coding: test both. Terra beat Sol on one science benchmark at lower cost.

Keep the safety picture in mind, too. OpenAI says Terra meets the Critical cybersecurity threshold under its Preparedness Framework, the highest of its two cyber risk levels, and keeps the safeguards it built for GPT-6 Astra. That mostly affects what the model will refuse, not everyday office tasks.