GPT-6.1 Terra Arrives at Half the Price, With 37 Fewer Error Points and No Flagship in Sight

OpenAI shipped GPT-6.1 Terra on October 26, 2026, at $1 per million input tokens and $5 per million output — half the price of GPT-5.6 Terra on both axes, and half the price of GPT-6.1 Sol four weeks after that model's own launch. The number that will be quoted most is a reliability one: on OpenAI's internal error-inducing-prompt test, mistakes at low reasoning effort fell from 53.6% to 16.7%. The number that will be quoted least is that Terra still trails Sol by four points on the independent Artificial Analysis index. Both are true, and the space between them is the whole story of this release.

That space matters because of what was predicted. When this model was still a string in a Microsoft evaluation-harness pull request, we wrote on October 20 that a cheaper mid-tier model would need to land below $2 input and $10 output to make sense, and that reports of it beating Sol were arithmetic, not evidence. The price half of that prediction landed almost exactly: $1/$5, close to the $1/$4 we called plausible. The performance half did not. Terra beats its predecessor clearly and beats Sol nowhere that has been independently measured.

What OpenAI announced, in the order it matters

The specification below comes from OpenAI's model documentation, read on October 27. The performance claims in the next section are OpenAI's own and are labeled as such throughout.

• Price — $1 per million input, $5 per million output, $0.05 cached input, against $2/$12 for GPT-5.6 Terra and $2/$10 for GPT-6.1 Sol.

• Context — 1,050,000 tokens, maximum input 922,000, maximum output 128,000. Identical to both predecessors; this is not a context-window release.

• Knowledge cutoff — June 1, 2026, up from February 16, 2026 on GPT-5.6 Terra.

• Reasoning effort — none, low, medium (default), high, xhigh, max. Note the "none" setting: GPT-6.1 Sol does not offer it, so Terra has one cheaper floor than Sol does.

• Modalities — text and image in, text out. No audio, no video.

• Endpoints — Chat Completions, Responses, and Batch. Tool calling requires the Responses API. No fine-tuning, no Realtime, no Assistants.

• Cost modifiers — batch and flex are 50% below standard; fast mode is 2× standard; regional processing carries a 10% premium where offered. Requests above 272K input tokens bill at 2× input and cache rates plus 1.5× output for the whole request.

The ChatGPT rollout is narrower than the API rollout. Terra is available on Plus, Pro, Business, Enterprise, and Edu plans, inside ChatGPT Work and Codex, selectable from the Advanced menu on the desktop app and on Work for the web, with Work also supporting it on mobile. It is explicitly not available in ChatGPT Chat, and it does not support UAE inference residency. Enterprise access still depends on workspace availability and admin controls. If your mental model of an OpenAI launch is "it shows up in the chat box," this one does not.

The 37-point claim is real but narrower than it sounds

OpenAI's headline result is a drop in errors on prompts designed to induce mistakes — a test the company runs internally, has not published, and which nobody outside has reproduced. At low reasoning effort the error rate went from 53.6% on GPT-5.6 Terra to 16.7% on Terra. At maximum effort the company reports a further 18-point fall to 4.3%, which it describes as close to GPT-6 Astra and GPT-6.1 Sol.

Read that carefully. The 37-point figure is not "37% fewer errors on your work." It is a 37-point swing on an adversarially constructed internal set, at the lowest effort setting, against a model that is being retired from some distribution channels. The honest summary is that Terra's floor got much less embarrassing. Whether its ceiling reached Sol's is a different question, and the internal test cannot answer it because Sol's number on that same test is quoted only as "close."

The rest of OpenAI's published claims, all vendor-run:

• Terminal-Bench 4.0 — 50.3% solved at maximum effort on Terra, against 21.5% for GPT-5.6 Terra. More than double, and the largest generation-over-generation jump in the release.

• Agents' Last Exam — 15 percentage points above GPT-5.6 Terra at medium effort, at roughly one-third the cost.

• MMMU-Pro — beats GPT-5.6 Terra at every reasoning level and approaches Astra and Sol, with no numerical scores published.

• Terminal-Bench-Science 0.1 — within 5 points of Claude Opus 5.5 at under one-fifth the estimated cost.

• Speed — the launch describes Terra as substantially faster than Astra and Sol. Artificial Analysis had not published a measured throughput figure for the model when we checked, so treat "faster" as a claim until someone runs it.

Independent scores: six points over its predecessor, four behind Sol

Artificial Analysis lists GPT-6.1 Terra at an Intelligence Index of 48, ranked 25th of 221 models, at $0.40 per index task. On the same live listing, GPT-5.6 Terra sits at 42 (53rd, $1.40 per task) and GPT-6.1 Sol at 52 (13th, $0.72 per task). Those comparisons are only valid against each other as currently published — the index is periodically renormalized, and older writeups quote different absolute numbers for the same models.

Three things follow from that. First, Terra is a genuine upgrade on its predecessor: six index points for less than a third of the cost per task. Second, it does not beat Sol, which is what the pre-launch reporting suggested it might. Third, and least discussed, Terra's cost per finished task is the lowest of the three — $0.40 against Sol's $0.72 — which is a more useful number than price per token, because it folds in how many tokens the model burns getting to an answer.

The measured speed gap is the one that cuts the other way. GPT-5.6 Terra is listed at 137 output tokens per second and 111.75 seconds to first token; GPT-6.1 Sol at 65.9 tokens per second and 325.58 seconds. If Terra really is the fastest of the three, it inherits the latency crown as well as the price crown, and that combination — not raw intelligence — is what a computer-use or document pipeline actually cares about.

Where "half price" survives, and where it quietly stops being true

A standard-tier workload of 100,000 input tokens and 5,000 output tokens cost $0.26 on GPT-5.6 Terra and costs $0.125 on GPT-6.1 Terra — a 52% saving, slightly better than the headline because output was the more expensive side of the old rate card.

The long-context multiplier does not erase the discount, because it applies to both models: over 272K input tokens, Terra bills at $2 input and $7.50 output for the whole request, GPT-5.6 Terra at $4 and $18. A 400K-input, 4K-output request runs $0.83 against $1.67. Still half.

Where the arithmetic changes is fast mode. Terra at 2× standard is $2/$10 — precisely GPT-6.1 Sol's standard list price, for a model four index points weaker. If your workload needs Terra's speed and you are paying the fast-mode multiplier, the price argument for Terra over Sol disappears, and the only remaining question is whether Terra's output is good enough for the task. Batch and flex are the reverse: 50% off takes Terra to $0.50/$2.50, which is the cheapest competent general-purpose rate on OpenAI's card and the setting most bulk pipelines should be testing first.

What this means if you route models for a living

One honesty note before the product note: GPT-6.1 Terra is not on the OrcaRouter catalogue at the time of writing. Our catalogue lists 214 models, including GPT-5.6 Terra, GPT-6.1 Sol, GPT-6 Astra and GPT-6 Luna, but not this one — so anyone telling you they route Terra today is either moving faster than the vendor's own documentation or guessing.

What that leaves you is a two-part decision that OrcaRouter does make cheaper. The comparison arms are both already live: GPT-5.6 Terra at its $2/$12 list and GPT-6.1 Sol at $2/$10, on one key, at 0% markup over provider list price. That matters more than usual here, because the interesting question is not "is Terra better than Sol" in the abstract but "at what point does Terra's $0.40 per task stop beating Sol's $0.72 on my traffic." You can measure the Sol arm on our side today and hold the fallback ladder there while Terra is trialled directly against OpenAI's API. If you later move the workload across, automatic failover means a bad day on one provider does not become a bad day on your product.

The awkward position in the lineup

The sharpest criticism of this release is not about benchmarks. It is that Terra occupies a band that its own siblings have squeezed. Hacker News commenters spent the launch arguing that Terra at xhigh effort costs more than Sol at low effort while scoring worse, which makes "pick the cheaper model" a much less obvious rule than the price card suggests. Others pointed out that GPT-6.1 Sol already matches older Astra on some work, and asked what is left for a mid-tier model when the tier below it keeps absorbing capability.

OpenAI's answer is workload shape rather than benchmark rank: routine documents, screenshots, forms, technical drawings, browser and desktop automation, drafting and editing — high-volume tasks where the model needs to be reliable and fast, not maximally clever. That is a real category, and the 37-point error reduction is aimed straight at it. It is also a category where a four-point index deficit to Sol is irrelevant and a $0.32-per-task saving per thousand calls is not.

Two things to watch. OpenAI has not published a shutdown date for gpt-5.6-terra, even though third-party documentation already flags it deprecated and GitHub Copilot retires it on November 16, 2026 with GPT-6.1 Sol or GPT-6 Luna named as migration targets — not Terra. And GPT-6.1 Astra has not shipped; until it does, the top of this family is still GPT-6 Astra, a model Terra is now explicitly benchmarked against at a tenth of the per-token price.

If your work is high-volume and well-specified, the case for moving is strong and the migration cost is a model string. If your work is the hardest 5% — long-horizon coding, deep research, agentic tasks where one wrong step ruins the run — nothing in this release changes your answer, and the four-point index gap is the reason. The price cut is real; it just is not a capability story.