Oct 27, 2026

Article

The most useful AI launch this week isn't the smartest model on the market. It's a mid-priced one that can look at a screen, understand what it sees, and fill in the form. That's the pitch behind GPT-6.1 Terra, which OpenAI released on October 26, and the price is what makes it worth your attention.

What actually happened

OpenAI made Terra available in the API as gpt-6.1-terra and inside ChatGPT Work and Codex for Plus, Pro, Enterprise and Edu users. Standard pricing is $1 per million input tokens, $0.05 per million cached input tokens and $5 per million output tokens. OpenAI says that's half the standard price of both GPT-5.6 Terra, the model it replaces, and GPT-6.1 Sol, the tier above it.

Terra takes text and images, has a context window of roughly 1.05 million tokens, and lets developers dial reasoning effort up or down. Its training knowledge runs to June 1, 2026. There's no free API tier.

What it's built for

OpenAI is unusually specific about the target: high-volume screen work. Its own example is reading incoming invoices and typing the details into accounting software. That's computer use, where a model views a screenshot, decides what to click or type, and repeats until the job is done.

This matters because computer-use agents are expensive by nature. Every step sends a fresh screenshot, and a simple task can take dozens of steps. Halving the per-token price doesn't only make each step cheaper. It changes which jobs make financial sense at all.

What the evidence shows, and what it doesn't

Read this part slowly: every performance number in this launch comes from OpenAI. The company says Terra beats GPT-5.6 Terra on the MMMU-Pro visual reasoning benchmark at every reasoning setting. It says Terra, at medium effort, scores 15 percentage points higher than its predecessor at about a third of the cost, and performs on par with its flagship GPT-6 Astra at a tenth of Astra's per-token price. OpenAI's figures, as published by LiteLLM, put Terra at 50.3% on Terminal-Bench 4.0 at maximum effort, versus 21.5% for GPT-5.6 Terra. On OpenAI's error-inducing prompts at low effort, the error rate fell from 53.6% to 16.7%.

Those are vendor benchmarks. Independent testing hasn't caught up yet, and real workloads rarely look like benchmark suites. Treat the numbers as a reason to test, not a reason to migrate. OpenAI also draws its own boundary: for complex coding, it points developers to Sol or Astra instead.

The fine print that hits your bill

• Long prompts cost more. Above 272,000 input tokens, rates rise to $2 input and $7.50 output per million.
• Batch and Flex processing are 50% off standard rates, so anything that can wait gets cheaper still.
• The faster priority mode costs double.

Where it fits in the market

Terra isn't the cheapest model around. Release trackers list several October models, including DeepSeek V4.1 Flash and MiniMax M3.1 Flash, at well under a dollar per million input tokens. Terra's bet is that reliability on screens and documents is worth paying for, and that teams already inside ChatGPT Work and Codex won't want to stitch together another vendor.

What founders can do with it

Picture a back office that processes 1,000 supplier invoices a month. Assume, for illustration only, that each invoice uses about 3,000 input tokens and 500 output tokens. At standard rates that's roughly half a cent per invoice, or about $5.50 a month. Reasoning tokens and multi-step screenshots can multiply the real figure several times. Even at ten times that estimate, the line item is smaller than a single afternoon of manual data entry.

That's the real story. The cost of AI that operates software is falling below the cost of doing the boring part yourself.

Your moves this week

1. Pick 50 real tasks from your own workflow (forms, invoices, CRM updates) and run them on Terra before believing any benchmark.
2. Route anything non-urgent through Batch or Flex for the 50% discount.
3. Keep complex coding and high-stakes reasoning on the larger tiers, as OpenAI itself recommends.
4. Track cost per completed task, not cost per token. That's the number that decides whether an agent earns its place.