GPT-6.1 Terra vs Claude Sonnet 5.5: Coding Tests, Cost, and Speed
Prefer to watch? Watch on YouTube · Subscribe for new builds · Read the transcript
OpenAI just released GPT-6.1 Terra at $1 in and $5 out per million tokens. That's half the price of Claude Sonnet 5.5 and Gemini 4 Argon. In this video I cover what OpenAI actually shipped, what the independent numbers say and what people built in the first 24 hours. I also ran three head-to-head coding agent tests of my own. If you build apps with AI and want to know whether Terra is worth switching to, this is for you.
What I cover in this video
- 0:00: The same Star Fox prompt given to Terra and Sonnet. One took 70 seconds and the other took 12 minutes.
- 0:47: What GPT-6.1 Terra is, where you can use it and how it's priced
- 2:18: OpenAI's own benchmark claims and what Artificial Analysis found
- 4:36: The twist: what OpenAI says Terra is actually for
- 5:21: What people built with it in the first 24 hours
- 7:39: How I set up the test bench, and what went wrong with Argon
- 8:25: Test 1 (Star Fox), Test 2 (the whale) and Test 3 (the bug hunt)
- 12:19: The verdict and who should switch
What GPT-6.1 Terra is and what it costs
Terra came out on October 26th. It sits in the middle of OpenAI's lineup. Luna is the cheap model, Sol and Astra are the heavy hitters, and Terra is the everyday model between them. It's available in ChatGPT on the Plus, Pro, Enterprise and Edu plans, in Codex and through the API. Free users don't get it. It was also in GitHub Copilot and Microsoft Foundry on day one.
- Price: $1 in and $5 out per million tokens. That's half of GPT-6.1 Sol, Claude Sonnet 5.5 and Gemini 4 Argon, which all cost $2 in and $10 out. Argon's intro price ends November 30th, and then it goes up to $4 and $20.
- Context: about a 1 million token context window. Watch out, though: once a request goes over 272K input tokens, the whole request is billed at 2x input and 1.5x output.
- OpenAI's claims (vendor reported): 50.3% on Terminal-Bench 4.0, up from 21.5% for the last Terra. OpenAI also says it finishes tasks over 3x faster than Sol and Astra.
The independent numbers and the twist
Artificial Analysis ran its own tests. On its Intelligence Index, Terra scored 48, which puts it in 25th place. Sonnet 5.5 scored 56 (#2 of 221), Argon scored 53 (#9) and Sol scored 52. That means the rumor that Terra would beat Sol didn't hold up. Terra lands four points behind its own big brother.
Cost is where it gets interesting. Running the index cost about $0.40 per task with Terra, $1.62 with Argon and $5.46 with Sonnet, so Sonnet is roughly 14x more expensive per task. The sticker price isn't the only reason. Sonnet used about 420 million tokens to get through the index, while Terra used 130 million and Argon used 110 million. On raw output speed, Sonnet is actually the fastest at about 140 tokens per second, compared with about 90 for Terra and about 80 for Argon. So when OpenAI calls Terra fast, they mean it finishes sooner, not that it types faster.
The twist is that OpenAI never claimed Terra beats Sol. Here's what they actually said:
For in-depth research or complex coding, choose GPT-6.1 Sol or GPT-6 Astra.
OpenAI pitches Terra as "fast and capable computer use at a lower price." It's built to click around your screen, fill out forms and drive apps cheaply. OpenAI's own safety card also rates Terra's cyber skills as "critical." In one of OpenAI's safety tests, when Terra hit a barrier it found a way around it in about 30% of runs. I think a fast, cheap computer use model is a smarter product than the rumored cheaper Sol.
In the first 24 hours, most posts were about speed. One person had Terra draw Sam Altman stroke by stroke in a paint app in about 1:40. Someone else ran Terra against Sol on a long web form, and Terra finished in 53.5 seconds while Sol took 3:21. People also built cars in Blender, landing pages and small games. Not everyone was impressed, though. One creator posted a whale scene with the caption "GPT-6.1 Terra is fast. That's about it." So I ran the whale test myself.
My three tests
Every model got the exact same prompt, one shot, with no help and no follow-ups. Each one ran as a coding agent in a fresh, empty folder on high effort with web search off. Terra ran in Codex on my ChatGPT plan, and Sonnet ran in Claude Code on my Claude plan. I don't pay per token on those plans, so the prices below are what the same run would cost on the API. (For more on the two harnesses, see Claude Code vs Codex.)
Argon didn't work in my setup. In Codex through OpenRouter, every run crashed with an error from Google's API (4 of 4). In Claude Code through OpenRouter, it stalled on both builds for 24 and 30 minutes without saving a file. That's most likely our setup and not the model, so Argon sat out the builds. It did finish the bug hunt.
Test 1: Star Fox
The prompt asked for a simple, beautiful Star Fox-style space shooter in one HTML file: a low-poly ship flying through a canyon and into space. Arrow keys steer, Space fires, Shift boosts and B does a barrel roll.
- Terra: 1:10, about $0.04, about 6K output tokens. It's plain, but it plays.
- Sonnet: 12:04, $1.43, about 86K tokens. It had canyon walls, sun glare, glowing lasers and rocks flying at you, so it looks like an actual game.
Test 2: The whale
The original poster never shared their prompt, so I wrote my own: a humpback whale breaching out of the ocean at sunset in 3D, with waves, a big splash, foam rings and a slow cinematic camera.
- Terra: 0:54, about $0.04. The ocean, sunset and foam rings are all there, but the whale looks like a paddle with fins.
- Sonnet: 22:05, $2.93, over 140K tokens. The light on the water and the spray look good, and the whale actually looks like a whale. It took about 25x longer and cost about 80x more.
Test 3: The bug hunt
Last time every model aced this test, so I made it much harder. It's a shared expenses app of about 740 lines across three files, with eight planted bugs and three decoys. Some bugs are obvious, like page one skipping the 20 newest expenses. Others are nasty: 19.99 gets saved as 19.98, anyone can delete another group's expenses, search can dump every group's data, and a failed delete still says "expense deleted." Each bug found and fixed is worth 2 points, and each made-up bug costs 1 point, for a max of 16. Two third-party models, Qwen and DeepSeek, judged the answers blind, and they gave the same scores.
- Terra: 16/16 in 37 seconds, about $0.03
- Sonnet: 16/16 in 74 seconds, about $0.23
- Argon: 13 in about 5 minutes, $0.39. It missed the race bug where switching groups quickly shows the wrong group's expenses. It also lost a point for flagging a CSS flaw that turned out to be real and that we left in by accident, so that point is on us.
Key takeaways
- If you're building something visual, like a game, a 3D scene or a landing page that has to look great, Sonnet 5.5 is still clearly better. It won both build tests.
- Terra tied Sonnet on the bug hunt in half the time for about a seventh of the price. My whole bench cost about 11 cents with Terra and about $4.60 with Sonnet.
- OpenAI itself says to use Sol or Astra for complex coding. Terra is meant for fast, cheap everyday work and computer use.
- I'd use Terra for the boring stuff in my reselling business: filling out forms and cleaning up listings. Opus 5.5 is still my daily driver for coding.
- One run per test is a vibe check, not a verdict.
FAQ
How much does GPT-6.1 Terra cost?
GPT-6.1 Terra costs $1 per million input tokens and $5 per million output tokens. That's half the price of GPT-6.1 Sol, Claude Sonnet 5.5 and Gemini 4 Argon. If a request goes over 272K input tokens, the whole request is billed at 2x input and 1.5x output.
Is GPT-6.1 Terra better than Claude Sonnet 5.5 for coding?
Not for visual builds. Sonnet clearly won my Star Fox and whale tests. On a hard bug hunt, though, both models scored a perfect 16/16, and Terra did it in half the time for about a seventh of the cost.
Does GPT-6.1 Terra beat GPT-6.1 Sol?
No. On the Artificial Analysis Intelligence Index, Terra scored 48 and Sol scored 52. OpenAI itself recommends Sol or GPT-6 Astra for in-depth research or complex coding.
What is GPT-6.1 Terra best used for?
OpenAI pitches Terra as fast, low-cost computer use, meaning clicking around your screen, filling out forms and driving apps. It's a good fit for lots of small jobs where speed and price matter more than polish.
Chapters
Transcript
Read the full transcript
0:00So, this is the exact same prompt given to two different AI models. Build me a Star Fox game. One shot, no help. On the left is GPT 6.1 Terra, the brand new model from OpenAI. On the right, Claude Sonnet 5.5. Terra built its game in 1 minute and 10 seconds, about 4. Sonnet took 12 minutes and $143. And you can see the difference. Sonnet's Canyon has the sun glare and the glowing lasers. Terra's looks like a rough draft, but it's a rough draft in 70 seconds. And there was supposed to be a third model in this video, Gemini 4 Argon, and Argon could not get a single game saved. Not one. I'll show you exactly what happened. So, the rumor going into this
0:47launch was that Terra would beat OpenAI's own GPT 6.1 Soul for less money. And in this video, we're going to be breaking it all down. what OpenAI actually shipped, what the independent numbers say, what people built with it in the first 24 hours, and our own head-to-head tests, and if it's even worth using. Let's jump right into it. So, what is GPT 6.1 Terra exactly? It dropped on October 26th. It's the middle of OpenAI's lineup. Luna is the cheap one, Soul and Astra are the heavy hitters, and Terra is the everyday model in between. You can use it in chat GPT on the plus pro enterprise and edu plans in codeex and through the API. Free users don't get it. And it was in GitHub
1:33copilot and Microsoft foundry on day one. Now the price $1 in, $5 out per million tokens. That's exactly half of Soul. And it's half of Sonnet 5.5 and half of Gemini 4 Argon, which are both $2 in, $10 out. At least for now. Google's Argon price goes up to 4 and20 on December 1st. It has about a 1 million token context window. But here's something to be wary about. Once you send it more than 272,000 tokens, the whole request costs double on the input and 1 and a half times on the output. And here are OpenAI's numbers. On Terminal Bench 4.0, which is basically can it do real coding agent work in a terminal? Terra scores 50.3%.
2:18The last Terara scored 21 12. And OpenAI says it finishes tasks over three times faster than Soul and Astra. These are OpenAI's numbers vendor reported. So before I show you who actually checked them, real quick, every video I make is built in code with Claude Code inside Shipping School Premium. I show you how step by step to turn your product into videos that get views. plus new courses every day, 500 plus hours of replays, live calls every week, and a Q&A with me every Saturday. Join free link below. So, Artificial Analysis ran their own tests, and on their intelligence index, Terra scored a 48. That's 25th place. Now, look at the other two in our matchup. Sonnet 5.5 scored a 56. That's
3:06second place out of 221 models. Gemini 4 Argon 53, ninth place. So Terra comes in last of our three, eight points behind Sonnet. And remember that rumor about beating Soul? Soul scored a 52. So Terra is four points behind its own big brother. But here's the part that actually matters for your wallet. To run that whole index, it cost about 40 cents a task with Terra. Argon 162. Sonnet 5.46. So, Sonnet is about 14 times more expensive per task. Look at that cost comparison. Absolutely insane. And here's why. It's not just the sticker price. Sonnet used 420 million tokens to
3:51get through the index. Terra used 130 million. Argon 110. So, Sonnet writes over three times as much as Terra. Remember that. It's going to come up in our tests. On raw speed, Sonnet is actually the fastest of the three. about 140 tokens a second. Terra is about 90 and Argon about 80. So when OpenAI says Terra is fast, they mean it finishes sooner, not that it types faster. So here's the twist with this launch. The information reported that Terra would beat Soul for less money. Everybody ran with it and then OpenAI shipped it. And OpenAI never said that. What OpenAI actually said is this quote for in-depth research or complex coding, choose GPT
4:366.1 Soul or GPT6 Astra. So OpenAI is telling you don't use this for your hardest work. So what is it for? Computer use. OpenAI's pitch is quote fast and capable computer use at a lower price. And Sam Alman called it a new era of computer use. So, this thing is built to click around your screen, fill out forms, drive apps, and do it cheap. And here's a weird one. In its own safety card, OpenAI rates Terara's cyber skills as critical. That's OpenAI's own label. And in one of OpenAI's safety tests, when Tara hit a barrier in the task, it found a way around it in about 30% of runs. So, it's not a cheaper soul. It's a fast, cheap computer use model. And honestly, I think that's a smarter
5:21product than what the rumor said. So, what are people actually doing with it? And the first day was all about speed. The most viewed Terra Test on X so far, over 130,000 views. Someone told it, "Make an image of Sam Alman on this paint app." And it drew him stroke by stroke. And he said, "This is not sped up at all." It took about a minute and 40 seconds. That's one run, but still. This one is wild. Same long web form, Tara versus Soul. In his first run, Tara finished in 53 and a half seconds. Soul took 3 minutes and 21 seconds. And in the recording he posted, you can see Soul is still signing the name while Terra is done. People built cars in
6:06Blender, a Viking axe in under 5 minutes, landing pages, little games, and the cost comparisons are where people lost it. One person built the same red supercar in Blender with Terra and with Sonnet 5.5. Terra 10 minutes and about 31. Sonnet an hour and 26 minutes and $15.20. That's his one run, not ours. But look at it. But not everyone was impressed. Artificial Analysis posted its number about 5 minutes after launch and the doubters ran with it. Terra scores 48 on the AA intelligence index. I can hear Daario laughing right now. Another person said, "It's somehow even behind Sonic 5.5. What's the point of this release?" And this one stuck with me.
6:53Someone gave it a whale breaching out of the ocean next to Soul and Astra. And his caption was, "GPT 6.1 Terra is fast. That's about it. Remember this whale. We're going to run it ourselves." The best take I saw was from a Japanese creator. If you use it like a top model, it's very disappointing. But it's not meant to be a cheap stand-in for the top models. It's built to do everyday computer tasks, fast and cheap. So, which is it? I wanted to find out for myself. So, we ran our own tests, three of them. GPT 6.1 Terra versus Clog Sonnet 5.5 versus Gemini 4 Argon. Same exact prompt for every model. One shot, no help, no follow-ups. And every model ran as a coding agent in a fresh empty
7:39folder on high effort with web search off. Terra ran in OpenAI's codeex on my chat GPT plan. Sonnet ran in Claude code on my cloud plan. And since we don't pay for those per token, the prices you'll see for them are what the same run would cost on the API. And Argon is where it got messy. We first ran Argon in codeex through open router and every single run crashed with an error from Google's own API four for four. So we moved it into cloud code through open router. And on both build tests, Argon started and then it just stalled. On Starf Fox, it wrote one sentence saying what it was about to build and that was it. 24 minutes on one, 30 on the other, and it never saved a file. Now, to be fair, that's most likely our setup, not the model.
8:25Google's API and these coding agents just didn't get along. So, I'm not calling it a loss for Argon on the builds. It just sits those two out. It did finish the bug hunt, though, and you'll see that. And one run each is a vibe check, not a verdict. Keep that in mind. Test one. Build a simple, beautiful Star Fox style space shooter. One HTML file. A low poly ship flying through a canyon and into space. Arrow keys steer. Space fires lasers. Shift boosts. and B does a barrel roll. So, here's Terra. 1 minute and 10 seconds, about 4 cents. You can see we've got the ship, the canyon, the rings, enemies coming in, and it plays with the same controls we gave every model. It's plain, but it plays. And here's Sonnet. 12 minutes and 4 seconds, $143. And look
9:13at this. The canyon walls, the sun glare, the lasers glowing, rocks flying at you. This looks like an actual game. And every game got the exact same inputs. Same moves, same shots at the same time. And look at how much each one wrote to get there. Terra about 6,000 tokens. Sonnet about 86,000. 14 times as much. Same story as artificial analysis. Which one looks better to you? Comment down below because I want to know. Test two. Remember that whale post? The author never shared the prompt, so we wrote our own. A humpback whale breaching out of the ocean at sunset in 3D. Waves, a big splash, foam rings, a slow cinematic camera. Terra, 54 seconds, about 4
10:00cents. And you can see we've got the ocean, the sunset, the whale comes up out of the water, and the foam rings spread out. But the whale looks like a paddle with fins. Honestly, that guy on X was right. It's fast, and that's about it. And then Sonnet, 22 minutes, $2.93, over 140,000 tokens. And look at this. The light on the water, the spray. The whale actually looks like a whale. It's not even close. So sonnet took about 25 times longer and cost about 80 times more. And it's clearly the better whale. So the question is, is a better whale worth 80 times the price? For a landing page? Maybe. for a hundred little jobs a day? Probably not. Test three, the bug
10:47hunt. And last time, this test was too easy. Every model got a perfect score. So, we made it way harder. It's a shared expenses app. About 740 lines across three files with eight bugs hidden in it on purpose. Some are obvious. Page one skips the 20 newest expenses, and some are nasty. 1999 gets saved as 1998. Anyone can delete another group's expenses. The search box can dump every group's data. And a failed delete still says expense deleted. And we added three decoys. Code that looks wrong but isn't. Flag one and you lose a point. Two points for every bug found and fixed, right? Minus one for every bug it makes up. 16 is a perfect score. And it was
11:34scored by two thirdparty models, Quen and Deepseek. Neither one is from OpenAI, Anthropic or Google. And they never saw which answer came from which model. So the result, Terra 16 out of 16 in 37 seconds for about 3 cents. Sonnet also 16 out of 16. 74 seconds about 23 cents. Both found all eight. Neither one fell for a decoy. And Argon scored 13. It took about 5 minutes and cost 39. It missed the race bug where switching groups fast shows you the wrong group's expenses. And it lost a point for flagging something that wasn't one of our eight. But here's the thing. We checked and that one is a real bug that we left in by accident. Sonnet noticed
12:19it, too. So that point is on us, not Argon. And both judges gave the exact same scores. So on finding bugs, Terra tied Sonnet in half the time for about a seventh of the price. So, what does this actually mean as of today? Who should switch? If you're building something visual, a game, a 3D scene, a landing page that has to look amazing, Sonnet 5.5 is still better. It's not close. Both of our build tests went to Sonnet. But if you're running lots of small jobs or you want an agent clicking around your computer, Terra is really interesting. It tied Sonnet on the bug hunt. It finished every test in about a minute. And our whole bench cost about 11 with Terra versus about $4.60 of
13:05Sonnet. Here's how I'd use it in my reselling business. The boring stuff. Filling out forms, cleaning up listings, the jobs where I need it done fast and cheap, and I don't need it to be beautiful. And Argon, I want to give it a fair shot. On artificial analysis, it's ahead of Terra. But until it plays nice with the coding agents I actually use, it's hard for me to recommend for this kind of work. And for my own coding, Opus 5.5 is still my daily driver. Terra didn't change that. But for the price, I'm going to keep it around. So, which one are you putting to work? Terra or Sonnet? Tell me below. And if this saved you a few bucks on your API bill, tap the like. That's how this one gets in front of other people. Have a blessed day.
Ready to start building with AI?
Join the Shipping Skool community free. Paid members get live bootcamp calls every week and weekly-updated courses.
Join Shipping SkoolKeep reading
Claude Opus 5 vs Fable 5: Cost, Speed, and 3 Real Build Tests
I ran Claude Opus 5 and Fable 5 on high effort through three builds: a Three.js MMO, a Remotion video, and a landing page rebuild. Opus 5 won the game and the UI test, the video was close to a tie, and Opus cost about half as much. It's my new daily driver, but Fable 5 is much faster.
Gemini 4 Carbon vs Claude Opus 5.5: Benchmarks and Speed Test
Gemini 4 Carbon tied Claude Opus 5.5 on both of my hidden-test builds (12/12 and 28/32 each), but in my runs through OpenRouter it took up to 7.6 times longer and twice saved no file. Carbon costs $2 in / $10 out until November 30, then the same as Opus.
Gemini 4 Argon vs Claude Opus 5.5: Benchmarks, Pricing, Real Tests
Gemini 4 Argon is cheap at launch ($2 in / $10 out per million tokens until November 30) and strong at vision, but independent testers rate it below Claude Opus 5.5, and it struggles with real coding work. My pick for building is still Claude Code. If you want to try Argon, keep it in a separate git worktree.
