GPT-6.1 Terra Review: Half the Price of Sol, and When It Actually Saves You Money

GPT-6.1 Terra is half the price of GPT-6.1 Sol on every line of the price sheet, and an independent test puts it at about fifteen times cheaper per task than Claude Opus 5.5. It also scores ten points lower than Opus on that same scoreboard. Here's when Terra saves you money, and when it burns it.

The promise

Two numbers frame this whole review. On the Artificial Analysis Intelligence Index, Terra costs about $0.40 per task and Claude Opus 5.5 costs about $5.98, roughly a fifteenfold gap. INDEPENDENT · ARTIFICIAL ANALYSIS On the same index, Opus scores 58 and Terra scores 48. INDEPENDENT · ARTIFICIAL ANALYSIS

So you're trading some capability for a lot of price. The rest of this post is about where that trade pays off.

What Terra is

GPT-6.1 Terra came out on October 26, 2026, twenty-seven days after GPT-6.1 Sol shipped on September 29. FIRST-PARTY · OPENAI

OpenAI has four tiers right now, from small to large:

Model	Input / output per 1M tokens
GPT-6 Luna	$0.10 / $0.50
GPT-6.1 Terra	$1 / $5
GPT-6.1 Sol	$2 / $10
GPT-6 Astra	$10 / $50

OpenAI pitches Terra for visual reasoning, precise computer use, and clear, natural writing. FIRST-PARTY · OPENAI

The specs, per OpenAI's model page:

• Context window: 1,050,000 tokens, so about one million.
• Max output: 128,000 tokens.
• Reasoning effort: six levels (none, low, medium, high, xhigh, max). Medium is the default.

The price

Terra costs $1 per million input tokens, $0.05 per million cached input tokens, and $5 per million output tokens. Sol costs $2, $0.10 and $10. Terra is exactly half on every line. FIRST-PARTY · OPENAI

Here's how that compares with the closest competitors, using each vendor's own pricing page:

Model	Input	Cached input	Output
GPT-6.1 Terra	$1.00	$0.05	$5.00
GPT-6.1 Sol	$2.00	$0.10	$10.00
Claude Sonnet 5.5	$2.00	$0.10	$10.00
Claude Opus 5.5	$4.00	$0.20	$20.00
Gemini 4 Argon (through Nov 30)	$2.00	$0.10	$10.00
Gemini 4 Argon (from Dec 1)	$4.00	$0.20	$20.00

Sonnet 5.5 costs the same as Sol, and Opus 5.5 is four times Terra per token. FIRST-PARTY · ANTHROPIC Gemini 4 Argon doubles on December 1. FIRST-PARTY · GOOGLE

The number I'd watch is cached input. Five cents is 5% of the regular input rate, which means 95% off. Agents resend the same context constantly, so cache hits are where the real savings live.

If your job can wait, Batch halves the price again: $0.50 in, $0.025 cached, $2.50 out. FIRST-PARTY · OPENAI

OpenAI's claims vs independent tests

This is the chapter where I put what OpenAI says next to what Artificial Analysis measured.

Terminal-Bench 4.0. OpenAI says Terra hits 50.3% at max effort, against GPT-5.6 Terra's published score of 21.5% (OpenAI doesn't state the effort for that older score). FIRST-PARTY · OPENAI Artificial Analysis ran the benchmark itself and scored Terra even higher, about 56%. It also scored the old Terra higher, at about 35%, so the jump in its runs was about twenty points. INDEPENDENT · ARTIFICIAL ANALYSIS Still a big gain, just smaller than OpenAI's +28.8. The two may use different harnesses or settings.

Errors and guessing. OpenAI also says errors on its internal set of tricky prompts fell from 53.6% to 16.7% at low effort, and to 4.3% at max. FIRST-PARTY · OPENAI That's an internal test, so nobody outside can check it. The closest independent check is the Artificial Analysis hallucination test: when Terra didn't know an answer, it guessed anyway 38% of the time. The old Terra guessed 88% of the time. INDEPENDENT · ARTIFICIAL ANALYSIS

The "9x cheaper" claim. OpenAI says that on Agents' Last Exam at max effort, Terra is more than three times as token-efficient as Opus 5.5, at a quarter of the price per token, for about nine times lower cost per task. FIRST-PARTY · OPENAI OpenAI's announcement text gives those ratios but no actual score.

Artificial Analysis measured cost per task on its own index: Terra $0.40, Opus 5.5 $5.98. That's about fifteen times cheaper. INDEPENDENT · ARTIFICIAL ANALYSIS It's a different test, and it shows an even bigger gap than OpenAI claimed.

The scores. Here's where the stories split. On the Artificial Analysis Intelligence Index:

Model	Score	Rank
Claude Opus 5.5 (max)	58	#1
Claude Sonnet 5.5 (max)	56	#2
Gemini 4 Argon (high)	53	#9
GPT-6.1 Sol (max)	52	#13
GPT-6.1 Terra (max)	48	#25 of 221

INDEPENDENT · ARTIFICIAL ANALYSIS

So you're buying price and giving up some brains.

The catches

There are five.

1. Max effort is slow. Artificial Analysis measured Terra's time to first token at max effort at about 134 seconds. INDEPENDENT · ARTIFICIAL ANALYSIS That's over two minutes before a single word. To be fair, Sol (about 284 s), Sonnet 5.5 (about 466 s) and Opus 5.5 (about 727 s) all take even longer at max, so Terra's wait is the shortest of the four. These are live metrics and can move.

2. The price cliff at 272K input tokens. Cross 272,000 input tokens and you pay double on input and cache, and one and a half times on output, for the entire request from the first token. FIRST-PARTY · OPENAI

3. Cybersecurity rating. OpenAI's own system card rates Terra Critical for cybersecurity, the same rating as Astra, its top model. FIRST-PARTY · OPENAI Separately, Codex CLI 0.163.0 (October 15) turned on Daybreak, OpenAI's approved-access program for security work, by default for eligible accounts. OpenAI's Daybreak guide says Terra gets standard request treatment by default, and Terra isn't on the guide's model list. FIRST-PARTY · OPENAI

4. Hard coding. OpenAI itself says Sol and Astra remain the better choice for the most complex coding tasks. FIRST-PARTY · OPENAI Artificial Analysis agrees: on its Coding Agent Index, Terra scores 48 and Sol scores 54. INDEPENDENT · ARTIFICIAL ANALYSIS

5. Effort level. On its main index, Artificial Analysis found low and medium are Terra's cost sweet spot. INDEPENDENT · ARTIFICIAL ANALYSIS On Hacker News, commenters reading OpenAI's Terminal-Bench chart said Terra at xhigh costs more than Sol at low and still scores lower. One put it as "Terra XHigh more expensive than Sol Low and still does worse." Artificial Analysis's own runs agree on the cost and disagree on the score: Terra xhigh scored about 46% at $0.66 per task, and Sol low scored about 31% at $0.38. INDEPENDENT · ARTIFICIAL ANALYSIS

Bonus from that thread. A commenter pointed out that switching models mid-session breaks your prompt cache. So routing to a cheaper model can cost more.

The math

Say you run 1,000 tasks like the ones in the Artificial Analysis index, at max effort, at its measured cost per task:

Model	Cost per task	1,000 tasks
GPT-6.1 Terra	$0.40	$400
GPT-6.1 Sol	$0.72	$720
Claude Opus 5.5	$5.98	$5,980

INDEPENDENT · ARTIFICIAL ANALYSIS

Terra over Opus saves $5,580 on that workload, for a model that scores ten points lower.

Now Sol. Per token, Terra is half. Per task, it's $0.40 against $0.72, about 56% of Sol's cost. Artificial Analysis breaks that cost down by token type. On output, Terra cost less than half what Sol did (about 17.5 cents vs 38.1 cents). The extra is on the input side, mostly cached context Terra read back (about 22.8 cents vs 34.3 cents). INDEPENDENT · ARTIFICIAL ANALYSIS

Price per token is half the story. How many tokens each task uses, input included, is the other half.

Now the cliff. Send 250,000 tokens in and 5,000 out at standard rates: $0.25 plus $0.025, about 28 cents. Send 300,000 in and the input rate doubles to $2 per million, which is 60 cents, while output goes to $7.50 per million, almost 4 cents. That's about 64 cents. Twenty percent more input, more than double the bill (about 2.3x).

Who should use it

Use Terra:

• At low or medium effort for volume work: business automation, form filling, document extraction, writing drafts and computer-use agents.
• Starting at low, and moving up only when your pass rate tells you to.
• On Batch for anything that can wait.
• With prompts kept under 272,000 tokens.

Skip Terra:

• For your hardest coding. Reach for Sol, Astra or Opus there.
• At max effort when a user is waiting.

And if you're running Terra at xhigh, price out Sol at low. On the Artificial Analysis index it's cheaper per task ($0.13 vs $0.19). INDEPENDENT · ARTIFICIAL ANALYSIS It also scores lower there, so test whether it's good enough for your job.

How to get paid with it

This part is my opinion: two service ideas to try.

Idea one: back-office automation for small businesses. Invoices in, forms filled, records updated. Terra is built for computer use and forms, and at a dollar in and five out, you can run a lot of it. Charge per workflow, or monthly. Before you quote, measure your cost per successful run. That sets your margin.

Idea two: a model-routing audit. Look for teams sending everything to their priciest model. Map the workload, move bulk jobs to Terra on Batch, and keep hard jobs on a bigger model. Then show them the before and after bill.

Verdict

Terra is the cheap workhorse. Run it at low and medium for volume, and keep the hard stuff on bigger models.