OpenAI cuts prices in half with a model that makes things up less

In short: on 26 October 2026 OpenAI released GPT-6.1 Terra, the middle model in its line-up. It is much cheaper than its predecessor GPT-5.6 Terra and, according to independent measurements by Artificial Analysis, makes up answers less often.

What changed

• Price per million tokens: $1 input, $5 output. The predecessor cost $2 and $12.
• Cached input: $0.05 per million (previously $0.20).
• Context window: about 1.05 million tokens, with up to 128K tokens of output.
• Artificial Analysis: at maximum effort the model scores 6 points higher on its index than the predecessor, and the cost per task falls from $1.40 to $0.40.
• The hallucination rate on the AA-Omniscience test dropped from 88% to 38%.

Where the catches are

• Prompts longer than 272K tokens cost 2x on input and 1.5x on output, for the whole request.
• Batch modes are 50% cheaper, but responses arrive more slowly.
• In its system card OpenAI rates the model “Critical” for cybersecurity capability and applies the same safeguards as its larger models, so some security-related requests may be blocked.
• The same system card shows mixed health results: higher than the predecessor on HealthBench Professional, but 1.9 points lower on HealthBench.
• For the hardest coding tasks, according to The Decoder, OpenAI still recommends its larger models.

How to read this

Prices come from OpenAI’s documentation and the scores from Artificial Analysis, an outside company, but on its own tests. Your workload may behave differently, so compare costs on a few dozen real examples before switching. I collect model price comparisons in the model pricing hub.

Sources

GPT-6.1 Terra model documentationOpenAI · Oct 27, 2026 · Primary sourceGPT-6.1 Terra system cardOpenAI · Oct 27, 2026 · Primary sourceGPT-6.1 Terra pushes the cost-efficiency frontierArtificial Analysis · Oct 27, 2026 · Primary sourceGPT-6.1 Terra halves OpenAI's API pricesThe Decoder · Oct 27, 2026