OpenAI Releases GPT-6.1 Terra: Tokens Cost Half as Much, Hallucinations Drop From 88% to 38%

OpenAI has released GPT-6.1 Terra, the first Terra-class model since GPT-5.6. The independent lab Artificial Analysis calls it a new frontier in price efficiency. The per-token price dropped by at least half: $1 per million input tokens and $5 per million output tokens, instead of $2 and $12. At the maximum effort level, the model also scores 6 points higher than its predecessor on the Artificial Analysis Intelligence Index, while the average cost of a single task fell from $1.40 to $0.40.

Pricing: what got cheaper and by how much

The cut applies to the entire price list, including the cache. For agentic systems that loop over a long context, the cache is what determines the final bill.

• Input tokens: $1 per million (GPT-5.6 Terra charged $2).
• Output tokens: $5 per million (previously $12).
• Cache writes: $1.25 per million tokens, half the previous price.
• Cached input: $0.05 per million tokens versus $0.20, which is four times cheaper.

Output tokens got 2.4 times cheaper, while the cost of a task fell by almost 3.5 times. The gap exists because 6.1 Terra solves tasks more efficiently, not just at a lower per-token price.

Where the model sits on the Pareto frontier

Artificial Analysis compares models on an "intelligence versus cost" basis. At low and medium effort levels, GPT-6.1 Terra sits on the Pareto frontier: you cannot get the same quality for less. The other modes remain off the frontier. At maximum effort, the gain over GPT-5.6 Terra is 6 points; at lower levels it is larger, from 7 to 13.

Within OpenAI's own lineup, the model sits in the middle. In max mode it trails GPT-6.1 Sol by 4 points and beats GPT-6 Luna by 10 points.

In the new Artificial Analysis Coding Agent Index v2.0, Terra scores 48 points. GPT-6.1 Sol scores 54 and GPT-6 Luna scores 31. Here, the xhigh and low modes are the ones on the efficiency frontier.

Hallucinations and terminal agents

The most striking number in the release concerns reliability. In AA-Omniscience, a closed benchmark of knowledge and hallucinations, GPT-6.1 Terra guesses at an answer 38% of the time when it does not know it. For GPT-5.6 Terra, that figure was 88%. The 50-point drop comes with a 2-point gain in accuracy, meaning the model now answers correctly more often and fabricates less.

The second notable jump is in Terminal-Bench 4.0, an agentic benchmark for working in a terminal. The gain was about 20 points, and at maximum effort Terra matched GPT-6.1 Sol. In most other tests the improvements are more modest: for example, 5 points in GDP.pdf. Scores are from Intelligence Index v4.3.2.

What else Artificial Analysis released

The lab published several other items in the same days:

• Code2Video-AA: an implementation of HeyGen's Code2Video Bench, in which models produce short product and brand videos from 114 design briefs using HyperFrames HTML (October 26).
• Kimi K3.1 Pro: debuted in 4th place on the Intelligence Index and will become the leader among open-weights models (October 23).
• AA-PersonalityBench: a panel of LLM judges compares responses pairwise across 30 personality traits that roll up into 8 indices, and separately measures 21 communication-style characteristics (October 19).

What to change in production

For teams building agents on GPT-5.6 Terra, the upgrade looks like an obvious call: the cost per task drops by more than a factor of three, and quality goes up. Choosing the effort level now matters more than choosing the model. Low and medium give the best price-to-quality ratio on the general index, while for coding agents the frontier is held by low and xhigh. The other modes cost more without delivering a gain.

The drop in fabricated answers from 88% to 38% helps in scenarios where an error is expensive: support, knowledge-base search, report automation. For complex agentic chains, Terra's 48 points on the Coding Agent Index cover a large share of what teams previously had to pay Sol for, and Sol's 54 points are needed only where every point matters. Caching at $0.05 per million tokens makes long sessions with repeated context economically sensible.