OpenAI Releases GPT-6.1 Terra for Computer Use Tasks

Let's get into what Terra actually is, because the positioning matters. This isn't OpenAI's flagship — that's still Astra. Terra is the workhorse tier, engineered explicitly around computer use: looking at a screen, parsing what's on it, and taking actions like a human would, whether that's clicking through a web form or pulling line items out of an invoice into accounting software.

OpenAI's own benchmarks claim Terra scores 15 points higher than its predecessor on the "Agents' Last Exam" while running at roughly a third of the cost, and that it matches Astra's computer-use performance at a tenth of the price, finishing tasks more than three times faster than both Astra and Sol. Independent testing complicates that story a bit, which is exactly why it's worth discussing. NerdsChalk ran a smaller, controlled test — a 15-field form, three runs per model — and found Terra's median completion time was 9.

4 seconds versus Sol's 17.3 seconds, a meaningful but less dramatic gap than OpenAI's headline number. More interesting: the cheapest model in the lineup, Luna, only got every field right in one of three runs.

That's the real technical story here — speed and cost are climbing fast, but reliability doesn't scale linearly with price cuts, and accuracy is still where cheaper models stumble. On the safety side, OpenAI rated Terra's cyber capabilities as "Critical" on its Preparedness Framework, meaning it ships with the same safeguards built for the full Astra model, despite being the budget option.

The pricing structure tells you exactly who OpenAI is trying to win over. Terra runs $1 per million input tokens and $5 per million output tokens — half the price of GPT-6.1 Sol, and by OpenAI's own internal testing, about one-ninth the per-task cost of Anthropic's Claude Opus 5.

5, even though Opus still outscores it on tasks where the two go head-to-head. That's a deliberate trade: OpenAI isn't trying to out-reason Anthropic right now, it's trying to out-cheap it on the high-volume, low-complexity work that actually dominates enterprise spend — the invoice processing, the data entry, the repetitive clicking that nobody wants to pay premium reasoning-model prices for. This matters because the AI cost curve has become the real competitive battleground.

Every major lab is racing toward the point where running an agent costs less than paying a human to do the task once, and whoever gets there first captures the operational workflows that generate the steadiest, most defensible revenue — not flashy demos, but the boring stuff companies do every single day. The risk, which OpenAI's own engineer reportedly acknowledged, is that Terra isn't a bombshell so much as a way to stretch usage limits for existing subscribers. That's a rational business move, but it also signals margin compression is coming across the mid-tier model market as every lab chases the same cost-per-task curve downward.

Here's the tension: OpenAI is optimizing for breadth and price, while Anthropic is still winning the high end. The Rundown pointed out that Claude's weekly token volume on OpenRouter is up 3.6 times since Opus 5.

5 launched — a cheaper mid-tier OpenAI model isn't going to flip that trend overnight, because Anthropic currently owns the "best at hard things" narrative, especially in coding, where Claude reportedly writes over 90 percent of Anthropic's own internal code. But computer use is a different battlefield entirely, and it's getting crowded fast. Google's Gemini line is undercutting on price, Moonshot's Kimi K3.

1 Pro is pushing an open-weight alternative that benchmarks near Opus 5.5 level, and now Terra is specifically targeting the office-automation use case that vendors like Oracle, with its 25 new Fusion agents, and Meta, with its Muse browsing agent, are also chasing. This is quickly becoming less about who has the smartest model and more about who can reliably automate the most mundane, high-volume digital labor at the lowest marginal cost.

Expect price wars in this specific segment — computer-use and form-filling agents — to intensify through the next two quarters, squeezing margins for anyone who can't differentiate on reliability.

There's a very human undercurrent to a model whose flagship demo is "reading incoming invoices and entering the details into accounting software." It's not glamorous, but it's exactly the kind of repetitive knowledge work that touches millions of jobs — accounts payable clerks, data entry specialists, administrative staff who spend Friday afternoons keying in numbers nobody enjoys typing. One widely shared reaction, flagged in coverage of this launch, joked that "somewhere, an engineer who spent four years learning CAD software just felt a chill" — a nod to a broader anxiety about specialized skills becoming commoditized by agents that are "good enough.

" The social question isn't just job displacement, though — it's trust calibration. OpenAI's own numbers show Terra's hallucination rate dropped from 53.6 percent to 16.

7 percent compared to its predecessor, which sounds like major progress until you sit with the fact that a one-in-six error rate is still being deployed into workflows touching real financial data. As these agents move from novelty to infrastructure, the public's tolerance for "mostly right" answers in consequential tasks is going to be tested, and trust will likely be won or lost not on benchmark scores but on whether people's invoices, forms, and records stay accurate without constant human double-checking.

First, don't take vendor benchmarks at face value — run your own bake-off. The gap between OpenAI's claimed 3x speed advantage and NerdsChalk's measured 1.8x difference is a reminder that your specific workflow is the only benchmark that matters.

Pick one gradable task — form-filling, data cleanup, invoice logging — run it three times on your current model and three times on Terra, and score both accuracy and time before switching anything in production. Second, build verification into the workflow before you build volume into it. Given Terra's "Critical" cyber risk rating and a still-meaningful hallucination rate, any deployment touching financial or customer data needs a human-in-the-loop checkpoint or an automated validation layer — don't let cost savings outrun your audit trail.

Third, segment your AI budget by task complexity rather than defaulting to one model everywhere. Use cheaper, faster models like Terra for high-volume, low-stakes clicking and data entry, and reserve premium reasoning models for judgment calls. That tiered approach, which is exactly what OpenAI is betting enterprises will adopt, is likely to become the standard architecture for AI spend over the next year — and getting ahead of it now saves both money and headaches later.