GPT-6.1 Terra Pro Is Not an OpenAI Model: What the Name Actually Points To

Search for GPT-6.1 Terra Pro and you will find it listed, priced and described, which is exactly why the name deserves a careful look. OpenAI does not sell a model under that name. The model that shipped on October 26, 2026 is GPT-6.1 Terra, and every figure attached to “Terra Pro” on third-party catalogues traces back to that one model, its one price card and its one API identifier, gpt-6.1-terra. There is no second set of weights behind the longer name.

The distinction matters beyond pedantry. A model ID is what you put in a request, what your bill is keyed to and what a router resolves. A phantom suffix is the kind of detail that turns into a failed deployment, a budget scoped around a product that does not exist, or a benchmark comparison between a model and a label. So here is what actually shipped, where “Pro” comes from, and how to check which model you are calling before you commit a workload to it.

What OpenAI actually shipped on October 26

GPT-6.1 Terra is the mid-tier model in the GPT-6 line, replacing GPT-5.6 Terra and sitting between GPT-6 Luna and GPT-6.1 Sol, with GPT-6 Astra above it. OpenAI’s announcement describes it as a fast, efficient model built for detailed visual reasoning, precise computer use and clearer, more natural writing.

• Price — $1 per million input tokens, $0.05 per million cached input tokens, $5 per million output tokens.

• Context — 1,050,000 tokens, with a 922,000-token maximum input and a 128,000-token maximum output.

• Knowledge cutoff — June 1, 2026.

• Modalities — text and image input, text output.

• Reasoning effort — none, low, medium, high, xhigh and max. The none setting is Terra-specific; GPT-6.1 Sol’s floor is low.

• Endpoints — Chat Completions, Responses and Batch. Batch and Flex are priced at 50% of standard.

• Availability — the API, plus ChatGPT Work and Codex for Plus, Pro, Enterprise and Edu plans. OpenAI’s announcement does not list the main Chat picker as a surface.

Two pricing details are easy to miss. Prompts above 272,000 input tokens are billed at twice the input and cache rate and 1.5 times the output rate for the whole request. And cache writes cost $1.25 per million tokens at 30-minute retention, or $2.50 at 12-hour retention, which is a different number from the $0.05 cached-input rate.

Where the “Pro” label comes from

OpenAI does use the word “Pro”, just not as a model suffix on Terra. Three separate things can produce the string a reader meets.

• ChatGPT Pro is a subscription family. As of October 27, 2026, OpenAI’s help centre lists three monthly plans: Pro 100 at $100, Pro 200 at $200 and Pro 500 at $500, above Plus at $20. “Pro” there describes the plan, not the weights.

• OpenAI has used “Pro” for a ChatGPT configuration rather than a model. GPT-6-Astra-Pro is the clearest precedent: an entitlement that changes which GPT-6 Astra configuration a Pro subscriber gets, with no separate model card and no separate API identifier. Coverage adopted the name; the API sells only gpt-6-astra.

• Third-party catalogues append their own labels. A listing can carry a plan name, a speed tier or a segment marker next to the vendor’s model name, and the result reads like a product.

Put those together and “GPT-6.1 Terra Pro” is best understood as the model name crossed with the plan name. No OpenAI model page, announcement, price card or help-centre article uses it. The vendor’s own documentation resolves gpt-6.1-terra and returns 404 for gpt-6.1-terra-pro.

The real model, by the numbers

The independent picture is narrower than launch coverage suggests, and more useful. Artificial Analysis measured GPT-6.1 Terra at an Intelligence Index of 48 at maximum reasoning effort, ranked 25th of 221 models, at $0.40 per index task and 90.9 output tokens per second.

• GPT-6.1 Terra at max effort — index 48, rank 25, $0.40 per task, 90.9 tokens per second, 131.8-second time to first token.

• GPT-6.1 Terra at medium effort — index 41.

• GPT-6.1 Terra with reasoning off — index 33.

• GPT-5.6 Terra at max effort — index 42, rank 53, $1.40 per task.

• GPT-6.1 Sol at max effort — index 52, rank 13, $0.72 per task.

• Coding Agent Index v2.0 — Terra 48, Sol 54, GPT-6 Luna 31.

So the honest summary is a six-point gain over its predecessor on the independent index, a four-point deficit to Sol, and a cost per completed task roughly a third of the model it replaces. The “half the price” line is true per token and roughly true per task, but it is not a capability claim. On the independent measurements, Terra is a cheaper model that is better than the one it replaces and still below the model above it.

What the vendor claims, and what that claim is worth

OpenAI’s launch numbers are vendor-reported and none has been independently reproduced yet. They are still the most detailed read on what the model is for, provided the label stays attached.

• Agents’ Last Exam — 15 percentage points above GPT-5.6 Terra at medium effort at about one-third the cost, and on par with GPT-6 Astra at one-tenth its price per token (OpenAI-reported).

• OSWorld 2.1 — at medium effort, above GPT-5.6 Terra’s best score at a fraction of the cost per task (OpenAI-reported; no direct Sol figure published).

• MMMU-Pro — above GPT-5.6 Terra at every reasoning setting, and nearly matching GPT-6 Astra and GPT-6.1 Sol at lower cost (OpenAI-reported).

• Terminal-Bench 4.0 — 50.3% at max effort against 21.5% for GPT-5.6 Terra (OpenAI-reported). OpenAI still names Sol and Astra as the better choice for the hardest coding tasks.

• Terminal-Bench-Science 0.1 — above GPT-6.1 Sol at max effort and lower cost, within five points of Claude Opus 5.5 at under one-fifth the estimated cost (OpenAI-reported).

• Internal error test — the error rate at low effort fell from 53.6% to 16.7%; at max effort it fell 18 points to 4.3%, which OpenAI says nearly matches Astra and Sol (OpenAI-reported).

• Speed — more than three times faster than GPT-6 Astra and GPT-6.1 Sol on OpenAI’s tested tasks (OpenAI-reported).

The one claim with an independent check is speed: Artificial Analysis’s 90.9 tokens per second is consistent with “fast”, though its 131.8-second time to first token is a reminder that throughput and responsiveness are different numbers. A model can produce tokens quickly and still take two minutes to begin.

Why a phantom name is a routing problem

If you paste gpt-6.1-terra-pro into a request, you are asking a provider for a model that does not exist. The failure is usually a 404 or a validation error, which is easy to catch in development and much less pleasant in a production path that was scoped, priced and approved around a name someone read in a catalogue.

The same problem applies to benchmarks. A comparison that pits “Terra Pro” against Sol is comparing a real model to a label, and any number attached to the label is really Terra’s number wearing a longer name. That is not a small accounting error when the entire decision is whether to pay twice as much for four more index points.

This is also where a routing layer earns its place. OrcaRouter exposes more than 200 models through one API at provider list price with no markup, so a vendor price change is live on the same terms rather than renegotiated, and automatic failover can move a request off a model that is erroring or rate-limited. The honest version of that pitch is that a router can only resolve models that exist. GPT-6.1 Sol is in our catalogue at $2 input and $10 output. GPT-6.1 Terra is not yet, as of October 27, so the correct route for Terra today is OpenAI’s own API. No router should be able to sell you a Terra Pro.

How to check which model you are actually calling

Five checks settle it, and none takes more than a minute.

• The identifier. gpt-6.1-terra resolves in OpenAI’s model documentation; gpt-6.1-terra-pro does not.

• The response object. Read the model field on the response rather than trusting the request you sent. Aliases and routing layers can resolve to a different string.

• The context window. Terra returns a 1,050,000-token window with a 128,000-token output ceiling; a materially different window means you are on a different model.

• The reasoning floor. Terra accepts none; Sol’s lowest setting is low. A model that rejects none is not Terra.

• The price. $1 input and $5 output is Terra’s standard card. A “Pro” surcharge is not.

What to do if you searched for Terra Pro

If you wanted the cheaper capable OpenAI model for computer use, document work and high-volume automation, you want GPT-6.1 Terra, at $1 and $5. If you wanted the strongest coding and long-horizon reasoning in the line, you want GPT-6.1 Sol at $2 and $10, or GPT-6 Astra above it. If you wanted a faster tier inside ChatGPT, “Pro” is a plan and an allowance, not a model, and the only Ultrafast options OpenAI documents are on GPT-6 Astra and GPT-6.1 Sol.

The name will keep circulating because it is a plausible collision of two real things. The model behind it is real, one day old and already well specified. The suffix is not.