GPT-6.1 Terra vs GPT-5.6 Terra: The Migration You Get Whether You Plan It or Not

Two migration notices went out this month, and they point in different directions. OpenAI's deprecation table lists GPT-5.6 Terra as the recommended replacement for older models including gpt-5-mini and o4-mini. GitHub Copilot, meanwhile, retires GPT-5.6 Terra on November 16, 2026 and names GPT-6.1 Sol or GPT-6 Luna as the migration targets — skipping over GPT-6.1 Terra entirely. OpenAI, in other words, is still steering traffic into the model that one of its largest distribution partners is steering traffic out of. Terra's own launch on October 26 is the third notice, and it is the only one that argues for staying in the middle.

This page is about the upgrade itself: what changed between the two Terra generations, what the change is worth in money, and what a migration actually involves beyond swapping a model string. It is not a launch recap — the model shipped yesterday and the launch coverage is already everywhere. The interesting question is whether a team with GPT-5.6 Terra in production should move to GPT-6.1 Terra, to Sol, or wait.

The delta, dimension by dimension

Everything below is from OpenAI's model documentation for the two models, read on October 27, 2026.

• Price — GPT-6.1 Terra $1 input / $5 output per million, against $2 / $12. Cached input drops from $0.20 to $0.05.

• Context — 1,050,000 tokens on both, 922,000 maximum input, 128,000 maximum output. Unchanged.

• Knowledge cutoff — June 1, 2026, against February 16, 2026. Roughly three and a half extra months of world.

• Reasoning effort — the same six settings on both: none, low, medium (default), high, xhigh, max. This is one place Terra inherits a Sol advantage: GPT-6.1 Sol has no "none" setting.

• Modalities and endpoints — identical. Text and image in, text out; Chat Completions, Responses and Batch; tool calling through Responses only; no fine-tuning on either.

• Cost modifiers — identical structure on both: batch and flex at 50% below standard, fast at 2×, regional at +10%, and the 272K-input threshold that doubles input and cache rates and multiplies output by 1.5 for the whole request.

So the specification sheet says: same shape, half the price, three more months of knowledge. Everything else is benchmark claims.

What OpenAI says the generation bought

The launch numbers are vendor-run and none of them have been independently reproduced yet. Stated as claims, not findings:

• Terminal-Bench 4.0 — 21.5% solved by GPT-5.6 Terra at maximum effort, 50.3% by GPT-6.1 Terra. This is the biggest relative jump in the release and the one most likely to matter to anyone using Terra for shell work or repository tasks.

• Error-inducing prompts — error rate at low effort falls from 53.6% to 16.7%, and to 4.3% at maximum effort. The test is OpenAI's own, unpublished, and constructed to provoke mistakes; it measures the floor, not the ceiling.

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

• MMMU-Pro — ahead of GPT-5.6 Terra at every effort level, approaching Astra and Sol, with no absolute scores published.

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

Read together, these describe a model that got much better at the tasks Terra was already sold for — computer use, screenshots, documents, terminal work — rather than one that moved into Sol's territory. That is consistent with how OpenAI positions it: a fast, efficient model for visual reasoning and precise computer use, not a reasoning flagship.

The independent check is less dramatic

Artificial Analysis currently lists GPT-5.6 Terra at an Intelligence Index of 42, ranked 53rd of 221 models, at $1.40 per index task and $2/$12 per million tokens. GPT-6.1 Terra sits at 48, ranked 25th, at $0.40 per task and $1/$5. That is a six-point gain and a 71% cut in cost per finished task — a good upgrade, and a smaller one than the vendor's Terminal-Bench number implies.

Both listings are current as of October 27 and are only comparable to each other on the same scale; the index is renormalized periodically, and older write-ups quote different absolutes for these models. The same page notes that GPT-5.6 Terra is marked deprecated and that only its default 10k-token benchmark continues to update, which is a quiet signal that the predecessor's published scores are becoming a snapshot rather than a live measurement.

Speed is the unresolved column. GPT-5.6 Terra is measured at 137 output tokens per second with 111.75 seconds to first token. Artificial Analysis has no throughput figure for GPT-6.1 Terra yet. OpenAI describes the new model as substantially faster, and if the claim holds then the migration improves latency as well as cost — but on this machine, today, that is a claim with no independent number behind it.

Rewriting the cost line, four ways

The headline is "half price," and on the standard tier that is exact: $1 against $2 input, $5 against $12 output. Output is where the saving is larger — a 58% cut — so output-heavy workloads do better than the headline.

A 100,000-input, 5,000-output request: $0.26 on GPT-5.6 Terra, $0.125 on GPT-6.1 Terra. A 400,000-input, 4,000-output request, which trips the long-context multiplier on both: $1.67 against $0.83. The multiplier is proportional, so it never eats the discount; it just makes both models more expensive than their headline rates suggest.

Batch and flex halve both: $1/$6 on the old model, $0.50/$2.50 on the new one. For any pipeline that can tolerate a turnaround window, this is the cheapest configuration in the comparison and the first thing to test.

Fast mode inverts the story. At 2× standard, GPT-6.1 Terra costs $2/$10 — exactly GPT-6.1 Sol's standard list price, for a model four index points weaker. If your workload needs the fast setting, the price argument for Terra over Sol collapses and the decision becomes purely about output quality. That is the single most important line in this comparison and it does not appear on either model's pricing page.

What you actually have to change

The migration is mostly a string swap, but four things deserve a check before you ship it.

• Model identifier and snapshot — the new ID is gpt-6.1-terra and there is one snapshot, aliased to the base name. If you pin dated snapshots today, you are moving to a floating alias, which is a policy change as much as a model change.

• Tool calling path — both models require the Responses API for tool calls. If you have been calling tools through Chat Completions on GPT-5.6 Terra, it was already unsupported; the migration is a good moment to confirm which endpoint your SDK is really hitting.

• Effort defaults — medium is the default on both, and the error-rate claims are made at low effort. If you run at none or low to save money, re-run your eval before assuming the 37-point improvement applies to your setting, because OpenAI's numbers move a long way between effort levels.

• Long-context billing — the 272K threshold is unchanged, but your prompts may have grown. A prompt that crossed it on the old model costs the same relative penalty on the new one, so the discount is real but the cliff is still there.

Nothing in the move requires a modality change, a context-window change, or a new tool. That is the good news: this is one of the cleaner generational swaps OpenAI has shipped, because the interface is genuinely identical.

Running the old arm and the exit arm on one key

Here is the practical problem with this comparison: the model you are leaving and the model you may be moving to are not necessarily the same one. GitHub Copilot's migration names GPT-6.1 Sol and GPT-6 Luna, not Terra. So a team that trusts its distribution partner ends up comparing GPT-5.6 Terra against Sol, while a team that trusts the vendor's recommendation ends up on Terra.

OrcaRouter carries both of the arms you can actually measure today — GPT-5.6 Terra at $2/$12 and GPT-6.1 Sol at $2/$10, at 0% markup over provider list price, on one key. That makes the real question testable: run your own traffic through the old Terra and through Sol, compare quality per dollar, and only then decide whether GPT-6.1 Terra's middle position is worth adopting. GPT-6.1 Terra itself is not on our catalogue as of October 27, so it has to be trialled against OpenAI's API directly; automatic failover on our side keeps the Sol arm available as a fallback while you do.

The 0% markup point is not decorative here. Terra's launch halved the list price of the model you are running, and on a pass-through router that price change reaches your bill without a renegotiation. On a platform that marks up inference, the same cut arrives as a smaller discount and a new contract.

Who should move, and who should wait a fortnight

Move now if your workload is the one Terra is aimed at: screenshots, forms, document extraction, browser and desktop automation, high-volume drafting. The price halves, the knowledge cutoff moves forward three and a half months, the context window is unchanged, and the vendor's largest claimed gain — Terminal-Bench — is in exactly the kind of task you run. The independent six-point gain supports the direction even if it deflates the size.

Wait if you are on GPT-5.6 Terra for hard reasoning and picked it as a cheaper Sol. In that case the honest comparison was never against the old Terra; it was against Sol, and Sol is still four index points ahead of the new Terra at twice the input price and the same output price as Terra's fast mode. Watch the first independent speed measurements, and watch what OpenAI does about a shutdown date — none is published for gpt-5.6-terra today, which is why the November 16 date in this article is Copilot's, not OpenAI's.

And if you are one of the teams OpenAI is pushing off gpt-5-mini and onto GPT-5.6 Terra: skip a step. Moving onto a model that a major distributor is already retiring in three weeks is not a migration, it is a delay.