GPT-6.1 Terra halves Sol's rate card at $1/$5. At high effort it beats Haiku 5.5 for 62% of the cost per task, but matching Sol medium takes max effort and costs nearly twice as much.
OpenAI shipped GPT-6.1 Terra yesterday at $1 in and $5 out, exactly half of Sol on every line of the card. The rate card is the easy part. What a finished task costs depends on how hard you make it think, and Terra gets expensive fast at the top.
Photo by Florian Schindler on Unsplash
- Between index scores of about 40 and 45, Terra is the cheapest OpenAI model per task, and at high effort it undercuts Haiku 5.5 and DeepSeek V4.1 Flash too.
- Above that, run Sol at medium. Terra max ties its 48 and costs $0.40 a task to Sol's $0.21.
The card, next to its neighbours
The changelog dates gpt-6.1-terra to October 26 and pitches it at visual reasoning, computer use and writing. There was never a GPT-6 Terra. The line jumps straight from 5.6 to 6.1, and the price drops a long way on the jump.
| Model | Input | Cached | Write 30m / 12h | Output |
|---|---|---|---|---|
| GPT-6.1 Terra | $1.00 | $0.05 | $1.25 / $2.50 | $5.00 |
| GPT-6.1 Terra, Fast | $2.00 | $0.10 | $2.50 / $5.00 | $10.00 |
| GPT-6.1 Sol | $2.00 | $0.10 | $2.50 / $5.00 | $10.00 |
| GPT-5.6 Terra | $2.00 | $0.20 | $2.50 / $5.00 | $12.00 |
| GPT-6 Luna | $0.10 | $0.01 | $0.125 / $0.25 | $0.50 |
Per million tokens, Standard tier unless noted, prompts up to 272K input tokens, from OpenAI's pricing page on October 27, 2026. Over 272K the whole request bills 2x on input-side lines and 1.5x on output: Terra becomes $2.00 / $0.10 / $7.50. Batch and Flex are half. Regional processing adds 10%.
Against GPT-5.6 Terra, input halves, output falls 58% and cached input drops to a quarter, because the read discount went from 90% to 95% the same way it did on Sol. Terra's Fast tier is a flat 2x, which puts it on precisely Sol's Standard card. There is no Ultrafast for Terra; OpenAI only sells that tier on Sol and Astra.
The window matches Sol's, 922K input plus 128K output, with text and image in. The knowledge cutoff is newer, June 1, 2026 against Sol's April 30. One difference matters for migrations. Terra accepts reasoning.effort: "none", which Sol 6.1 dropped.
Half the card is not half the task
Artificial Analysis already lists Terra at six effort levels. We put each one on the same axis as Sol and the cheap models people will actually weigh it against. Bars are dollars per index task; the number before the name is the Intelligence Index score.
- 36 Terra, low$0.05
- 41 Terra, medium$0.10
- 42 Sol, low$0.13
- 43 Haiku 5.5, max$0.21
- 43 MiniMax M3.1 Flash$0.22
- 44 Terra, high$0.13
- 45 Terra, xhigh$0.19
- 45 DeepSeek V4.1 Flash, max$0.26
- 48 Sol, medium$0.21
- 48 Terra, max$0.40
- 52 Sol, max$0.72
Artificial Analysis Intelligence Index and cost to run it, divided per task, read October 27, 2026. Terra in accent. AA also lists a non-reasoning Terra at 33 for $0.11, more than Terra low costs; we left it out until that settles.
Between about 40 and 45 points, Terra is the cheapest way OpenAI sells you a given score. Terra high scores 44 for $0.13, two points above Sol low at the same price. It beats Haiku 5.5 at max effort by a point for 62% of the cost, and DeepSeek V4.1 Flash at max ties Terra xhigh while costing 37% more. Haiku's rate card is a tenth of Terra's, and DeepSeek's is about a quarter at the peak rate AA prices it on. They spend more tokens getting there.
Above 45 it flips. Terra max reaches 48, the score Sol gets at medium, and costs $0.40 a task against Sol's $0.21. The reason is volume. Terra max wrote 130 million output tokens across the index. Sol medium wrote 15 million. At $5 against $10, that's $650 against $150 of output before input is counted. If your workload needs Sol-medium quality, buy Sol at medium.
The December 11 migration just got cheaper
OpenAI shuts down gpt-5-mini on December 11, and its deprecations page still names GPT-5.6 Terra as the replacement. It didn't change yesterday. In our October price roundup that swap took a 500M-input, 100M-output month from $325 to $2,200. Here is the same month with Terra 6.1 on the list:
| Model | 500M in | 100M out | Month | vs gpt-5-mini |
|---|---|---|---|---|
| gpt-5-mini (until Dec 11) | $125 | $200 | $325 | 1.0x |
| GPT-6 Luna | $50 | $50 | $100 | 0.3x |
| Gemini 3.8 Flash | $375 | $375 | $750 | 2.3x |
| GPT-6.1 Terra | $500 | $500 | $1,000 | 3.1x |
| GPT-6.1 Sol | $1,000 | $1,000 | $2,000 | 6.2x |
| GPT-5.6 Terra (named replacement) | $1,000 | $1,200 | $2,200 | 6.8x |
Standard tier, no caching, prompts under 272K. Gemini 3.8 Flash is on an introductory price until December 31 and doubles to $1.50 / $7.50 on January 1, which makes this month $1,500.
Following OpenAI's own pointer now costs $1,200 a month more than the newer model in the same family. We'd test Luna first, since gpt-5-mini work often fits a model that cheap, and Terra 6.1 second. GPT-5.6 Terra has no shutdown date yet, but we see no reason to pick it for new work.
An agent month on Terra
We keep reusing one agent workload so posts stay comparable: 1 billion input tokens a month, 85% of them cache hits, plus 50 million output, all on the 30-minute cache. On Sol yesterday that came to $960.
On Terra 6.1 it's 150M written at $1.25 ($187.50), 850M read at $0.05 ($42.50) and $250 of output: $480. Exactly half, as you'd expect from a card that halves every line. GPT-5.6 Terra runs the same month for $1,145, mostly because of its $12 output and $0.20 reads. Push it through Batch and Terra 6.1 lands at $240.
The token count is the part to watch. That $480 assumes Terra writes the same 50 million output tokens Sol would. At max effort it won't: on the index it wrote 130 million output tokens to Sol max's 67 million. Set the effort level before you trust any monthly estimate, including ours.
What we couldn't check
OpenAI's launch post blocked our fetches, so its benchmark claims reach us through DataCamp and others. The headline one is Terminal-Bench 4.0 at max effort, reported as 50.3% against 21.5% for GPT-5.6 Terra. Treat it as OpenAI's number until the post is readable. GitHub, which put Terra in Copilot the same day, only says it solved "noticeably more" terminal and C# tasks than GPT-5.6 Terra.
Amazon Bedrock also went live on day one and, as with Sol, AWS prints no per-token rate. Its announcement only says Terra costs at most half of Sol and GPT-5.6 Terra. LiteLLM carries Bedrock Global at OpenAI list and US in-region at 1.1x.
Where we'd put it
Terra at high effort is where we'd move anything running on Sol at low: two more index points for the same $0.13 a task. It should also replace GPT-5.6 Terra, which costs more on every line of the card. We haven't seen AA numbers for 5.6 Terra at matched effort yet, so check your own evals.
Don't crank it to max to chase Sol. Past xhigh you pay more per task for a score Sol medium already reaches. Compare both against your own token mix on the GPT-6.1 Terra page or in the calculator.
Sources
- OpenAI pricing and GPT-6.1 Terra model page - every tier, long-context rates, context, cutoff, effort levels
- OpenAI changelog and deprecations - October 26 launch, gpt-5-mini replacement
- Artificial Analysis - Intelligence Index scores, cost to run and output tokens by effort level
- GitHub changelog and AWS announcement - Copilot and Bedrock availability
- DataCamp - OpenAI's reported Terminal-Bench 4.0 figures
- Anthropic pricing and Gemini API pricing - Haiku 5.5 and Gemini 3.8 Flash rates