GPT-6.1 Terra: Half the Price of Sol, Yet Not Better Value in Most Benchmarks

October 26, 2026, evening

OpenAI launches GPT-6.1 Terra: a model for working on a computer, at half the price of GPT-6.1 Sol. Its post on X closes in on 4.9 million views.

The release notes slip in something else, in one small line: some Plus and Pro subscribers may now see Terra as the default in ChatGPT Work, instead of Sol at the Light reasoning setting.

So I asked the simple question: what does it cost to get the same result from Terra as from Sol?

GPT-6.1 Terra is OpenAI's new model, released on October 26, 2026, and it sits between the small GPT-6 Luna and GPT-6.1 Sol. We took the charts OpenAI itself published and pulled out the cost and score at each of the five reasoning settings shown. That let us calculate something the launch doesn't show: what each model costs when you ask it for the same level of quality. We added independent measurements and developer tests.

One term that keeps coming up: reasoning level. Every OpenAI model has a "dial" that sets how much it thinks before answering. The levels are Low, Medium, High, Xhigh and Max (there's also None, with no reasoning at all), and Medium is the default. More reasoning means more quality, more tokens, more money and more time. When I write "Terra Max", I mean Terra at maximum reasoning.

In short

GPT-6.1 Terra costs $1 per million input tokens and $5 per million output: half of GPT-6.1 Sol.
At the same setting it's faster and cheaper. But to match Sol's quality it has to think much harder, and then in most benchmarks Sol is cheaper and faster.
The exception: on OpenAI's test of long professional on-screen tasks, Terra gets close to Sol's top quality at a fifth to a seventh of the cost.
For a business agent running on OpenAI: Sol at Medium remains the pick for most work. Terra at Medium is worth testing on simple, high-volume tasks, like entering invoices.

My take, at a glance

×7Cheaper on long professional tasksOn OpenAI's test: nearly Sol's top quality, at a seventh of the cost

×1.8Pricier at codeTerra at Max costs more than Sol at Medium, and scores lower

×2.5Slower at the same quality385 seconds per task vs. 151 for Sol, at an identical score

PlansWho really gainsChatGPT subscribers, whose quota goes further. And OpenAI

My take: a good model, a confusing launch

Terra isn't a bad model. It's just not a cheap substitute for Sol. Peter Gostev, who tests models, wrote that it feels like a "slightly simpler 6.1-Sol", and that's a big leap from the previous model in this tier: GPT-5.6 Terra got 22.3% of hard questions factually wrong; the new Terra, 4.3%. But price per token isn't what you end up paying. You pay per task, and Terra's task runs longer when you ask it for the same quality.

Picture two couriers. One charges $10 a mile, the other $5. Easy call. But the second one barely knows the city, and to reach the same address he drives three times the distance. That's Terra on most tasks. Only in one neighborhood, professional on-screen work, does it know shortcuts Sol doesn't.

What GPT-6.1 Terra is, in a minute

$1 / $5per million input and output tokens. Sol: 2 and 10. Opus 5.5: 4 and 20

1Mtokens at once, and up to 128K tokens in a single answer

91tokens per second. Sol: 55 to 77, depending on reasoning level

Plus and upPlus, Pro, Business, Enterprise and Edu, in ChatGPT Work and Codex, plus the API. Not on the free plan

Sources: OpenAI's launch page, model page and pricing page; speed per Artificial Analysis, October 27, 2026. We explain tokens in What is a token.

According to OpenAI, it was built for the work that fills a workday: reading what's on screen, clicking and filling in fields in apps, drafting quotes and reading drawings. What it means for a model to "use a computer", we explain in What is Computer Use. And OpenAI itself writes that for the most complex coding tasks, Sol and Astra remain the better choice.

GPT-6.1 Terra vs. Sol: which is cheaper at the same quality

OpenAI compared Terra and Sol at the same setting. But if Terra at Medium delivers less than Sol at Medium, that comparison isn't fair. So for each benchmark, we worked out the cheapest setting at which each model reaches the same score.

The four benchmarks: Agents' Last Exam tests long professional tasks on screen. OSWorld tests tasks in everyday desktop software, like files, a browser and a spreadsheet. Terminal-Bench tests writing code and running commands, the work an agent does when it fixes a website. And the Artificial Analysis index averages ten general benchmarks, run by an independent company.

Same quality, who's cheaper: four benchmarks

Professional computer workTerra Medium: 57.6% at $0.29. Sol needs Max to beat that: 58.6% at $2. Terra is 5 to 7x cheaper

Desktop softwareSol Medium: 73.5% at $0.65. Terra needs Max: 74.2% at $1.19. Sol is about 45% cheaper

Code and server commandsSol Medium: 52.7% at $1.02. Terra Max: 50.3% at $1.84. Sol is 1.8x cheaper, and better

Independent general indexBoth score 48: Sol Medium at $0.21, Terra Max at $0.40. Sol costs nearly half as much

First three: Platin Digital's calculation from the charts on OpenAI's launch page (Agents' Last Exam, OSWorld 2.1, Terminal-Bench 4.0), estimated cost per task. Last: Artificial Analysis. Green: Terra is better value. Red: Sol.

Three out of four go to Sol. And the one benchmark Terra wins big, Agents' Last Exam, is the one OpenAI chose to show first. That doesn't make the number wrong. It means it holds for one kind of work: a long professional task on screen, like turning an architectural drawing into a 3D model. That's what Terra was built for.

One caveat: according to an OpenAI employee, the margin of error on these benchmarks is about 5 points either way, so a one-point gap settles nothing. I trust the trend, because it repeats across every benchmark.

Code and server commands: cost per task at five selected settings

Sol Low · 47.1%

$0.76

Sol Medium · 52.7%

$1.02

Terra Medium · 35.8%

$0.46

Terra Xhigh · 46.4%

$0.99

Terra Max · 50.3%

$1.84

Source: the Terminal-Bench 4.0 chart on OpenAI's launch page. The percentage next to each row is the score. Terra at Max costs more than Sol at Low, and still scores below Sol at Medium.

Is GPT-6.1 Terra faster than Sol? At the same setting, yes. At the same quality, no

That's exactly what I wondered: Terra generates text faster, but burns more tokens to reach the same result. Which wins? Artificial Analysis, which measures models independently, timed each task:

Average time per task, in seconds, and index score

Terra Medium · score 41

Sol Medium · score 48

151

Terra Max · score 48

385

Sol Max · score 52

496

Source: Artificial Analysis, October 27, 2026. The second and third rows have exactly the same score.

Same setting

Medium vs. Medium. At Max vs. Max, Terra is also faster and cheaper: 385 seconds and 40 cents, vs. 496 and 72

Terra is faster: 63 seconds vs. 151
And half the cost: 10 cents vs. 21
But the result is weaker: 41 vs. 48

Same quality

Terra Max vs. Sol Medium, both at 48

Terra is 2.5x slower: 385 seconds vs. 151
And nearly 2x pricier: 40 cents vs. 21
Because at max reasoning it writes far more

The extra tokens eat up the speed. On the independent AI Benchy test at Max, Terra wrote about 391K tokens vs. about 64K for Sol, six times as many. It cost $2.21 vs. $1.01, ran slightly slower and scored lower: 9.4 vs. 9.9. At Medium the picture flips: the same score, 9.6, at about half the price.

Real tests: what developers found

Four tests, and the results

One long request, three modelsSol: $3.62, 43 min. Terra: $4.90, 43 min. Opus 5.5: $31.85, 2 h 45 min

A car in BlenderTerra: 10 min, about 31 cents. Sonnet 5.5: 1 h 26 min, $15.20

An island with windmillsTerra: 5 min, about 32 cents. Opus 5.5: 1 h 41 min, $30.23

43 data-analysis questionsTerra: 40 correct, $1.31. Sol: 43, at $2.61

Green: Terra is better value. Red: worse. Orange: depends on what matters to you. Sources: EnvolDev and MrTacticalX on X, and Plotly, October 26–27, 2026.

Against Sol on the same task: Terra costs more. A developer ran the same request on three models. Both finished in 43 minutes, and Terra cost $4.90 vs. $3.62 for Sol. His headline raves about Terra, but only because he was comparing it with Opus.

EnvolDev on X: Opus 5.5 vs. GPT-6.1 Sol vs. GPT-6.1 Terra, October 26, 2026

Look at Sol's row: same time, less money.

Against Claude in Blender: a huge gap. Here Terra does what it was built for: driving a graphics program on its own. A car at night in Miami: 10 minutes and about 31 cents, vs. an hour and a half and $15 for Sonnet 5.5. The developer didn't compare it with Sol here.

MrTacticalX on X: GPT-6.1 Terra vs. Claude Sonnet 5.5 in Blender, October 26, 2026

Both models at High. His comparison, not ours.

The view from someone who used it. Peter Gostev called it the "first non-frontier model" he doesn't mind using, one that feels like a "slightly simpler 6.1-Sol". His advice: if you're on the $20 Plus plan, stick with Terra, because your quota will go further.

Peter Gostev on X: first impressions of GPT-6.1 Terra, October 26, 2026

"I'm not gonna tell you it will change your life", but at half Sol's price, it's a good option.

Why OpenAI launched it at all

You can read the launch two ways, and both are true:

What OpenAI says

Fast computer work at half the price
High-volume automation, like entering invoices
The plan's quota covers more tasks than with GPT-5.6 Terra
Fewer factual errors than the previous version

What else it gains

Some subscribers get Terra instead of Sol as the default
Each subscriber task costs it less
A middle model that stops people dropping to Luna or switching to rivals

Sources: OpenAI's launch page and the ChatGPT release notes, October 26, 2026: "Some Plus and Pro users may see GPT-6.1 Terra at Medium instead of GPT-6.1 Sol at Light in the default Power presets in Work." Overall usage limits haven't changed.

Even OpenAI isn't sure it will last. Ted Sanders, an OpenAI employee, wrote on Hacker News that Terra is "similar on vision and computer use (but much faster!)" to Sol and Astra, and a bit worse at most other things. And: "if traction is poor, maybe we'll go back to 3 [models]."

When to switch on Terra in your business agent

Terra, at MediumWorth testing: the cheapest, fastest in the family

Entering invoices into software
Filling in forms
Reading screenshots and documents
A Plus plan that keeps hitting its limit

✓

Sol, at MediumThe default for most work

Website fixes
Data analysis and reports
Desktop software
Multi-step tasks

Not Terra at MaxYou pay more and wait longer

Sol at Medium comes out cheaper and faster
For complex code: Sol, Astra or Claude

My rule: if Terra at Medium gets the job done, it's the cheapest and fastest option in OpenAI's lineup. The moment you need to push it to Max to succeed, switch to Sol. And for tasks where it drives a graphics program itself, like the Blender tests, it's worth trying first.

If your agent already runs on Sol at Medium, there's no reason to switch. If it runs on Sonnet 5.5 or Opus 5.5 and operates software on the computer, Terra is worth testing: on OpenAI's charts it costs about half as much as Sonnet at Medium for the same level, and against Opus on long tasks the gap reaches 9x. But on desktop software, Sonnet at max reaches 83.9%, while Terra stops at 74.2%. Test it on a few of your real tasks, not on a benchmark.

One privacy note: according to a developer on Hacker News, OpenRouter, a service that connects you to many models through one account, doesn't yet have a Terra provider that commits not to retain the data you send, unlike Sol. If you work directly with OpenAI, this doesn't affect you.

What I'm doing with it

I usually build agents on Hermes Agent, an open-source framework from Nous Research, and pick the model that fits each business: OpenAI, Anthropic, Google and others. Platin Digital is an OpenAI Select Partner, and that doesn't change the choice. Terra goes into testing on simple data-entry and form tasks. In agents already running on Sol, it doesn't replace it.

In our ranking, the Sigma AI Ranking, GPT-6.1 Terra now appears in the "New in the ranking" strip, and will get a place once the open independent measurements are in. Also from this week: what OpenAI actually proved in math, and Kimi K3.1 Pro, the Moonshot model whose weights are due to open in early November. What an agent costs beyond the model price, we break down in How much an AI agent costs a business.

Frequently asked questions

What is GPT-6.1 Terra?

GPT-6.1 Terra is an OpenAI model released on October 26, 2026. It sits between GPT-6 Luna, the small, cheap one, and GPT-6.1 Sol, and was built specifically for computer work: reading the screen, operating apps, filling in forms and writing. It's available to Plus, Pro, Business, Enterprise and Edu subscribers in ChatGPT Work and Codex, and in the API as gpt-6.1-terra.

How much does GPT-6.1 Terra cost?

In the API: $1 per million input tokens, 5 cents for cached input, and $5 per million output tokens. That's half the price of GPT-6.1 Sol and a tenth of GPT-6 Astra. On a ChatGPT plan there's no extra charge. According to OpenAI, the quota covers more tasks than with GPT-5.6 Terra, and people who hit the limit with Sol will benefit most.

Is GPT-6.1 Terra better than GPT-6.1 Sol?

At the same reasoning setting, Terra is cheaper and faster but weaker. To match Sol's quality it needs a high setting, and then in most benchmarks Sol is cheaper and faster. By our calculation from OpenAI's charts, on code Sol at Medium is 1.8x cheaper than Terra at Max and still scores higher, and on desktop software Sol is about 45% cheaper at the same level. The exception is long professional on-screen tasks on Agents' Last Exam, where Terra is 5 to 7x cheaper.

Is GPT-6.1 Terra faster?

It generates 91 tokens per second, more than Sol. At the same setting it finishes a task faster: 63 seconds vs. 151 at Medium, according to Artificial Analysis. But at the same quality it's slower: Terra at Max needs 385 seconds to reach the score Sol at Medium achieves in 151.

Who is GPT-6.1 Terra for?

ChatGPT subscribers who keep running out of quota, simple high-volume tasks like entering invoices and filling in forms, and agents that operate graphics software and until now ran on expensive Claude models. For complex code, OpenAI itself recommends Sol or Astra.

What's the difference between Luna, Terra, Sol and Astra?

These are OpenAI's four tiers, from cheapest to priciest: Luna ($0.10 and $0.50 per million tokens), Terra (1 and 5), Sol (2 and 10) and Astra (10 and 50). An OpenAI employee summed it up: if you care only about quality, Astra; if you're a bit price-sensitive, Sol; if you're very price-sensitive, Terra or Luna.