• Signal: GPT-6.1 Terra lists at $1 input and $5 output per million tokens, against $2 and $12 for GPT-5.6 Terra, per Artificial Analysis.
• Artificial Analysis reports a max-effort cost per task of $0.40 versus $1.40, with a 6-point Intelligence Index gain at max effort.
• Noise risk: the AA-Omniscience guess rate fell from 88% to 38%, but accuracy rose only 2 points, so the model abstains more rather than knowing more.
• OpenAI's system card says it is treating GPT-6.1 Terra as Critical in cybersecurity, which matters for anyone giving it tool access.
• The Terminal-Bench 4.0 gain is described as about 20 points by Artificial Analysis, while The Decoder cites 50.3% versus 21.5%, roughly 29 points; the two sources do not reconcile.

OpenAI's GPT-6.1 Terra, published October 26, 2026, is a genuine price-performance move: listed API prices are about half those of GPT-5.6 Terra, and Artificial Analysis reports higher scores at lower cost per task. The verdict is Signal for agentic and coding workloads, with a caveat. Several of the sharper claims are OpenAI's own or rest on a newly released index, so buyers should rerun their own tasks before migrating.

The price cut is confirmed by independent listings

GPT-6.1 Terra lists at $1 per million input tokens and $5 per million output tokens, against $2 and $12 for GPT-5.6 Terra, according to Artificial Analysis. Cached input falls to $0.05 from $0.20, and cache writes are halved to $1.25. The Decoder reports the same figures and notes Claude Sonnet 5.5 at $2 and $10. Output tokens dominate agent bills, so the 58% output cut matters more than the headline 'half'.

Quality rose, though the sources disagree on one gain

Artificial Analysis reports a 6-point Intelligence Index gain over GPT-5.6 Terra at max effort, 4 points below GPT-6.1 Sol, with 7 to 13 point gains at lower effort. Max-effort cost per task drops to $0.40 from $1.40. On Terminal-Bench 4.0, Artificial Analysis says about 20 points, while The Decoder cites 50.3% versus 21.5%, about 29 points. The article does not explain the gap, so treat the size of that gain as unsettled.

Hallucination is better, but not solved

On AA-Omniscience, Artificial Analysis reports that GPT-6.1 Terra's hallucination rate, meaning how often it guesses when it lacks an answer, fell to 38% from 88% for GPT-5.6 Terra, while accuracy rose by 2 points. The improvement comes mainly from abstaining more often, not from knowing more. For regulated or fact-sensitive use, that is progress but not a reason to remove grounding or review steps.

Treat the vendor comparisons as leads

The Decoder relays claims that Terra matches GPT-6 Astra on Agents' Last Exam at one-tenth the token price and runs computer-control tasks at roughly nine times lower cost than Claude Opus 5.5 at max effort. OpenAI's announcement page returned HTTP 403 when we tried to read it, and we could not reproduce these results, so they remain unverified vendor figures. Artificial Analysis also describes its Coding Agent Index v2.0 as new.

The security classification needs a careful read

OpenAI's system card states that it is treating GPT-6.1 Terra as Critical in cybersecurity and High for biological and chemical capability. The Decoder reports indirect prompt-injection robustness of 98.8%, up from 96.2%, attributing it to the system card. The card's text does not quantify that figure; it appears only in a chart we could not read, so we treat it as reported. Even at 98.8%, agents reading untrusted content need least-privilege tooling.

Verdict: Signal, with a test-first condition

GPT-6.1 Terra is worth benchmarking now if you run high-volume agent or coding workloads on GPT-5.6 Terra, because the price drop and third-party cost-per-task numbers point the same way. It is not a reason to skip your own evaluation, and it is a poor fit where a 38% guess rate on unknowns is unacceptable. Watch for independent reproductions of the Terminal-Bench and injection results before treating the gains as settled.