OpenAI releases low-cost "GPT-6.1 Terra" model, cutting API prices to less than half of its predecessor

OpenAI releases GPT-6.1 Terra

OpenAI on Oct. 26 (Oct. 27 Japan time) released the AI model "GPT-6.1 Terra." It positioned the model for everyday work at a lower price than the higher-end "GPT-6.1 Sol," and cut developer API pricing to less than half that of its predecessor. The company said it improved the ability to read screen content and operate apps, as well as writing ability. The announcement came on the same day rival Anthropic filed to list its shares.

A few days before the release, its name had appeared in Microsoft and GitHub configuration files as a cheaper alternative to GPT-6.1 Sol.

API pricing undercuts two rivals

API pricing was set at $1 for input and $5 for output per 1 million tokens, the units of text that AI processes. That is down from $2 for input and $12 for output for the predecessor, "GPT-5.6 Terra." The higher-end Sol, Anthropic's "Claude Sonnet 5.5" and Google's "Gemini 4 Argon" are all priced at $2 for input and $10 for output, and Terra undercuts them. Google plans to raise Argon's price to $4 for input and $20 for output in December.

The cost of having a model solve one test question, as measured by evaluation firm Artificial Analysis, fell to $0.40 for Terra from $1.40 for its predecessor. Gemini 4 Argon was $1.62 and is expected to be $3.98 after the December price increase. Terra was measured at the setting where the AI reasons the longest, and Argon at the setting one level below.

When text exceeding 272,000 tokens is fed in at once, pricing becomes $2 for input and $7.50 for output per 1 million tokens. The discounted rate for repeatedly feeding in the same text (cached input) is $0.05.

Available in ChatGPT paid plans and Amazon Bedrock

In the conversational AI "ChatGPT," paid Plus, Pro, Enterprise and Edu users can choose Terra in "ChatGPT Work," a feature for delegating tasks, and in the coding assistant "Codex." The free version and the low-cost Go plan are not included. No change to ChatGPT's monthly fees has been announced.

It also became available on Amazon's cloud service "Amazon Bedrock" on the day of release.

OpenAI's measurements put it close to the top model

According to OpenAI's own measurements, Terra scored 60.0% on "Agents' Last Exam," a test of delegating work to AI, beating the 59.3% of OpenAI's top-end "GPT-6 Astra." On "OSWorld 2.1," a screen-operation test, it scored 74.2%, short of Astra's 80.9%. On "Terminal-Bench 4.0," a development task test, it scored 50.3%, more than double its predecessor's 21.5% but below Astra's 58.2%. OpenAI said Terra can achieve results on par with Astra on Agents' Last Exam at one-tenth of Astra's per-token price.

On the same Agents' Last Exam, a table published by Google shows Astra at 34.2%. Results vary widely depending on which company does the measuring.

OpenAI said Terra's cyberattack-related capabilities reached "Critical," the highest level under its internal standards, and applied the same safety measures as for Astra.

Falls short of Sol in third-party overall evaluation

In Artificial Analysis's overall evaluation (out of 100), Terra scored 48. That was up 6 points from its predecessor's 42, but 4 points short of Sol's 52 and 10 points short of Anthropic's "Claude Opus 5.5" at 58. In an evaluation focused on coding, Terra also scored 48, below Sol's 54. OpenAI recommends Sol or Astra for difficult development work.

Response speed also remains an issue. OpenAI says it completes tasks more than three times faster than Sol or Astra, but at the setting where it reasons the longest, it takes nearly 40 seconds on average before an answer begins to appear. The median across all models compared by Artificial Analysis is 3.65 seconds. The rate at which it guesses answers to questions it does not know fell from 88% for its predecessor to 38%.

As of Oct. 31, Artificial Analysis's web page showed Terra's longest-reasoning setting at 51 points.

Anthropic gains share of enterprise spending

Anthropic released its listing application documents on the same day, Oct. 26. According to reports, it is aiming for a Nasdaq listing in mid-November at a valuation of up to around $2 trillion. Revenue for the January-June period of 2026 was $16.3 billion, up sharply from about $1.2 billion a year earlier.

In corporate spending, Anthropic has been gaining share. According to an index compiled by corporate card company Ramp based on transactions on its cards, OpenAI's share of spending on AI models fell from 44.2% in late September to 35.1% in late October. Anthropic's share rose from 51.3% to 60.5%.

Over the same period, the average amount per token that companies paid OpenAI fell from $0.80 to $0.46. This average also reflects companies switching to cheaper models.

Ramp's figures count only payments by companies using Ramp's corporate cards.