GPT-6.1 Terra vs Kimi K3.1 Pro: Seven Index Points for Twice the Price

GPT-6.1 Terra and Kimi K3.1 Pro sit two rungs apart on the same leaderboard and two factors apart on price. Artificial Analysis puts Kimi K3.1 Pro at 55 on its Intelligence Index — fourth overall — and GPT-6.1 Terra at 48, twenty-fifth. Kimi costs $2 per million input tokens and $10 per million output on the OrcaRouter catalogue; Terra costs $1 and $5. The closed, cheaper, faster model is seven points less capable; the open-weight challenger is twice the price and, on desktop automation, 5.4 points better. Neither of those facts settles the purchase.

The comparison is worth doing carefully because the two models are not substitutes in the way the price suggests. They differ on output length by a factor of seven, on licensing by open versus closed, and on where they sit in their vendors' lineups. This piece prices the same workloads on both, marks which numbers are independent and which are vendor claims, and ends with the split we would actually use.

The cards

• GPT-6.1 Terra — $1 input / $5 output per million, $0.05 cached input, 1,050,000-token context, 922,000 maximum input, 128,000 maximum output, knowledge cutoff June 1, 2026, six reasoning-effort settings, text and image in, text out.

• Kimi K3.1 Pro — $2 input / $10 output per million on OrcaRouter, 1,048,576-token context, 943,718 maximum output, open weights promised by November 4, 2026, released October 23, 2026.

The context windows are effectively identical at about a million tokens. The output ceilings are not: Terra stops at 128,000 tokens, Kimi allows 943,718. For chat that difference is invisible; for long-form generation, whole-file code output or transcript-scale summarisation, it is the difference between one call and a chunking pipeline.

Why seven index points is a lot here

On a leaderboard that spans roughly 38 to 60, seven points is more than a rounding difference — it is the gap between 25th and 4th. GPT-6.1 Sol, the model OpenAI sells as the high-performance tier, sits at 52. Kimi K3.1 Pro at 55 is above it. GPT-6 Astra, the flagship, is at 53. So the open-weight model from Moonshot is not competing with Terra; on the composite it is competing with OpenAI's top two models and winning by two to three points.

On the dimensions that matter for agents, the pattern holds. On OSWorld 2.1, Kimi K3.1 Pro scores 79.6% against 74.2% for Terra — a 5.4-point gap, and above GPT-6 Astra's 75.8%. On coding, coverage reports describe Kimi as closely trailing Claude Opus 5.5, which at 58 is the highest score in this comparison. Terra's Coding Agent Index of 48 is six points below Sol and 29 below Astra; Kimi is not in that neighbourhood.

The counterweight is that Terra is not trying to win those comparisons. OpenAI's positioning is bounded work — forms, screenshots, invoices, drafting — where 48 is enough and $0.40 per task is the point. A seven-point intelligence gap matters most on the tasks Terra was not sold for.

Price math: when does twice the price win?

A standard 100,000-input, 5,000-output request costs $0.125 on Terra and $0.25 on Kimi. A 400,000-input, 4,000-output request, above the long-context threshold, costs $0.83 on Terra and $1.66 on Kimi. If both models solve a task in one attempt with the same token count, Terra is exactly half the price and there is nothing else to discuss.

The interesting case is when they do not use the same number of tokens or attempts. Take a document-extraction workload where Terra succeeds on the first attempt 80% of the time and Kimi on 92%. With a flat 3,000-input, 400-output call per attempt, Terra's expected cost per success is about $0.0055; Kimi's is about $0.0092. Terra still wins on cost — because the accuracy gap would have to be enormous for a 2× token price to be overcome by fewer retries.

Now take an agentic coding task where the models need 30,000 input and 2,000 output tokens per attempt, and the failure mode is a wrong patch rather than a retry. Terra at 48-level coding on a task that needs 77-level coding does not fail 20% more often; it fails qualitatively, producing changes that run but do not solve the problem. That is not a retry cost, it is an engineer-hour cost, and it is the case where twice the price is the cheap option.

Open weights change the second year, not the first

Moonshot has promised Kimi K3.1 Pro weights by November 4, 2026. That matters for three reasons and none of them is today's invoice. First, self-hosting removes the per-token price from the marginal cost calculation entirely, replacing it with GPU-hours — which is cheaper at sustained high volume and much more expensive at low volume. Second, open weights remove the vendor as a single point of failure, which is a procurement argument rather than a technical one. Third, a downloadable model can be fine-tuned, evaluated internally and pinned to a version, none of which is possible with a closed API.

Two caveats belong next to the promise. The weights are not out yet, so every statement about them is a claim. And the released model is not identical to the API model: reports say Moonshot stripped cyber-related training data before release, which changes behaviour on security-adjacent prompts — a deliberate choice, and one that means the open weights and the API endpoint are not interchangeable for every workload.

GPT-6.1 Terra has no equivalent path. It is closed, API-only, not available for fine-tuning, and not offered under any self-host arrangement. If portability is a requirement, the comparison is over before the benchmarks start.

Latency is not comparable, and that is the point

Kimi K3.1 Pro shows a p50 of about 6.95 seconds on its OrcaRouter model page — a platform-level latency figure across real traffic. GPT-6.1 Terra's published latency number is 133.64 seconds to first token on Artificial Analysis's max-effort reasoning benchmark, which measures how long the model thinks before answering a hard question, not how long a typical request takes. Putting those two numbers side by side would be meaningless.

What can be compared is throughput on the same benchmark: Artificial Analysis measures Terra at 90.9 output tokens per second and GPT-6.1 Sol at 73.5. Kimi's equivalent figure is not on the same page. The honest position is that Terra is demonstrably fast among OpenAI models and Kimi's speed on your workload is something you measure rather than read.

Where each wins

• Terra — high-volume bounded tasks where a 74% computer-use pass rate is acceptable behind a review step; document and invoice pipelines; drafting and editing; anything where the newer June 2026 cutoff helps and $0.40 per task compounds.

• Kimi K3.1 Pro — agentic coding and long-horizon work where a seven-point intelligence gap changes the outcome; desktop automation where 79.6 beats 74.2; long-output generation where a 943,718-token ceiling removes a chunking layer; and any deployment that needs open weights, self-hosting or a pinnable version.

• Neither — unsupervised, irreversible computer use. Claude Opus 5.5 at 81.8 on OSWorld is the leader, and the gap to both of these models is larger than the gap between them.

Running them side by side

Kimi K3.1 Pro is on the OrcaRouter catalogue at $2/$10 with a 1,048,576-token context, and it can be called through the OpenAI SDK by pointing the base URL at our endpoint. GPT-6.1 Terra is not on the catalogue at the time of writing, so the Terra arm has to be trialled directly against OpenAI.

That asymmetry is temporary but useful: it means the Kimi arm can be measured on live traffic today, at list price with no markup, alongside the rest of the catalogue — GPT-6.1 Sol, GPT-6 Astra, Claude Opus 5.5, Claude Sonnet 5.5 — and the Terra arm can be measured in parallel against OpenAI's API. When Terra lands, the comparison is already built, and the routing layer can send each task to whichever arm wins on it rather than to whichever model a blog post ranked higher.

The split we would use

If the task is bounded, repeated and reviewed, Terra's half price beats Kimi's seven points. If the task is open-ended, code-shaped or needs to run without a vendor, Kimi's seven points beat Terra's half price — and the open weights make the second year cheaper than the token math suggests.

The mistake is picking one model for the whole workload. The evidence supports a split: Terra for the volume tier, Kimi for the hard tier, and a measured fallback between them. That is a routing decision, not a model decision, and it is the one this comparison actually earns.