OpenAI's biggest week was in the cheap middle of the market

Facts are attributed and linked to their sources. The analysis and opinions are the author's own.

What it is

On 26 October OpenAI released GPT-6.1 Terra, a fast model aimed at computer use, visual reasoning and writing. Its standard API price is $1 per million input tokens, $0.05 per million cached input tokens and $5 per million output tokens, which OpenAI says is 50% lower than GPT-5.6 Terra and GPT-6.1 Sol. The example OpenAI gives is plain office work: "reading incoming invoices and entering their details into accounting software".

OpenAI's own figures show a large accuracy gain. On its internal benchmark of error-inducing prompts, the error rate at low reasoning effort fell from 53.6% with GPT-5.6 Terra to 16.7%, and to 4.3% at maximum effort. On Terminal-Bench 4.0 at maximum effort it solves 50.3% of tasks against GPT-5.6 Terra's 21.5%. OpenAI also says Terra "meets the Critical threshold in cybersecurity under our Preparedness Framework".

Four days earlier OpenAI made its Decisions API generally available. It is built on GPT-6 Luna for fast, bounded answers such as choices, scores and probabilities, handles up to 400 questions per request, and costs $0.10 per million input tokens with no charge for output tokens. API product lead Nikunj Handa told TechCrunch that it is "shrinking things from one to three seconds to 100, 120, 180 milliseconds".

Why it matters

Most AI coverage tracks the top of the range. For anyone running operations, the middle matters more: that is where invoice capture, document checks, routing and triage happen. Halving the price of a capable mid-tier model, while cutting its error rate sharply, changes which tasks are worth automating. These are OpenAI's numbers, though, and independent evaluations should follow before anyone budgets on them.

When a model's decision arrives in about 150 milliseconds, the slow part of the pipeline becomes the people who must validate it.

The Decisions API is the more interesting product for finance, in my view. Fast, bounded scores look a lot like what risk teams already build with traditional models. But a language model's answer is not a validated scorecard. Teams using it for anything that touches customers will need the usual model-risk discipline: documented inputs, stability testing, monitoring for drift and a clear owner. The Critical cybersecurity rating is a reminder that cheaper capability is cheaper for attackers too, which is why the formal-verification work in this issue matters.

What I'd watch

• Independent measurements of Terra's error rates on business documents rather than OpenAI's internal benchmark.
• Whether Decisions API outputs are stable enough, call to call, to pass a model validation review.
• How rivals respond on mid-tier pricing over the next month.