TL;DR
OpenAI’s GPT-6.1 Terra combines visual reasoning and computer use at published API rates of $1 per million input tokens and $5 per million output tokens. Lower token prices matter, but the meaningful business metric is the cost of a successfully completed, reviewed task—not the price of a click.
A cheaper model is not automatically a cheaper workflow
On October 26, OpenAI released GPT-6.1 Terra for developers through its API and made it available to specified ChatGPT Work and Codex plans. The model targets the mix of visual interpretation, computer interaction and writing that makes up many office workflows. (Source : OpenAI — GPT-6.1 Terra; OpenAI — API changelog)
This is more consequential than a simple model refresh. Agents that operate graphical interfaces repeatedly interpret screenshots, decide on an action, wait for the interface to respond and inspect the result. A task that looks like one instruction to a user may require many model calls and consume a substantial amount of context. Lowering a component price can therefore change which workflows are practical to test.
It does not follow that the price of every action falls by half. OpenAI’s published prices are token rates, not a fixed price per successful computer action. The number and size of calls, retries, cached context and validation work all influence the final bill.
Start with the published rates
OpenAI lists standard pricing of $1 per million input tokens, $0.05 per million cached input tokens, and $5 per million output tokens. The company describes its standard API prices as 50% lower than those of GPT-5.6 Terra and GPT-6.1 Sol. For an operational estimate, use the actual input, cached-input and output rates separately instead of assuming that every line item decreases by the same percentage. (Source : OpenAI — Pricing and availability)
Consider an illustrative task that uses one million uncached input tokens and 100,000 output tokens. At the listed rates, the model-token component is $1.50. That example is not a benchmark or a typical job: it excludes cached inputs, any separately priced tools, hosting, monitoring, human review and failed attempts. It is useful because it exposes the levers a team can measure rather than treating a token discount as an outcome guarantee.
A comparison with another model should use the same task set and record how many tokens each system actually consumes. The cheaper rate can lose its advantage if the agent requires more steps, repeats work or fails more frequently.
Benchmark gains, with the right caveats
On Agents’ Last Exam, OpenAI reports that Terra at medium reasoning effort scores 15 percentage points higher than GPT-5.6 Terra at roughly one-third the cost. On Terminal-Bench 4.0, the company reports 50.3% at maximum reasoning, versus 21.5% for the predecessor’s published score. OpenAI also reports lower error rates on an internal set of error-inducing prompts. These are vendor-reported results under specified settings, not proof that a particular enterprise workflow will improve by the same margin. (Source : OpenAI — GPT-6.1 Terra)
The distinction between task cost and token price matters here. Model behavior can change how much work is needed: a higher-performing agent may finish with fewer corrections, while a more deliberate reasoning setting can spend more tokens. A valid experiment tracks the full trajectory, the number of successful outcomes and the intervention rate, not just a leaderboard score.
A practical scorecard for buyers
Run the old and new models on the same representative tasks, with the same permissions and a clear definition of completion. Include difficult cases: ambiguous invoices, missing approvals and unexpected interface changes. Record total model tokens, tool costs, elapsed time, reversals and the time a human spends reviewing or fixing the result.
Then calculate cost per accepted outcome: total measured operating cost divided by the number of outcomes that pass the agreed review. Report the failure rate alongside that figure. A low average cost can hide a small number of expensive mistakes, particularly when an agent can send messages or change financial records.
Caching can lower input costs, but it only helps when the relevant input is eligible for the provider’s caching rules. Treat it as a measured benefit, not a blanket discount. Likewise, speed can increase throughput, but faster execution is useful only if the workflow has controls on irreversible actions.
The deployment decision
Start with read-only or draft-producing tasks. For actions such as sending a purchase order or changing a payment, require explicit authorization and a clear audit trail. Evaluate prompt injection as a workflow risk: an agent may encounter hostile instructions in content it is asked to read. OpenAI reports resistance in its own tests, but those results are not a substitute for testing the specific tools, data and permissions in use. (Source : OpenAI — Safety and alignment)
The important shift is not that desktop automation suddenly has a universal price. It is that better visual models and lower published token rates make another round of carefully measured automation experiments worth running.
FAQ
Is computer use billed at half price per click?
The published figures are token rates. They do not establish a fixed per-click charge or guarantee a 50% reduction in the cost of a completed task.
Can Terra replace human review?
Not on the evidence presented. Review requirements should reflect the risk and reversibility of each action, and should be tested in the intended environment.
Which cost metric matters most?
Cost per accepted outcome, accompanied by failure rate and review time, is more informative than price per token alone.
Further Reading
— The Agent Report