GPT-6.1 Terra only pays off at half throttle

OpenAI released GPT-6.1 Terra on October 26, its first mid-tier model since GPT-5.6 Terra. I’m writing about it because two days in, enough independent numbers have piled up to compare it with the cheap models I covered here recently: Anthropic’s Haiku 5.5 and DeepSeek V4.1 Flash 1017. Also, Artificial Analysis changed its index the day after Terra’s release, and that shuffled the rankings quite a bit.

What Terra is

Half the price of its predecessor. A million input tokens cost $1, output tokens $5, and cached input $0.05. Batch processing comes out at half price. The context window is a million tokens, and the model accepts text and images as input. Above 272K input tokens, the whole request is billed at a higher rate (input at double, output at 50% more).

Where to find it. In the API under the name gpt-6.1-terra, and in ChatGPT Work and Codex for Plus, Pro, Enterprise and Edu subscribers. If you’ve seen “GPT-6.1 Terra Pro,” that’s just a label from catalogs and pricing plans; there is no separate model by that name.

What OpenAI intends it for. Computer use, visual reasoning and “professional tasks” in general. For complex programming, OpenAI itself recommends the more expensive Sol or Astra.

The jump over its predecessor is big. In the original version of the Artificial Analysis index, Terra added six points at a third of the cost per task. And the hallucination rate (how often a model answers when it doesn’t know the answer) dropped from 88 to 38 percent.

Watch out, the yardstick changed

On October 27, Artificial Analysis released version 5.0 of its index. It dropped three tests, including the office-work GDPval-AA, and added new tests for programming (FrontierCode 2.0) and scientific work in the terminal (Terminal-Bench-Science). Science now makes up a quarter of the index. And that is exactly where the cheap models fall flat.

Terra improved from 48 to 51 points. Haiku 5.5 fell from 43 to 36, DeepSeek 1017 from 45 to 32. The models are the same; only the tests changed. An index is only as good as it resembles your own work (and few people do scientific tasks in the terminal). I’m going with version 5.0 from here on, because it’s the current one and relies more on private tests that can’t be trained on.

Independent measurement: the effort level decides

The index is version 5.0 from Artificial Analysis.

From low to high, Terra wins. At medium it scores 38 points for 17 cents (about four Czech crowns), more than Haiku and DeepSeek at max, for less than half their price. At high it scores 42 points for roughly six crowns. At low it scores 29 points for 7 cents, which according to Artificial Analysis is the lowest cost per task in its price class, and it beats even Luna at max for half the price.

From xhigh up, take Sol. Sol at medium scores a point more than Terra at xhigh and costs a tenth less. Sol at xhigh scores a point more than Terra at max and costs a fifth less. So at the top levels, Terra competes with its own older sibling for customers, and loses.

The triple speed holds only in OpenAI’s own test. The vendor promises that Terra finishes a task more than three times faster than Sol and Astra. In Artificial Analysis’s measurements, a task at medium takes two minutes, and Sol at medium takes four. At max, Terra takes 11 minutes, longer than Haiku and DeepSeek, and you wait almost 80 seconds for the first token.

How it differs from Haiku and DeepSeek

Source: Artificial Analysis (index 5.0) and Vals AI, as of October 28.

A token ten times more expensive, a task twice as expensive. Terra is stingy with words. It outputs 60K tokens per task, while Haiku outputs almost four times as many. That’s why the tenfold per-token price shrinks to only about double the cost per task at max. And at medium, Terra is cheaper than both rivals at max.

It clearly leads in the terminal and in programming, but only at max. On Terminal-Bench 4.0, Artificial Analysis measured 56 percent, even more than OpenAI itself reports (50.3). At medium, though, it gets only 32 percent, roughly what Haiku and DeepSeek get at max. Sol at medium gets 48. In Artificial Analysis’s coding agent index, Terra scores 48 points, Sol 54, DeepSeek 46 and Haiku 37.

It lags in office work. This surprised me, because OpenAI sells Terra for professional tasks. In the AA-Briefcase office test, both Haiku and DeepSeek beat it by roughly 170 Elo points. In business process automation, DeepSeek is better and Haiku is significantly worse.

It doesn’t make much up. In the knowledge test, Terra is the best of the three. DeepSeek scores one point, so it isn’t suited to research without source material. Haiku makes things up about as rarely as Terra, but it knows fewer facts.

On Vals, DeepSeek wins at a quarter of the price. In the Vals Index (finance, law, taxes, programming), DeepSeek ranks ninth, Terra tenth and Haiku thirteenth. One test run costs $0.66 on DeepSeek and $2.92 on Terra. On this type of work, DeepSeek fares much better than in the Artificial Analysis index.

So are they competitors? On cost per task, yes; on performance in the index, not really. DeepSeek remains a rival where office and professional work, price and open weights are what matter. Haiku is a cheap, fast helper for subtasks inside a larger agent. But Terra’s toughest competitor is Sol.

Anthropic has nothing to put against Terra. Haiku 5.5 is cheaper and weaker. Claude Sonnet 5.5 scores three points more (54), but a task on it costs $10, ten times more than on Terra, and takes 22 minutes. Curiously, per token ($2 and $10) Sonnet costs only twice as much as Terra. But it outputs 285K tokens per task. Sonnet leads in office work, automation and the terminal, Terra in scientific tasks, and in FrontierCode programming the two are tied. How Sonnet does at lower effort levels, nobody has measured yet.

OpenAI’s table

Agents’ Last Exam is the main argument. In this test of professional tasks involving computer use, Terra at medium scores 57.6 percent, Astra 59.3 and Claude Opus 5.5 63.2. By OpenAI’s estimate, a task costs $0.29 at medium, $1.01 at max, and $8.82 on Opus. That’s where the slogan about a task nine times cheaper than on Opus comes from. But Opus is still three points better, and it’s among the most expensive models on the market, so beating it on price is easy.

Behind Sonnet in computer use. On OSWorld 2.1, Terra scores 74 percent according to OpenAI, Sol 78 and Sonnet 5.5 84. Here a task on Terra even came out slightly more expensive than on Sol.

First impressions from X are mixed. A web demo from DesignArena took four times longer than on the predecessor, yet came out a third cheaper. The author of a pixel-art scene paid $2.38 versus $8.38 on Opus, but liked the Opus result more. None of this can be verified.

The system card has one oddity. According to the system card, in a test with messages from foreign agents, Terra tried to engage with them in 65 percent of cases and performed an unauthorized action in 13.75 percent. Sol 6.1 did so in only 1.25 percent, but the older GPT-6 Sol in 16.25. Anyone who lets an agent read content from other bots should keep approvals turned on with Terra.

Which OpenAI model for what

GPT-6.1 Terra: use it for computer use and bulk agent work with forms, browsers and PDFs at the low to high levels, where it delivers the most performance per crown and doesn’t make much up.

GPT-6.1 Sol: use it for coding agents, terminal work, and anywhere you’d run Terra at xhigh and above, because there Sol is both better and cheaper.

GPT-6 Luna: use it for sorting and routing huge volumes of short queries without reasoning, where the price per token decides. As soon as the model has to think, Terra at low comes out both better and cheaper.

GPT-6 Astra: use it for scientific and research tasks in the terminal, where it leads in independent tests and where it pays off to pay ten times more per token than for Terra.

Bottom line: what to take away

Set Terra to medium or high. That’s where it’s the best deal, and for four to six crowns per task you get more than from Haiku or DeepSeek at max.

For office documents, spreadsheets and presentations, reach for DeepSeek or Haiku. Both are better in the office test, and DeepSeek is significantly cheaper too. DeepSeek also has open weights, so you can run it with a European provider if the data must not leave the EU.

For a cheap helper inside an agent with a short context, Haiku remains the pick. It’s two and a half times faster at generation and ten times cheaper per token; just watch its 100K-token limit.

For the hardest programming, none of these three. Sol, Opus or Sonnet.

For me, Terra is the most sensible mid-tier model OpenAI has released. But it’s being sold for office work, where independent measurements show more of a weakness, and at the highest effort level it mainly competes with OpenAI’s own Sol. Set it to medium and you have a cheap, sober agent that doesn’t babble much and doesn’t make much up.

Sources: OpenAI’s announcement and the GPT-6.1 Terra system card from October 26, 2026; Artificial Analysis measurements (index version 5.0 from October 27; for the predecessor and the hallucination rate, version 4.3.2; Coding Agent Index v2.0); Vals Index; roundups from Kingy, eesel and OrcaRouter; reactions on X as reported by ChainCatcher. OpenAI’s table is measured by the vendor. Terra doesn’t have a Code Arena rating yet. Data valid as of October 28, 2026. This text was written with the assistance of Claude Opus 5.5, that is, a competing model.