GPT-6.1 Terra costs exactly half of GPT-6.1 Sol — $1 against $2 input, $0.05 against $0.10 cached, $5 against $10 output. Because all three rates halve by the same factor, the saving is 50% on any token mix, in every pricing tier, on both sides of the 272K surcharge.
Pick GPT-6.1 Terra for vision work, document and screenshot interpretation, computer use and prose — OpenAI’s stated targets for it — and for high-volume mechanical work, where its none reasoning-effort setting makes it a cheap extractor that Sol cannot be configured to be at any price.
Pick GPT-6.1 Sol for complex coding, which is its stated territory, or when your latency budget might tighten later: Sol has an Ultrafast tier and Terra has none.
Prices verified 27 October 2026 against OpenAI’s pricing page and both model pages.
The rate card, in full
Per million tokens, short context (prompts of 272,000 input tokens or fewer).
| GPT-6.1 Terra | GPT-6.1 Sol | |
|---|---|---|
| Input | $1.00 | $2.00 |
| Cached input | $0.05 | $0.10 |
| Output | $5.00 | $10.00 |
| Cache write, 30-minute | $1.25 | $2.50 |
| Cache write, 12-hour | $2.50 | $5.00 |
| Batch / Flex | $0.50 / $0.025 / $2.50 | $1.00 / $0.05 / $5.00 |
| Fast mode | $2.00 / $0.10 / $10.00 | $4.00 / $0.20 / $20.00 |
| Ultrafast mode | not offered | $12.00 / $0.60 / $60.00 |
| Above 272K input | 2x in, 2x cached, 1.5x out | 2x in, 2x cached, 1.5x out |
| Knowledge cutoff | 1 June 2026 | 30 April 2026 |
| Reasoning effort floor | none | low |
Batch and Flex are priced identically to each other at 50% below standard on both models. Fast mode — renamed from Priority processing on 30 July 2026 — is 2x standard. Regional-processing endpoints for data residency add 10%, as do FedRAMP endpoints.
The 50% holds everywhere except the one place Terra does not exist
Price cuts between two models in a family are usually uneven: input moves by one factor, cached input by another, output by a third. That is why a headline percentage rarely survives contact with a real workload, and why the cost calculator asks for your token mix rather than quoting a single number.
This pair is the exception. Terra is at a clean 2:1 against Sol on every rate in every tier:
| Tier | Terra | Sol | Ratio |
|---|---|---|---|
| Standard | $1.00 / $0.05 / $5.00 | $2.00 / $0.10 / $10.00 | 2.00x |
| Batch and Flex | $0.50 / $0.025 / $2.50 | $1.00 / $0.05 / $5.00 | 2.00x |
| Fast | $2.00 / $0.10 / $10.00 | $4.00 / $0.20 / $20.00 | 2.00x |
| Standard, above 272K | $2.00 / $0.10 / $7.50 | $4.00 / $0.20 / $15.00 | 2.00x |
| Ultrafast | — | $12.00 / $0.60 / $60.00 | no comparison |
So the Sol-to-Terra decision is the rare one where you do not need to know your own traffic shape to know the saving. Cache-dominated, output-dominated, over the surcharge cliff or under it, batched or interactive: it is half, to the cent.
The exception is the interesting part. Terra is absent from the Ultrafast table entirely — only GPT-6 Astra and GPT-6.1 Sol appear there. A team that standardises on Terra for the rate card, and then finds eighteen months later that its latency requirement has tightened, cannot buy its way up a tier. It has to re-qualify a different model, with the evaluation and regression work that implies. Sol keeps that door open at twice the price.
What none buys, and why it is not a price difference
The difference most likely to decide this choice is not on the pricing page at all. It is the reasoning-effort ladder, and the two models do not have the same one.
- GPT-6.1 Terra:
none,low,medium(default),high,xhigh,max— six levels. - GPT-6.1 Sol:
low,medium(default),high,xhigh,max— five levels.
Terra’s extra rung is at the bottom. Setting reasoning.effort to none stops the model spending output tokens on internal reasoning before it answers. On classification, extraction, routing, tagging and short-form rewriting, those reasoning tokens are pure cost against no benefit — and since output bills at five times input on both models, they are the expensive kind of cost.
Sol’s floor is low. There is no setting that switches its reasoning off. So for a high-volume mechanical pipeline, Terra is not simply the cheaper of two interchangeable models; it is configurable for a job Sol cannot be configured for at any price. That makes the gap on such work much wider than 2x, and the true figure depends on how many reasoning tokens Sol emits at low on your prompts — which is measurable in an afternoon and inferable from nothing published.
Worth knowing if cost is the whole story: GPT-6 Luna also accepts none, at $0.10/$0.50, a tenth of Terra. The Luna comparison has that model’s arithmetic against Anthropic’s budget tier. Terra’s case over Luna is the newer cutoff, the vision tuning and the higher effort ceiling, not the price.
Everything that is identical
Buyers lose time hunting for differences that are not there. On these specifications the two models match exactly:
- Context window 1,050,000 tokens, with a 922,000-token maximum input and a 128,000-token output cap.
- Modalities text and image in, text out. Audio and video are unsupported on both.
- Endpoints Chat Completions, Responses and Batch. Tool calling requires the Responses API on both.
- Features streaming, structured outputs, function calling, file search, image input, web search and prompt caching. Neither supports fine-tuning or predicted outputs.
- Tools web search, file search, image generation, code interpreter, hosted shell, apply patch, skills, computer use, MCP and tool search. Computer use is available on both, despite being named in only Terra’s positioning line.
- Rate limits 5,000 RPM and 1M TPM on Build, 10,000 and 4M on Launch, 15,000 and 40M on Grow.
- Long-prompt surcharge 2x input, 2x cached input and 1.5x output on the whole request once input passes 272,000 tokens. The long-prompt guide covers how that compares with other vendors’ cliffs.
There is also no published capability grade separating them. OpenAI’s catalogue carries no intelligence or reasoning rating — the field labelled “Reasoning” on each card is a link to the effort guide followed by that model’s ladder. Sol’s page offers “Near-Astra performance for complex work at a lower cost,” a cost-performance claim rather than a score, and OpenAI’s model-selection guide does not mention Terra at all.
A worked job: 10,000 documents a month
Document extraction is squarely Terra’s stated territory, so price it there. Each request sends 4,000 cached input tokens (fixed instructions and a stable output schema), 3,000 fresh input tokens (the page) and emits 800 output tokens.
| Cached input | Fresh input | Output | Per request | 10,000 requests | |
|---|---|---|---|---|---|
| GPT-6.1 Terra, standard | $0.00020 | $0.00300 | $0.00400 | $0.00720 | $72.00 |
| GPT-6.1 Sol, standard | $0.00040 | $0.00600 | $0.00800 | $0.01440 | $144.00 |
| GPT-6.1 Terra, Batch | $0.00010 | $0.00150 | $0.00200 | $0.00360 | $36.00 |
| GPT-6.1 Terra, Fast | $0.00040 | $0.00600 | $0.00800 | $0.01440 | $144.00 |
Two things fall out of the bottom two rows. Batch halves Terra again, to a quarter of Sol at list — and for an overnight extraction run there is usually no reason not to take it. And Terra in Fast mode lands on $144.00, which is Sol at standard rates to the cent.
Terra in Fast mode is Sol at list price
That collision is the cleanest decision rule on this page. Fast mode is 2x standard, so Terra-Fast is $2.00 input and $10.00 output per million — exactly GPT-6.1 Sol’s standard rate.
Buying lower latency on Terra therefore costs the same as running Sol at list, which turns a tier upgrade into a model choice. At that identical price point:
- Terra-Fast gives the newer knowledge cutoff, the vision and prose tuning, the
noneeffort floor, and materially higher throughput than standard. - Sol-Standard gives the coding-oriented tuning and the option to escalate to Ultrafast later.
Neither dominates. The deciding question is whether your latency requirement is stable, because only one of these two options has a rung above it. The cache retention guide has the 30-minute and 12-hour cache-write arithmetic, which is the other lever on effective cost at this price point.
Knowledge cutoff runs backwards against price
Terra’s cutoff is 1 June 2026. Sol’s is 30 April 2026. The cheaper model knows more recent things — a month and a day more.
This matters less than it looks for most work and more than it looks for some. If your prompts carry their own context, a cutoff difference of five weeks is noise. If the model is answering about library versions, APIs, product names or anything that moved in May 2026, it is not noise, and the usual mitigation — retrieval — costs input tokens that Terra also meters at half Sol’s rate.
Which one to pick
| Your work | Model | Why |
|---|---|---|
| Vision, screenshots, technical drawings | Terra | OpenAI’s stated target, half the price |
| Document extraction at volume | Terra, Batch, effort none | $36 per 10,000 in the job above; Sol cannot set none |
| Drafting and revising prose | Terra | Stated target; the cutoff helps |
| Complex coding | Sol, then test Terra | Sol’s stated territory; verify before moving |
| Computer use | Either, measure | Named in Terra’s positioning, supported as a tool on both |
| Latency may tighten later | Sol | Only Sol has Ultrafast above it |
| Pure cost per mechanical task | Terra, or Luna | Both accept none; Luna is a tenth of Terra |
Neither model carries a shutdown date and neither exposes a dated snapshot, so the lifecycle risk is identical and should not break the tie. The model retirements tracker carries the dates for the models that do have them, and the pricing tracker logs each change to these rates as it happens.
For the plan-level picture — which of these models reaches ChatGPT’s consumer tiers and when — see the ChatGPT review. For the cross-vendor version of this decision, Claude vs ChatGPT sets the same arithmetic against Anthropic’s rate card, where the tokenizer difference changes the comparison before the rates do.
What changed
- 27 Oct 2026: Page created. GPT-6.1 Terra launched 26 October at $1/$5, half of Sol on every rate — see launch analysis. Terra accepts a
nonereasoning effort that Sol does not, and is absent from the Ultrafast tier.
GPT-6.1 Terra vs GPT-6.1 Sol — frequently asked questions
Is GPT-6.1 Terra just a cheaper GPT-6.1 Sol?
Close to it, with two real exceptions. Terra is exactly half of Sol on input ($1 against $2), cached input ($0.05 against $0.10) and output ($5 against $10), and it carries a newer knowledge cutoff (1 June 2026 against 30 April 2026). The context window, the 922,000-token maximum input, the 128,000-token output cap, the modalities, the endpoints, the feature list, the tool list, the 272K surcharge and the published rate limits are all identical. The two exceptions are that Sol has an Ultrafast tier and Terra has none, and that Terra accepts a `none` reasoning effort that Sol does not. OpenAI also points them at different work: Sol at complex coding, Terra at visual reasoning, computer use and prose.
Does the 50% saving depend on my token mix?
No, and that is unusual. Most model-to-model price cuts move input, cached input and output by different factors, so the real saving depends on how cache-heavy or output-heavy your traffic is. Terra halves all three rates by the same factor, so the saving is 50% whatever the mix. It also holds in every pricing tier — Standard, Batch, Flex and Fast all sit at a 2:1 ratio — and on both sides of the 272K surcharge, because the long-context multipliers (2x input, 2x cached, 1.5x output) are identical on both models. The one place the rule breaks is Ultrafast, where Sol is priced at $12/$60 and Terra has no price at all.
What does a reasoning effort of `none` actually change?
It turns the model into a non-reasoning one for that request, so it stops spending output tokens on internal reasoning before it answers. For classification, extraction, routing, tagging and short-form rewriting — jobs where the reasoning tokens are pure cost — that is the difference between a viable unit price and an unviable one. GPT-6.1 Terra's `reasoning.effort` accepts `none`, `low`, `medium` (the default), `high`, `xhigh` and `max`. GPT-6.1 Sol accepts only `low`, `medium` (default), `high`, `xhigh` and `max`: its floor is `low`, and there is no setting that switches reasoning off. So on high-volume mechanical work Terra is not merely cheaper than Sol, it is doing a job Sol cannot be configured to do at any price.
Should a team running GPT-6.1 Sol migrate to Terra?
For vision work, document and screenshot interpretation, computer use and drafting or revising prose, the burden of proof has moved onto staying with Sol: Terra is what OpenAI built for those jobs, at half the price, knowing more recent things. For complex coding, Sol is still the stated target — test before moving rather than assuming the cut carries across a task type the vendor points elsewhere. Two things to measure rather than assume. Output bills at five times input on both models, so a difference in verbosity between them will move the invoice more than the halved rate card does. And reasoning effort swings cost per task by more than the entire Sol-to-Terra price difference, so compare the two at the effort level you actually ship, not at their defaults.
Does OpenAI say one of these models is more capable?
Not in any graded field. OpenAI's model catalogue has no qualitative capability or intelligence rating — the column labelled "Reasoning" on each model card is a link to the reasoning-effort guide followed by that model's effort ladder, not a grade. The qualitative ranking language lives in prose, and it is about the family rather than this pair: Astra is described as the most powerful model and Luna as the most cost-efficient. Sol's own page frames it as "Near-Astra performance for complex work at a lower cost," which is a cost-performance claim rather than a capability score. OpenAI's model-selection guide, which pairs models with effort settings for seven workload rungs, does not name Terra at all. Nothing published separates the two on capability; the separations that are published are price, cutoff, speed tiers, the effort floor and stated task fit.
Can I pin a snapshot, and does either model have a retirement date?
No to the first and no to the second. Neither model publishes a dated snapshot — `gpt-6.1-terra` and `gpt-6.1-sol` are the only identifiers OpenAI lists, and each model's sole listed snapshot is itself. Neither appears anywhere on OpenAI's deprecations page, so neither carries a shutdown date. OpenAI's published policy is at least six months' notice for generally available models, three months for specialised variants and as little as two weeks for preview models. The practical consequence is the same for both: a build on either model cannot be pinned to a fixed version, so a change in behaviour underneath you can be neither prevented nor reproduced. That risk is identical on both sides and should not influence the choice between them.