GPT-6.1 Terra vs GPT-6.1 Sol: Where the $5 Output Stops Being the Cheaper Answer

The comparison most people run between GPT-6.1 Terra and GPT-6.1 Sol is wrong, and it is wrong in a way that flatters the cheaper model. They set both to maximum reasoning effort, note that Terra scores a little lower, note that Terra costs half as much, and conclude that Terra is the sensible default. But nobody runs a production pipeline at maximum effort on every call. The real choice is Terra at xhigh against Sol at low — and on that comparison, commenters on Hacker News pointed out within hours of launch, Terra can cost more than Sol while scoring worse. The model you pick matters less than the effort level you pin it to, and the effort level is the variable neither launch page puts in the headline.

Both models are from the same generation and the same month: GPT-6.1 Sol shipped September 29, 2026, GPT-6.1 Terra on October 26. They share a 1.05M context window, a 128K output ceiling, the same three endpoints, and the same text-and-image input. What separates them is a four-point gap on the independent index and a 2× gap on the price card — and the question of whether four points is worth double depends entirely on what your workload does with the last four points.

The same model, two price cards

Specifications from OpenAI's model documentation, read October 27, 2026.

• Price — GPT-6.1 Terra $1 input / $5 output per million; GPT-6.1 Sol $2 / $10. Terra's cached input is $0.05, Sol's $0.10.

• Context — 1,050,000 tokens each, 922,000 maximum input, 128,000 maximum output. Identical.

• Knowledge cutoff — Terra June 1, 2026; Sol April 30, 2026. The cheaper, newer model has the fresher world by about a month.

• Reasoning effort — Terra offers none, low, medium, high, xhigh, max; Sol offers low through max. Terra's extra "none" rung is a genuine cost floor Sol does not have.

• Modifiers — both have batch and flex at 50% below standard, fast at 2×, regional at +10%, and the 272K-input threshold. Sol alone has an Ultrafast tier at 6× standard, which at list works out to $12 input and $60 output per million.

• Tools and endpoints — identical set. Responses API for tool calling, Chat Completions without tools, Batch for both, no fine-tuning on either.

Four points, twice the price

Artificial Analysis currently lists GPT-6.1 Terra at an Intelligence Index of 48 — 25th of 221 models — and GPT-6.1 Sol at 52, 13th. That is the entire independent capability gap: four points, at maximum effort, on a composite index. It is not nothing. It is also not the gulf that "flagship versus mid-tier" language implies, and the two models share a generation, an architecture family, and a context window.

The cost gap is easier to state: exactly 2× on both input and output at standard rates. Terra is not a discount-bin model; it is Sol's price cut in half with four index points removed.

Cost per finished task, not per token

The number that should drive the decision is cost per completed task, because it folds in how many tokens each model spends thinking. On Artificial Analysis's Intelligence Index, GPT-6.1 Terra costs $0.40 per task and GPT-6.1 Sol $0.72 — a 44% saving for Terra, smaller than the 50% headline because Terra burns slightly more tokens to finish the same work at max effort.

Scale that. At a million tasks a month, the difference is roughly $320,000. At ten thousand tasks, it is $3,200 — real money, but smaller than one engineer-week spent rebuilding an eval harness. The threshold where Terra's price advantage justifies a migration is a volume question, not a quality question, and it is the first number to compute before reading any benchmark table.

There is a second-order effect in Terra's favour that does not show up in either price: the knowledge cutoff. Terra's June 1 training horizon is a month fresher than Sol's April 30. For workloads that depend on recent APIs, library versions, or documentation, that month is worth more than four index points, and it costs half as much to get.

The effort-level trap, worked through

Here is the failure mode in concrete terms. Terra at xhigh effort spends more tokens per task than Sol at low effort. If Terra's tokens cost half as much but it uses more than twice as many of them, the cheaper model is no longer cheaper — and because xhigh is a weaker setting than Sol's low on the hardest tasks, you are paying more for a worse answer. Hacker News commenters described exactly this on Terminal-Bench 4.0 within a day of launch.

The fix is not to avoid Terra. It is to compare at equal spend rather than equal settings. Concretely: pick your quality bar first — the pass rate your product needs on your own eval — then find the cheapest (model, effort) pair that clears it. That search space has twelve practical points across these two models, not two, and the answer is usually not "Terra at max" or "Sol at max."

OpenAI's own numbers illustrate how much the setting moves. Terra's error rate on the vendor's internal error-inducing set falls from 16.7% at low effort to 4.3% at maximum — a 12-point swing inside one model. The gap between Terra and Sol on the independent index is four points. The effort dial is a bigger lever than the model choice, and it is free to turn.

Where Sol still earns the markup

OpenAI's positioning for Sol is "near-Astra performance for complex work at a lower cost," and the vendor benchmarks where it separates from Terra are consistent with that: long-horizon agentic work, the hardest coding, and dense professional documents. Those are also the tasks where a failed run costs more than the tokens, which is precisely when a 2× price premium is easy to justify.

Terra's counter-argument is narrower than the marketing suggests but sharper than the index implies. OpenAI reports Terra within 5 points of Claude Opus 5.5 on Terminal-Bench-Science 0.1 at under one-fifth the estimated cost, and 15 points above GPT-5.6 Terra on Agents' Last Exam at roughly a third of the cost. Those are vendor-run numbers on vendor-selected tests, so they establish direction rather than magnitude — but the direction is "good enough at science-adjacent and agentic tasks, much cheaper," which is a real niche.

The honest split is this: Sol is for the tasks where you would notice a four-point drop in quality, and Terra is for the tasks where you would notice the invoice. Very few workloads are entirely one or the other, which is why the routing question matters more than the model question.

Latency may decide it before quality does

Artificial Analysis measures GPT-6.1 Sol at 65.9 output tokens per second with 325.58 seconds to first token at max effort. That time-to-first-token figure is not a typo, and it is disqualifying for anything interactive. Terra has no independent throughput measurement yet; OpenAI describes it as substantially faster, and if that holds, the gap is large enough to settle the argument for chat-shaped or agent-loop products regardless of index scores.

For batch work the calculus reverses: 325 seconds of thinking is irrelevant if the job runs overnight, and Sol's higher ceiling is worth paying for. Latency is a workload property, not a model property, and it belongs in the decision before benchmark rank does.

Holding both arms on one key

OrcaRouter carries GPT-6.1 Sol at its $2/$10 list price with 0% markup, on the same key as 200-plus other models. GPT-6.1 Terra is not on our catalogue as of October 27 — it is a day old and available through OpenAI's API — so the practical pattern today is to measure the Sol arm on our side and set it as the escalation target: run Terra where it wins, fail over to Sol when a request needs the extra four points or when a provider has a bad hour. Both are the same generation, the same context window, and the same endpoint shape, so the routing rule is a model string and an effort setting, not a rewrite.

The routing DSL matters here for a specific reason. The interesting production configuration is not "Terra or Sol" but "Terra at high, escalating to Sol at medium when confidence is low" — a composition that would otherwise mean two integrations, two retry policies, and two bills. That is the kind of decision this pair invites, and it is the reason the two models should be evaluated as a ladder rather than as an either-or.

The decision rule

• High-volume, well-specified, latency-sensitive — screenshots, forms, extraction, browser automation, routine drafting. Terra, at the lowest effort that clears your eval. The four-point deficit will not show up and the 44% per-task saving will.

• Hard coding, long-horizon agents, dense professional documents, anything where a failed run costs more than the tokens. Sol. This is the workload the 2× premium exists for.

• Interactive chat with a strict time-to-first-token budget. Terra, unless you have measured Sol at low effort and found the latency acceptable. Sol at max effort is a batch instrument wearing a chat interface.

• Freshness-sensitive work — recent APIs, library versions, current documentation. Terra's June cutoff beats Sol's April cutoff, and it does so at half the price.

• Unsure. Benchmark at equal spend, not equal settings. That single change will move more teams to Terra than any benchmark table will.

The open question is not whether Terra is good enough to replace Sol — it plainly is not, on the only independent measure available. It is whether the four-point gap survives the arrival of GPT-6.1 Astra, which has not shipped and which will push Sol down the ladder the same way Sol pushed Astra down it. When that happens, the interesting comparison will be Terra against a discounted Sol, and the effort dial will matter even more than it does today.