GPT-6.1 Terra

What it claims

OpenAI released GPT-6.1 Terra on 26 October 2026 as the midrange model in its lineup of GPT-6 Astra, GPT-6.1 Sol, GPT-6.1 Terra and GPT-6 Luna. OpenAI calls it "a significant step up from GPT-5.6 Terra." The announcement names its target tasks: drafting proposals, interpreting technical drawings, filling out forms and editing videos. Simon Willison compares its slot to Anthropic's Sonnet: neither the cheapest nor the most expensive model, which he says makes it harder to decide when to use.

Evidence

The page has no benchmark numbers. The only comparative judgement comes from OpenAI's Ted Sanders in a Hacker News comment. Sanders rates Terra probably better than Astra and Sol at technical communication. He rates it similar to them on vision and computer use, but much faster. He rates it "a bit worse on most other things." Against Luna, he calls it "a lot smarter," especially for vision, computer use and communication.

Willison ran his pelican drawing test at every reasoning level. At the "max" level, one pelican cost 11.3161 cents and took 4 minutes and 11 seconds. He also published a comparison grid of pelicans for every reasoning level across all four current models.

How to use it

API pricing is $1 per million input tokens and $5 per million output tokens. Sanders positions it for API developers who want something cheaper than Sol but stronger than Luna. Willison argues computer use suits a cheaper model because it often runs automated, repetitive tasks. The model offers several reasoning levels, with "max" the highest; the page does not list the others.

Limits

The quality claims are a vendor employee's informal opinion, not measured results. By Sanders' account, Terra trades general capability for speed and price: tasks outside communication, vision and computer use will likely do better on Sol or Astra. For ChatGPT subscribers, Sanders says "it's not a bombshell release." He suggests it may stretch usage limits or give faster computer use. The 4-minute latency and 11-cent cost of one "max" reasoning drawing show that the highest reasoning level is neither fast nor trivially cheap per call.