OpenAI released GPT-6.1 Terra on October 26, 2026, and the number that jumps out of the independent analysis is not the price, though the price halved. It is that Artificial Analysis measured the model's hallucination rate on its AA-Omniscience test falling from 88% for GPT-5.6 Terra to 38%.

That is a big claim, so this guide takes it apart: what the tests measure, what the price sheet looks like, how Terra compares with the bigger GPT-6.1 Sol, where it is available, and what OpenAI itself says versus what third parties measured.

The price sheet

Artificial Analysis lists GPT-6.1 Terra at $1 per million input tokens and $5 per million output tokens, against $2 and $12 for GPT-5.6 Terra. Cached input is $0.05 versus $0.20, and the cache-write price of $1.25 is described as halved. (Source: Artificial Analysis)

A separate pricing write-up adds details: a 1,050,000-token context window, 128,000 maximum output tokens, text and image input with text output, and six reasoning-effort settings from none to max. Requests above 272K input tokens are billed at 2 times the input and cache rates and 1.5 times the output rate for the whole request, not just the excess. Batch and Flex processing are 50% below Standard, and Fast processing is 2 times Standard. By that write-up's worked example, 100K uncached input plus 5K output costs $0.125. (Source: AI Cybr)

Reading the 88% to 38% hallucination claim

On AA-Omniscience, Artificial Analysis reports a hallucination rate of 38% for GPT-6.1 Terra, down from 88% for GPT-5.6 Terra, a 50-point drop, while accuracy rose only 2 points. The pattern implies the gain came mostly from the model being less likely to give a wrong answer when it does not know, not from knowing more. The source article does not break down the mechanism, so that is an inference. (Source: Artificial Analysis)

Do not confuse this with OpenAI's own figures. The pricing write-up reports an OpenAI-internal factuality test on difficult prompts, with an error rate of 16.7% at low reasoning and 4.3% at maximum reasoning. That is vendor-reported on a different test and shows the same direction. It is not directly comparable to the 38% number. (Source: AI Cybr)

Also on LumiChats

AI Models

Claude Sonnet 5.5: Nearly 7x the Coding Score, Same $2 Price

8 min read→

AI Models

GPT-6 Sol and Luna: OpenAI Cuts Prices in Half

7 min read→

AI Models

The Best AI Model Scores 12% on This New Test. Here's Why.

8 min read→

Where it sits next to Sol and Luna

• Coding Agent Index v2.0: GPT-6.1 Terra 48, GPT-6.1 Sol 54, GPT-6 Luna 31 (Artificial Analysis).
• Intelligence Index at max effort: 6 points above GPT-5.6 Terra, 4 points below GPT-6.1 Sol and 10 points above GPT-6 Luna; at lower effort levels, gains over GPT-5.6 Terra are 7 to 13 points.
• Terminal-Bench 4.0: Artificial Analysis says it rose about 20 points and is close to GPT-6.1 Sol at max effort. A separate write-up lists a vendor-reported 50.3% at maximum reasoning versus 21.5% for GPT-5.6 Terra; the two measurements use different setups.
• GDP.pdf: up 5 points; most other evaluations showed smaller gains.
• On its Intelligence Index, Artificial Analysis puts the low and medium effort levels on the cost-efficiency frontier; on its Coding Agent Index, the xhigh and low levels sit on it.

In plain terms, Sol remains the stronger choice for demanding coding. One write-up notes that OpenAI itself positions Sol and Astra as the better picks for the hardest coding work, while Terra targets everyday agentic and computer-use tasks at lower cost. (Sources: Artificial Analysis, AI Cybr)

Computer use and speed (vendor-reported)

The same write-up relays OpenAI-reported results: Agents' Last Exam up 15 percentage points over GPT-5.6 Terra at medium reasoning, at roughly one-third of the predecessor's task cost; OSWorld 2.1 above the predecessor's best score without an exact figure given; and computer-use tasks completed more than three times faster than with GPT-6 Astra and GPT-6.1 Sol in OpenAI's tested evaluation. These are vendor numbers, and independent replication has not been cited. (Source: AI Cybr)

Where you can use it

• API: model ID gpt-6.1-terra, with tool calling requiring the Responses API.
• ChatGPT and Codex: rolling out to Plus, Pro, Enterprise and Edu users, per a pricing write-up summarizing OpenAI.
• GitHub Copilot: generally available on Pro, Pro+, Max, Business and Enterprise, with gradual rollout. GitHub says it is billed at provider list pricing under usage-based billing and was aimed at everyday agentic coding and terminal work at lower cost than Sol. In early testing it solved noticeably more terminal and C# tasks than GPT-5.6 Terra.
• Limits: fine-tuning is not supported in the current model listing, no open weights have been announced, and computer use needs a separate driver to execute actions.

For Copilot Business and Enterprise teams, GitHub notes that new models are enabled automatically under default settings unless an administrator turns off the global default or disables the model. Check your model policy if cost control matters. (Source: GitHub Changelog)

What the numbers do not tell you

• Artificial Analysis does not publish token usage or the total cost to run its indices in the cited article, so real-world cost depends on your prompts.
• A lower hallucination rate on one benchmark does not mean the model is reliable on your documents.
• Price cuts are part of a longer trend: a daily digest called this OpenAI's fourth consecutive price cut amid pressure from Chinese open-weight providers, though that framing comes from the digest, not from OpenAI.

Cheaper models only help if they fit your task. LumiChats lets you run 40+ premium AI models side by side in one subscription, so you can compare a budget model like Terra against bigger rivals on your own questions before deciding.