OpenAI's GPT-6.1 Terra launched on 26 October 2026 as the cheaper sibling of GPT-6.1 Sol, which launched at DevDay on 29 September. Sol lists at $2 input / $10 output per million tokens, and Terra at $1 / $5. Half the price, same 1.05M context window, same effort ladder. The question is how much intelligence the discount removes, and whether it is the same amount on every task.
TL;DR
- Sol is the stronger model. Artificial Analysis (AA) scores Sol 51.8 and Terra 48.1 at max effort. On AA's Coding Agent Index it is 54 vs 48.
- Terra is cheaper per task at every matched effort level, by 1.8× to 2.6× (AA). At max effort it costs $0.40 vs Sol's $0.72.
- Effort is worth more than the model at the top end. Sol at xhigh costs $0.39 and scores 51.0, which beats Terra at max ($0.40, 48.1).
- Terra hallucinates less on AA-Omniscience (38% vs 54%). At max effort the two are tied on OpenAI's hard factual-error set (4.3% vs 4.4%).
- Speed claims need care. OpenAI says Terra is more than 3× faster on its own tests. On AA's max-effort runs the gap is 1.5× (384 s vs 569 s per task).
Sol vs Terra, max effort
Terra $0.40 vs Sol $0.72
Terra lower is better
Sol leads on coding
Same family, two different bets
Sol is built for depth and Terra for throughput. OpenAI's launch framing puts Sol or GPT-6 Astra on the hardest coding work, and Terra on professional computer use and high-volume agent loops. The table below covers what can be checked from the public listings.
| Specification | GPT-6.1 Sol | GPT-6.1 Terra |
|---|---|---|
| Released | 29 September 2026 (DevDay) | 26 October 2026 |
| API model ID | gpt-6.1-sol | gpt-6.1-terra |
| Context window | 1.05M tokens | 1.05M tokens |
| Lowest effort setting | low (none and minimal not supported) | none supported |
| Effort levels | low, medium, high, xhigh, max | none, low, medium, high, xhigh, max |
| Output speed (AA, max effort) | 69 tok/s | 91 tok/s |
Price: the rate card and the bill
Terra is exactly half of Sol on every standard line we could confirm. Both models switch to a higher long-context rate once a single prompt passes 272K input tokens, and the step reprices the whole request, not just the excess.
| Per 1M tokens (standard) | GPT-6.1 Sol | GPT-6.1 Terra | Terra vs Sol |
|---|---|---|---|
| Input | $2.00 | $1.00 | Half |
| Cached input | $0.10 | $0.05 | Half |
| Output (incl. reasoning) | $10.00 | $5.00 | Half |
| Input over 272K (whole request) | $4.00 | $2.00 | Half |
| Output over 272K (whole request) | $15.00 | $7.50 | Half |
The bill that matters is the one with reasoning in it. Sol's cheapest setting is low, so every Sol request bills some reasoning tokens. Terra can run with none. On a high-volume classifier, that setting alone can decide which model is cheaper in practice. Test your own prompts before you switch the model ID.
Independent results: the Artificial Analysis effort ladder
AA runs the full Intelligence Index (v4.3.2) once per effort level, so this is the cleanest like-for-like comparison available. Every row below is the same test at the same effort.
| Effort | Terra index | Sol index | Score gap | Terra cost / task | Sol cost / task | Cost ratio |
|---|---|---|---|---|---|---|
| Low | 36.2 | 42.1 | +5.9 Sol | $0.05 | $0.13 | 2.6× |
| Medium | 41.3 | 47.8 | +6.5 Sol | $0.10 | $0.21 | 2.1× |
| High | 43.9 | 50.2 | +6.3 Sol | $0.13 | $0.32 | 2.5× |
| xHigh | 45.4 | 51.0 | +5.6 Sol | $0.19 | $0.39 | 2.1× |
| Max | 48.1 | 51.8 | +3.7 Sol | $0.40 | $0.72 | 1.8× |
Two patterns stand out. The score gap stays between 3.7 and 6.5 points at every effort level, so Sol's advantage does not disappear with more reasoning. The cost gap shrinks at max, so Sol's extra reasoning buys it more at the top end than at the bottom.
The matched-money showdown
The useful question for a budget is not "which model is better" but "what does each dollar buy". Pairing rows by similar cost gives four clear answers.
| Cost per task (AA) | Terra option | Sol option | Winner at this price |
|---|---|---|---|
| About $0.13 | Terra high: 43.9 | Sol low: 42.1 | Terra, +1.8 points at the same cost |
| About $0.10 vs $0.13 | Terra medium: 41.3 at $0.10 | Sol low: 42.1 at $0.13 | Sol, +0.8 points for 30% more money. Close call. |
| About $0.19 vs $0.21 | Terra xhigh: 45.4 at $0.19 | Sol medium: 47.8 at $0.21 | Sol, +2.4 points for 10% more money |
| About $0.40 | Terra max: 48.1 at $0.40 | Sol xhigh: 51.0 at $0.39 | Sol, +2.9 points for the same money |
The surprise is at the top. The most expensive Terra setting costs as much as the cheaper Sol setting, and Sol at xhigh scores 3 points higher. If your work needs a top-end score, buy Sol at xhigh before you pay for Terra at max.
Vendor-reported benchmarks at max effort
OpenAI's launch charts are the only source for several of these rows. They were transcribed from the charts by secondary sites, and nobody has reproduced them independently yet. Terra's medium-effort results are shown where they exist, because OpenAI's launch framing leans on them.
| Benchmark | Terra medium | Terra max | Sol max | Astra max (ref) | Read |
|---|---|---|---|---|---|
| Agents' Last Exam v1 professional computer tasks | 57.6% | 60.0% | 58.6% | 59.3% | Terra ahead of Sol and Astra at max |
| MMMU-Pro visual reasoning | 84.6% | 85.5% | 86.3% | 86.7% | Sol ahead by 0.8 points |
| OSWorld 2.1 offline desktop agent | 66.7% | 74.2% | 78.0% | 80.9% | Sol ahead by 3.8 points |
| Terminal-Bench 4.0 agentic CLI work | 35.8% | 50.3% | 55.9% | 58.2% | Sol ahead by 5.6 points |
| Terminal-Bench-Science 0.1 scientific coding | 34.3% | 58.3% | 57.0% | 68.1% | Terra ahead of Sol |
| Factual errors, hard prompts lower is better | 10.1% | 4.3% | 4.4% | 4.2% | Tied within 0.1 points |
Terra wins on Agents' Last Exam and Terminal-Bench-Science, and loses on the other four vendor rows. Sol's lead on Terminal-Bench 4.0 is the one that matters most for coding agents, and the AA Coding Agent Index (54 vs 48) points the same way.
Accuracy: the trade-off nobody mentions
Terra's factual story is better than its score suggests. On AA-Omniscience, which rewards admitting ignorance, Terra's hallucination rate is 38% against roughly 54% for Sol at max effort. That is the opposite of the capability ranking. A model that says "I don't know" more often can score lower on raw knowledge tests and still be the safer choice for support answers and research summaries.
| Accuracy signal | Terra | Sol | Better |
|---|---|---|---|
| AA-Omniscience hallucination rate | 38% | ~54% | Terra |
| Factual errors, hard prompts, max | 4.3% | 4.4% | Tie |
| Factual errors, hard prompts, low | 16.7% | Not published | Not comparable |
Two cautions. Sol's AA hallucination figure is taken from AA's max-effort data, as relayed by OpenRouter and Emergent. Terra's figure is from AA's launch article, and AA's effort setting for it is not stated in the text we read. And The Decoder reports that the Terra system card places it under OpenAI's "critical" safety threshold, so check that document before you deploy either model in a sensitive domain.
Routing guide
| Your workload | Pick | Why |
|---|---|---|
| Hard coding agents, multi-file refactors | Sol at xhigh | Coding Agent Index 54 vs 48; Terminal-Bench 4.0 55.9% vs 50.3%. Cheaper than Terra at max. |
| High-volume agent loops at low or medium effort | Terra | Half the list price, and Terra medium costs $0.10 per AA task. |
| Desktop and browser automation | Sol, unless cost dominates | OSWorld 2.1 offline 78.0% vs 74.2%. Terra's gap is small enough to test on your own tasks. |
| Professional computer tasks (Agents' Last Exam style) | Terra | 60.0% vs 58.6% at max, for half the list price. |
| Support answers and research where a wrong answer costs more than a missing one | Terra | AA-Omniscience hallucination 38% vs 54%. |
| Visual and screenshot reasoning | Sol | MMMU-Pro 86.3% vs 85.5% at max. Within a point, so the price should decide it. |
| Batch and offline document pipelines | Terra | Half the price with Batch or Flex. Largest saving per dollar of any row on this page. |
| Anything needing a top-end AA score | Sol at xhigh or max | Terra's ceiling on AA is 48.1. Sol reaches 51.8 and costs $0.39 at xhigh. |
The short version
Sol is the better model. Terra is the better deal below the top of the ladder. Terra gives you about 86 to 93% of Sol's AA score at matched effort, for roughly half the list price and 1.8× to 2.6× less per task. Pay for Sol when the job needs its top-end coding and screen-control scores, or when you can run it at xhigh and get more than Terra can at max.
The cheapest test is to run 200 of your own tasks at Terra medium and Sol low. Those two settings cost about the same and score within a point of each other on AA, so the results will tell you which model fits your work.
FAQ
Is GPT-6.1 Terra a good replacement for GPT-6.1 Sol?
For high-volume work that does not need the top score, usually yes. Terra is 3.7 to 6.5 points behind Sol on AA at matched effort, and it is 1.8× to 2.6× cheaper per task. For hard coding and top-end accuracy, keep Sol, or use Sol at xhigh.
Which effort level should I start with?
For Terra, start at medium. It scores 41.3 on AA at $0.10 per task, which is close to Sol at low (42.1 at $0.13). For Sol, start at medium for everyday work and xhigh for coding agents. AA data shows xhigh to max adds less than a point for nearly double the cost.
Why is the speed claim different from the AA numbers?
OpenAI's "more than 3× faster" claim compares its own runs on its own evaluations. AA's max-effort runs show 384 s for Terra and 569 s for Sol per Intelligence Index task, a 1.5× gap. Both can be true because the workloads differ. Measure your own tasks before you rely on either figure.
Does Terra's lower price apply to cached input too?
Yes. Cached input is $0.05 for Terra and $0.10 for Sol, both half the uncached rate. Agents that resend a large, stable prefix on every call will pay mostly the cached rate, which narrows the gap between the models in practice.
Sources
- Artificial Analysis — GPT-6.1 Terra pushes the cost-efficiency frontier: Terra's max-effort Intelligence Index (48), cost per task, AA-Omniscience hallucination rate, Terminal-Bench gain.
- Artificial Analysis — GPT-6.1 Sol release page: Sol's effort ladder, cost per task and token use.
- Artificial Analysis — GPT-6.1 Sol vs GPT-6.1 Terra comparison: the head-to-head page (figures load dynamically, so the values above come from the sources below).
- eesel — GPT-6.1 Terra: pricing, benchmarks, and when to pick it over Sol: Terra's effort ladder from AA, output speed, pricing tiers and the Coding Agent Index.
- eesel — GPT-6.1 Sol pricing: Sol's rate card, service tiers, 272K cliff, and AA effort ladder with cost and token use.
- Kingy AI — GPT-6.1 Terra: Benchmarks, Pricing & Model Comparisons: OpenAI's launch-chart values for Terra at medium and max, Sol, and Astra.
- AIHubMix — GPT-6.1 Terra vs GPT-5.6 Terra: the same OpenAI benchmark table, and the factual-error comparison with GPT-5.6 Terra.
- The Decoder — GPT-6.1 Terra halves OpenAI's API prices: OpenAI's speed and cost claims, Terminal-Bench-Science, and the "critical" safety threshold note.
- OpenRouter — GPT-6.1 Sol (batch): AA data for Sol at low, high, xhigh and max, including AA-Omniscience and Terminal-Bench 4.0.
- Emergent — GPT-6.1 Sol benchmarks: Sol's hallucination rate change from 60% to 54% on AA-Omniscience.
Vendor scores are labelled as vendor-reported and come from OpenAI's launch charts via secondary sites. AA scores are independent runs in AA's Intelligence Index v4.3.2 harness at the stated effort. Cost per task is AA's figure and excludes any tier discounts. The bill chart uses list prices and excludes reasoning tokens. Prices are USD per million tokens, October 2026.