GPT-6.1 Terra's Benchmarks, Sorted: Where 48 Ties 52 and Where It Doesn't

The composite number for GPT-6.1 Terra is 48. GPT-6.1 Sol is 52, GPT-6 Astra is 53, GPT-5.6 Terra is 42, and GPT-6 Luna is 38. That is the whole story if you read only the top line of the Artificial Analysis Intelligence Index — a mid-tier model that lands six points above its predecessor and four below its more expensive sibling. It is also the least useful way to read the release, because the same index hides a 29-point coding gap to Astra, a 7.6-point computer-use gap to Claude Opus 5.5, and a hallucination improvement that is larger than every intelligence gain in the launch combined.

What follows separates three kinds of number: independent measurements from Artificial Analysis, vendor-reported results from OpenAI that nobody outside has reproduced, and derived figures — the arithmetic that turns either of the first two into a cost or a ratio. The distinctions matter, because the launch coverage has already started blending them.

The three scores everyone will quote

Artificial Analysis lists GPT-6.1 Terra at 48 on its Intelligence Index, ranked 25th of 221 models, at $0.40 per index task. GPT-6.1 Sol sits at 52 and 13th, at $0.72 per task. GPT-6 Astra, the current flagship, is at 53. GPT-5.6 Terra is at 42 and 53rd, at $1.40 per task. GPT-6 Luna is at 38.

Read as a ladder, Terra is a genuine upgrade: six points for less than a third of the cost per finished task, and ten points above the budget tier. Read as a comparison against Sol, it is exactly what the price card implies — four points weaker, at roughly half the token price. Nothing in the composite contradicts OpenAI's positioning.

The composite also flattens things it should not. Four index points between Terra and Sol is 8% of the scale; 29 points between Terra and Astra on the Coding Agent Index is 60% of that scale. A single number cannot carry both, and the launch coverage that quotes only the first has quietly dropped the second.

Where the gap widens: coding

On Artificial Analysis's Coding Agent Index v2.0, GPT-6.1 Terra scores 48, GPT-6.1 Sol scores 54, and GPT-6 Luna scores 31. That is a six-point deficit to Sol and a 17-point lead over Luna — the same shape as the intelligence index, but with the gap to the top of the family larger in relative terms.

The vendor numbers tell a more flattering story, and they are vendor numbers. OpenAI reports 50.3% solved on Terminal-Bench 4.0 at maximum effort, against 21.5% for GPT-5.6 Terra. That is more than double, and it is the single largest generation-over-generation jump in the release. It is also a terminal-task benchmark, not a broad coding-agent measure: Terra can be close to Sol on short, well-scoped shell work and still be six points behind on the wider index, and both things are true at once.

The practical reading is that Terra's coding gains are concentrated where the task is bounded. "Run this test suite, read the failure, patch the file" is Terra's shape. "Hold a fifteen-file refactor in your head for forty steps" is not, and OpenAI's own recommendation still points complex programming work at Sol or Astra.

Where the gap is smaller than the index suggests: computer use

On OSWorld 2.1, the standard desktop-automation benchmark, GPT-6.1 Terra scores 74.2% at maximum reasoning. GPT-6.1 Sol scores 75.9% at xhigh. GPT-6 Astra scores 75.8%. Claude Opus 5.5 leads the models we can compare at 81.8%, with Claude Fable 5.1 at 80.7%, Claude Sonnet 5.5 at 80.1% and Kimi K3.1 Pro at 79.6%.

Two things follow. First, Terra is within 1.7 points of Sol on the task it was built for, at half the token price — a much better trade than the four-point intelligence gap implies. Second, the whole GPT-6 family trails Anthropic on desktop automation by six to seven points, and Terra does not close that gap. If your product is a browser agent, the interesting comparison is not Terra against Sol; it is Terra against Claude Sonnet 5.5, which is cheaper per token than Sol and scores 5.9 points higher on OSWorld.

The vendor's own agentic numbers point the same way. On Agents' Last Exam v1, OpenAI reports 57.6% for Terra at medium effort, against 50.7% for GPT-5.6 Terra at maximum. The claim attached to it — comparable to GPT-6 Astra at medium, at roughly one-tenth the per-token price — is a per-task claim, and it is the most economically meaningful sentence in the launch.

The number that actually moved: hallucination

Artificial Analysis measures an Omniscience hallucination rate: how often a model guesses when it does not know. GPT-5.6 Terra scored 88%. GPT-6.1 Terra scores 38%. Accuracy rose two points over the same period.

A 50-point drop in guessing is bigger than any intelligence gain in this release, and it is the one number that changes what a model is safe to use for. A model that invents an answer 88% of the time when uncertain cannot be trusted with invoice extraction or form-filling without a human check on every output. At 38%, it still guesses more than a third of the time — this is an improvement, not a solution — but the failure mode moves from "assume it is wrong" to "sample and verify."

That has a direct cost consequence. Verification is the hidden line item in every cheap-model deployment: if you have to check 88% of answers, the labour dwarfs the token saving. At 38%, spot-checking becomes viable, and the $0.40-per-task figure starts to look like the real cost rather than the visible half of it.

Vendor claims, and what is missing from them

OpenAI's published results, all vendor-run and none independently reproduced at the time of writing:

• Terminal-Bench 4.0 — 50.3% at max effort, against 21.5% for GPT-5.6 Terra.

• 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.

• Error-inducing prompts — 16.7% at low effort against 53.6%, and 4.3% at max, on an internal test the company has not published.

The pattern in that list is worth naming. The two strongest claims — MMMU-Pro and the error-rate test — are the two with no absolute numbers attached. "Approaches Astra" and "close to Sol" are comparative statements that cannot be checked without the numbers behind them. The claims with hard figures are the ones a third party can reproduce, and Artificial Analysis has already reproduced the cost side: $0.40 per task, $1.40 for the predecessor.

Speed is the independent number that is not in the index

Artificial Analysis measures GPT-6.1 Terra at 90.9 output tokens per second with 133.64 seconds to first token on its reasoning benchmark. GPT-6.1 Sol is measured at 73.5 tokens per second and 284.31 seconds. That is roughly 24% more throughput and a little over twice the time-to-first-token advantage — an independent measurement, not a vendor claim, and the strongest objective case for Terra in the whole release.

For a chat product, throughput barely matters; for a computer-use loop that takes twenty steps per task, it is the difference between a 40-second task and a 90-second one. The index does not price latency, which is why a model ranked 25th can be the right buy for an agent that ranks 13th on quality.

How to read the index without being misled

• Treat the composite as a filter, not a ranking. It tells you which models are in the conversation; it does not tell you which one wins your task.

• Look at the coding and computer-use sub-scores separately. Terra's shape is unusually wide: near-Sol on desktop automation, six points behind on coding, four behind overall.

• Re-check absolutes against a single date. The index is renormalized periodically, and write-ups from different weeks quote different numbers for the same model.

• Weight the hallucination rate as heavily as the intelligence score if your workload has a verification cost. A 50-point improvement changes the operating model, not just the score.

• Measure your own task. The cost-per-task figures from Artificial Analysis are a proxy built on their prompts; your prompts will produce a different ratio between token price and attempts.

None of this makes GPT-6.1 Terra a bad model. It makes it a specific one: a fast, cheap, much-less-hallucinating model that is close to its more expensive sibling on screens and far from the flagship on code. The composite says 48. The sub-scores say which 48 you are buying.