OpenAI · GPT-6.1 · Computer Use · October 2026

GPT-6.1 Terra: The Model You Buy When the Job Is Volume, Not Difficulty

OpenAI released GPT-6.1 Terra on October 26, 2026, and the pitch is deliberately unglamorous: not the smartest model, but the cheapest model that is good enough for the work that fills a day. At $1 per million input tokens and $5 per million output tokens, it is half the price of GPT-6.1 Sol and a tenth of GPT-6 Astra. The question this page answers is not whether Terra is impressive — it is whether Terra is enough for your specific job, and what the 272,000-token surcharge does to the bill when it is not.

Last updated: October 27, 2026. Pricing and availability checked against OpenAI's October 26, 2026 announcement and model documentation as reported by multiple outlets; benchmark figures are vendor-reported unless stated otherwise.

At a Glance

The Price Ladder, in One Table

All four current tiers share a 1.05M-token context window, so the choice is about capability per dollar, not context size. Prices below are standard API rates per million tokens.

A worked example makes the gap concrete. A job using 10,000 input tokens and 2,000 billable output tokens costs $0.02 on Terra (10,000 x $1 + 2,000 x $5, per million), $0.04 on Sol, $0.20 on Astra, and $0.002 on Luna. The absolute numbers look trivial; multiply by thousands of jobs a day and the tier choice becomes the budget. One caveat: reasoning tokens that the user never sees are still billed as output tokens, so a model that reasons longer can cost more than its visible answer suggests.

When Terra Is Enough

Use Terra when the task is well-specified, repetitive, and verifiable — the kind of work where you can check the result cheaply and retry cheaply. OpenAI's own example is reading incoming invoices and typing the details into accounting software. The same logic covers: drafting proposals from a template, reading technical drawings for specified facts, filling in forms from source documents, routine screen-based automation across applications, and first-pass triage of documents where a human or a stronger model reviews the edge cases.

The deciding test is cost per acceptable finished result, not cost per token. If Terra completes the job on the first attempt at half Sol's price, it wins outright. If Terra needs three attempts and a human fix, Sol's higher token price may still be the cheaper job. Run both on the same 50 real jobs before committing a pipeline to either tier.

When You Must Pay for Sol (or Astra)

Pay up when failure is expensive, the spec is ambiguous, or the work is genuinely hard: complex multi-file coding, deep research across conflicting sources, novel problem-solving with no template, and anything where a wrong answer ships to a customer before a human sees it. OpenAI says Sol and Astra remain better choices for the most complex coding tasks, and an OpenAI staff comment framed Terra as better than Sol and Astra at technical communication, similar on vision and computer use (but much faster), and a bit worse on most other things. That is a useful rule of thumb: Terra for communication and screen work, Sol for hard reasoning.

A practical split works well: Terra does the volume pass, Sol handles the exceptions. Route the routine 90% to Terra, escalate failures, low-confidence outputs, and flagged edge cases to Sol, and keep Astra for the rare job that justifies $50-per-million output. You get most of Terra's savings without betting the whole pipeline on the cheaper tier.

The 272,000-Token Rule, and Other Fine Print

The surcharge is the one pricing detail that can quietly erase Terra's advantage. Above 272,000 input tokens, the entire request — not just the tokens past the threshold — bills at $2 input, $0.10 cached input, and $7.50 output per million. A job that looked like a Terra bargain at headline rates can cost nearly as much as Sol once context balloons. If your agent carries full conversation history, large documents, or many tool outputs, measure actual input tokens per request before forecasting spend, and test context trimming or job splitting against the surcharge.

Two more qualifications matter. First, cached input at $0.05 per million is what makes repeated-context agents affordable — but cache writes are billed separately, so the saving depends on how often the same context is reused. Second, Batch and Flex modes are reported at 50% below Standard rates, which can push Terra to $0.50 input / $2.50 output for work that does not need an immediate answer. Neither changes the tier decision, but both change the monthly bill.

The Honest Caveats

Three things to keep straight. First, the benchmark gains are OpenAI's own figures: 50.3% on Terminal-Bench 4.0 at maximum effort versus 21.5% for GPT-5.6 Terra, and an internal difficult-prompts error rate down to 4.3% at maximum effort. Artificial Analysis adds an independent data point — 6 points higher on its Intelligence Index at maximum effort, at about $0.40 per task versus $1.40 — but broad independent testing was not yet out at launch. Second, Terra is not in regular Chat, so a ChatGPT subscriber who never opens Work or Codex gains nothing from this release. Third, prices here are API prices: they do not change what a Plus or Pro subscription costs, and plan allowances still govern how much Work and Codex usage you get.

If you are building agent workflows rather than picking a model for one-off chat, pair this page with our Cindy guide on harness-switching agents and our Claude Duet guide on proactive agents — the model is only half the cost decision; the harness and the escalation policy are the other half. For voice-agent work on the same stack, see our Voxtral 3 guide.

FAQ