Stop Counting Tokens—GPT-6.1 Terra Is the Cheapest Frontier Model When Measured by Task

OpenAI has released GPT-6.1 Terra. Bindu Reddy's positioning is very direct: the frontier model with the lowest cost per successful task. Its overall score is close to Argon and Sonnet; browser operations are a strength, making it suitable for personal agents and chatbots.

First, let's explain the metric of 'cost per successful task.' Frontier models are usually billed by token, but for the same task, different models take different numbers of steps; a low token unit price does not mean it is truly cheap. Terra's selling point is making 'how much it costs to complete a task' the comparison standard.

The most striking thing in the table is not the scores, but the cost. Terra's price per successful task is clearly lower than several other models. What really got developers discussing was the pricing method itself.

Comparing prices by token is meaningless

Several developers expressed the same view in replies: for the same task, if the model takes a few extra steps, a lower token unit price is useless. What should be looked at is how much it costs to complete a task.

One user did the math: Terra costs $0.670 per successful task, while Sonnet 5.5 costs $3.718. Another user mentioned that in OpenAI's own data, Terra's cost per task is about 9 times cheaper than Opus 5.5. Someone else said that people using Sonnet in personal agents will quietly switch over as long as Terra holds up on browser operations, without posting to celebrate.

Personal agents spend most of their time clicking

Agents spend most of their time handling browser operations, not reasoning. One user proposed this division of labor: Terra handles the browser, and tasks that are hard to judge are passed to a larger model. Others pointed out that being cheap plus strong at browser operations is exactly what personal agents are missing; price was originally the limiting factor.

One developer put the division of labor more concretely: Terra runs the agent, Opus 5.5 writes code.

The counterintuitive finding in speed tests

Kajetan Mastela ran a set of speed tests: a higher tokens-per-second rate does not mean faster output. Sonnet 5.5's token rate is 96.6, higher than Terra Fast's 74.7; but Terra Fast actually delivers 266.2 characters per second of text, higher than Sonnet's 232.6. The reason is that the two models use different tokenizers; a token is not a uniform unit of text.

Terra Normal mode averages 192.8 characters per second, and Fast mode averages 266.2 characters per second; the test includes waiting and reasoning time. Fast is 38% faster than Normal, but uses 2.5 times the subscription usage, so Normal is more cost-effective for everyday use.

Those not yet convinced

One user asked: in the cost per successful task, are failed tasks counted in the denominator? Someone warned of a pricing trap: after the prompt exceeds 272K tokens, the entire request is billed at 2x input and 1.5x output; with a large window, the degree of cheapness has to be discounted. Another person said he would rather see 'cheapest per correct task.' Someone else asked: so can it book a dinner reservation yet?

GPT-6.1 Terra did not take first place on every benchmark, but it got two practical things right: cost and browser operations. For people running personal agents, these two things matter more than scores. Some users are also waiting for something else: for Sol to get a bit faster.