OpenAI's New Terra Model Cuts AI Agent Costs 50% — What Owners Should Test Before Switching

AI agents that click through software for you just got cheaper. On October 26, 2026, OpenAI released GPT-6.1 Terra, a mid-tier model built for computer use, visual reasoning, and writing, priced at $1.00 per million input tokens and $5.00 per million output tokens. OpenAI says that is 50% below both GPT-5.6 Terra and GPT-6.1 Sol.

For a small business, the headline is not the model name. It is the cost of letting software handle repetitive screen work: entering invoices, filling forms, and reading documents. Below is what was announced, which claims come from OpenAI alone, and a low-risk way to test it.

The Price, in Plain Terms

OpenAI's standard API pricing for Terra is:

• Input: $1.00 per million tokens.
• Cached input: $0.05 per million tokens.
• Output: $5.00 per million tokens.

For comparison, GPT-6.1 Sol is listed at $2 input and $10 output per million tokens, which is where the "50% cheaper" figure comes from. A token is roughly a small chunk of a word, so a million tokens covers a very large amount of text. Most owners will never buy tokens directly; they will feel the price through the software that uses the model.

Where You Can Get It

Terra is available in the API and in ChatGPT Work and Codex on Plus, Pro, Enterprise, and Edu plans. GitHub also made it generally available on October 26 in Copilot for Pro, Pro+, Max, Business, and Enterprise, with a gradual rollout and the same list pricing. OpenAI says subscription plans "stretch further" with Terra but has not published exact usage multipliers.

What OpenAI Says It Can Do

These are OpenAI's own claims, not independent results:

• On computer-use tasks at medium effort, Terra scores 15 points higher than GPT-5.6 Terra at about one-third the cost.
• It performs on par with the older GPT-6 Astra at one-tenth of Astra's per-token price.
• It completes tasks more than three times faster than Astra and Sol.
• In an internal test using error-inducing prompts, the error rate at low effort fell from 53.6% to 16.7%, and at max effort it was 4.3%.

OpenAI also says Sol and Astra remain better for the most complex coding, so Terra is pitched as the high-volume, lower-cost option rather than the smartest one.

The Independent Picture Is More Mixed

Artificial Analysis, an independent benchmarking service, scored Terra 48 on its Intelligence Index, ranking it 25th of 221 models against a class median of 28. That is strong. But two details matter for your budget. Running the index cost about $0.40 per task, ranking 32nd of 120 on cost. And Terra generated 130 million output tokens on the test versus a median of 81 million, meaning it is wordy. Output tokens cost five times more than input tokens, so a verbose model can eat into the headline savings.

One coding benchmark OpenAI cites, Terminal-Bench 4.0, has faced an audit questioning its task quality, so do not lean on that figure.

What a Small Business Could Automate

The launch coverage does not address small businesses directly, but the use cases it highlights map cleanly onto owner headaches:

• Invoice entry: reading incoming invoices and keying the details into accounting software.
• Form filling: completing repetitive forms in screen-driven workflows.
• Document review: interpreting drawings or scanned documents using visual reasoning.
• Drafting: proposals, customer replies, and everyday business writing.

If you already track your books carefully, an agent that drafts entries still needs your review. Tie any automation back to your profit and loss statement so mistakes show up before they compound.

A Five-Step Test Before You Switch

1. Pick one boring, repeatable task

Choose something you do dozens of times a week with a clear right answer, such as entering vendor invoices.

2. Test on 50 to 100 real examples

Analysts reviewing the launch suggest testing on your own tasks rather than trusting benchmarks. A sample that size shows how often the agent gets it wrong.

3. Measure cost per completed task

Do not compare price per token. Divide your total spend by the number of tasks finished correctly. A cheaper model that needs retries or runs long can cost more.

4. Keep a stronger model as a fallback

Route the hard cases to a more capable model and let Terra handle the easy volume.

5. Test for prompt injection

OpenAI reports strong resistance to prompt-injection tests, but that is a vendor claim. If your agent reads emails, web pages, or uploaded files, a malicious instruction hidden in that content could redirect it. Limit what the agent can access and require your approval before it sends money or messages.

Cost and Risk Beyond the Model

Cheaper automation does not erase compliance duties. If an agent touches customer data or payments, check your business insurance for cyber coverage questions, and keep AI-handled transactions flowing through a dedicated business bank account so everything is traceable.

The Bottom Line

Terra lowers the price of automating screen-based busywork, and the independent ranking suggests it is genuinely capable. But the biggest claims are OpenAI's own, the model is wordy, and exact plan limits are unpublished. Run a small pilot, measure the real cost per finished task, and keep a human checking the output. If the numbers hold, the savings are real. If not, you will have spent an afternoon, not a quarter.

Photo by Igor Omilaev on Unsplash