GPT-6.1 Terra: One Line in a Microsoft Pull Request, and No Model to Call

The most specific public evidence about G​PT-6.1 Terra is a single changed value in a Microsoft evaluation harness. On October 16, 2026, a pull request merged into a Microsoft repository set judge_model to the string gpt-6.1-terra across a set of evaluation configuration files. That is not an announcement, not a model card and not a rate card. O​penAI has published nothing under that name. What O​penAI does publish is GPT-5.6 Terra, the mid-tier model it shipped on July 9, 2026 at $2 per million input and $12 per million output tokens, and GPT-6.1 Sol, the September 29, 2026 release at $2 and $10. G​PT-6.1 Terra sits between the two in the naming scheme and, as of today, nowhere else. Five days after the pull request merged, The Information's Stephanie Palazzolo reported that a new Terra was in the works, that it should outperform GPT-6.1 Sol at a lower price, and that launches of this kind might slip into November. This page is about the distance between those two facts: a string in a config file and a paywalled sentence about a model that has not shipped. The documented AI incidents and agent failures that model launches like this one sit alongside are collected in the OrcaRouter AI Incident Archive, where each entry carries its own sources, severity, confidence grade and disputed flag.

What the pull request actually contains

The change is small enough to describe completely. The diff sets judge_model: gpt-6.1-terra, and alongside it an executor field reading copilot-sdk, a judge reasoning effort of medium, and a ten-minute timeout. A judge model in an evaluation harness is the model that scores other models' outputs — the thing that decides whether a candidate answer was good. Swapping it is an infrastructure decision, made by a team that needed something capable enough to grade other models on a deadline.

The write-ups that read the diff — Lukas Kretschmann's and LuminaBench's, both published October 19 — put the scope at eighteen configuration files, seventeen of which had previously named GPT-5.6 Terra as their judge. That ratio is the interesting part: it reads as a like-for-like replacement rather than a new capability being wired in. RuntimeWire, checking the same material the same day, noted that the eighteen-config figure had not been independently verified by anyone outside the people reading the same diff. Treat the count as reported rather than confirmed. The string itself is public and checkable; the interpretation of what eighteen means is not.

What the pull request does not contain is just as important. No endpoint. No price. No context window. No benchmark. No launch date, no deprecation notice for GPT-5.6 Terra, no changelog entry anywhere in O​penAI's documentation. A configuration identifier is a name a system is willing to send traffic to, which is weaker evidence than it first appears: internal harnesses routinely point at working labels that later ship under different names, or not at all. As one third-party tracker put it when checking the same week, this is a model identifier that leaked out of an evaluation config, not a model you can call today. That distinction is the whole article.

What The Information added, and what it did not

Palazzolo's report, posted October 19 at 2:52 PM, named a potential new Terra model from O​penAI among a set of expected October releases and said it should outperform GPT-6.1 Sol at a lower price, framed as competing with A​nthropic on cost. It also carried its own hedge: that launches might slip into November. TestingCatalog picked the thread up on October 20, pairing the paywalled report with the Microsoft pull request — which is how the two independent traces became one story.

Two traces converging is meaningfully better than either alone. A config identifier says someone inside a partner organisation was pointing at the name; a reported roadmap says the company behind the name has a plan for it. But neither trace is a specification, and the second one is paywalled, which means what circulates publicly is the summary rather than the reporting. No date, no price, no context window, no benchmark, no parameter count has been published by anyone. Everything below that is arithmetic and inference is labelled as such.

Three names, three different statuses

The disambiguation matters more than usual here, because two of these names have shipped and one is a string in a diff, and a search engine will cheerfully return any of them for the others.

• GPT-5.6 Terra — released July 9, 2026, repriced July 30. Listed at $2.00 per million input tokens, $0.20 cached input, $2.50 cache write and $12.00 output, with a 1,050,000-token context window and 128,000 maximum output tokens. Past 272,000 prompt tokens the input price doubles and the output price rises by half, to $4.00 and $18.00. It is callable today, and on our catalogue it carries exactly those vendor list prices.

• GPT-6.1 Sol — released September 29, 2026. Listed at $2.00 input, $0.10 cached input, $2.50 cache write and $10.00 output, with the same 1,050,000-token context, the same 128,000 maximum output and the same 272,000-token cliff. Artificial Analysis scores it 55 on the Intelligence Index, 71.8 on coding, 99 on math and 88 on science, as displayed on its model page and read on October 20, 2026. Callable today.

• G​PT-6.1 Terra — no release date, no entry in any vendor documentation, no rate card, no context window, no benchmark score, no model card. A string in a merged pull request, a sentence in a paywalled report, and the subject of everything that follows.

The price arithmetic, and why "cheaper than Sol" is a bigger claim than it sounds

In the 5.6 generation, Terra was the middle of three tiers: Luna at $0.20 input and $1.20 output, Terra at $2 and $12, Sol at $4 and $20. In the 6.1 generation there is one tier. GPT-6.1 Sol launched at $2 and $10 — the same input price as the older Terra, with a lower output price. The generation's only published model already sits where the previous generation's middle used to be.

That makes the reported positioning awkward in a useful way. If G​PT-6.1 Terra is meant to be the middle of its generation and it is meant to undercut GPT-6.1 Sol, it cannot land at $2 and $12. Twelve dollars of output is twenty per cent above Sol's ten, which would make the cheaper model the more expensive one on the line most workloads are billed on. For the reported claim to hold, the new model has to come in below $2.00 input and below $10.00 output, or Sol has to move first and make room.

Below Sol there is a real gap. GPT-6 Luna — released September 22, 2026, a week before Sol — is $0.10 input and $0.50 output. Between fifty cents and ten dollars of output per million tokens, O​penAI's current price list has nothing. A mid-tier landing somewhere in that band, say around $1 and $4, would be a genuinely new price point rather than a relabelled one, and it would explain why the reported framing emphasises cost rather than capability. This is arithmetic on published prices, not a forecast. But it does tell you what to check when a number finally appears: if the announced price lands at $2 and $12, the model is not a mid-tier undercut and the report meant something looser than it sounded.

What you can call today, and what a leaked name changes about it

Two of the three names above are live right now on our catalogue: openai/gpt-6.1-sol and openai/gpt-5.6-terra. Both run through one API key, at the providers' own list prices with no markup, so a vendor price cut shows up in the catalogue figure the same day rather than whenever a reseller gets around to it. If either model is rate-limited or degraded, routing can fail over to another provider for the same identifier without the application noticing. G​PT-6.1 Terra is not on the catalogue, is not callable through us, and we have no date for it.

That gap matters for planning. Anyone who needs 6.1-generation quality at a lower price has exactly one option today, and it is Sol. Anyone who needs the Terra price point has GPT-5.6 Terra, which is a generation behind and still bills twelve dollars per million output tokens — the exact figure a new Terra would have to beat. There is no third door, and a name in a Microsoft config file does not open one.

What would confirm this, and what would falsify it

Confirmation arrives in a fairly predictable order, and the order is what makes it useful. First the identifier appears in O​penAI's own documentation or model list — that single step turns a config value into a product listing, and it is usually the earliest unambiguous signal. Then a vendor rate card with a price and a context window, which makes the model real even before anyone announces it. Then an Artificial Analysis entry, which is the first number produced by someone who is not the vendor. And finally, a second unrelated organisation pointing configuration at the same name: one Microsoft repository is a data point, three would be a pattern.

Falsification is not hypothetical, and it comes in three strengths. The mild version is timing: the report's own caveat about November covers it. The middle version is naming — the model ships, the shape is right, and the string on the rate card is not the string in the diff, because the identifier in a working config was a label rather than a brand. The strongest version is that G​PT-6.1 Terra never exists as such, and the next mid-tier model arrives under a different name or as a re-tuned GPT-5.6 Terra with a new suffix. The string alone cannot distinguish new weights from a new label; nothing short of a model page can.

What to do with this in the meantime

The honest summary is one sentence long. A Microsoft evaluation harness started grading with a model called gpt-6.1-terra on October 16, 2026, a reporter with credible sources says a new Terra is coming and may undercut Sol, and nobody outside those circles has seen a single specification. That is enough to watch and not enough to plan around. If a decision has to be made this week, the 6.1 generation is GPT-6.1 Sol and the mid-tier price point is GPT-5.6 Terra, both callable now at list price. The cheap preparation for being right about the next one is to keep the model name in configuration rather than in code, because that is the only form of preparation that survives the name changing.