OpenAI’s GPT-6.1 Terra is the most useful kind of release: not the smartest model in the lineup, but the one that makes a specific job affordable. That job is computer use, meaning clicking through forms, reading screens and pushing invoice data into accounting software. If you’re building an agent that drives a desktop, Terra should be your new default. I’d still want to see one more number before calling it the best option on the market.

The per-task math

Pricing is $1 per million input tokens and $5 per million output, with cached input at $0.05. That’s half of GPT-6.1 Sol ($2/$10) and a tenth of GPT-6 Astra ($10/$50). Batch and Flex cut it in half again.

Per-token price is only half of it, though. On Agents’ Last Exam v1, OpenAI’s benchmark of professional computer tasks (one involves turning architectural drawings into a 3D model), Terra at medium effort scores 57.6% at $0.29 per task. Astra scores the same 57.6%, at $4.82. GPT-5.6 Terra managed 42.6% at $0.84. So the new model is better than its predecessor and cheaper per task, and it ties the flagship at about a sixteenth of the cost.

At max effort Terra hits 60.0% for $1.01 per task, against Claude Opus 5.5’s 63.2% at $8.82. Three points for roughly nine times the money is a bad trade for anything repetitive. OpenAI also says Terra finishes tasks more than three times faster than Astra and Sol, which matters as much as cost when an agent is stepping through a UI one screenshot at a time.

The hallucination numbers stand out too. On OpenAI’s adversarial error-inducing prompt set, low-effort errors fell from 53.6% to 16.7%. That’s an internal benchmark built to trip models up, so treat it as directional. For unattended form-filling, though, fewer confident mistakes is the feature I care about most.

Where the chart goes quiet

All of these scores are self-reported, and the comparison set is selective. On OSWorld 2.1, the desktop-agent benchmark, Terra scores 74.2% at max. Sol gets 78.0%, Astra 80.9%, and Claude Sonnet 5.5 gets 83.9%. Sonnet 5.5 is priced at $2/$10, the same as Sol.

Sonnet is Terra’s natural rival. Simon Willison points out that Terra occupies the same middle slot in OpenAI’s lineup that Sonnet holds in Anthropic’s. Yet on the Agents’ Last Exam chart, where per-task cost is front and center, Sonnet 5.5 isn’t shown. OpenAI compared Terra to Opus, a model priced at $4/$20, and won handily on cost. The comparison buyers actually need is missing.

Coding is the clearer weakness. Terminal-Bench 4.0 has Terra at 50.3%, a big jump from 21.5% for GPT-5.6 Terra, but well behind Sonnet 5.5’s 70.6%. OpenAI itself steers hard coding work to Sol or Astra. OpenAI’s Ted Sanders was refreshingly blunt on Hacker News, calling it “not a bombshell release” for subscribers and pitching it as a midrange API option that’s “a lot smarter than Luna” for vision, computer use and communication.

That’s the right way to read it. Terra is a workhorse for screen and document automation, with a 1.05 million token context window. Watch the billing, though: prompts above 272K input tokens cost 2x on input and 1.5x on output, so long agent transcripts get pricier fast.

Willison’s max-effort pelican cost 11.3161 cents. What would Sonnet 5.5 cost per task on Agents’ Last Exam?