OpenAI ships GPT-6.1 Terra at half the price, and early users wanted more

A mid-tier model sold on price and computer use

OpenAI released GPT-6.1 Terra, a model it describes as faster and more cost-effective for everyday work including writing, visual analysis, coding and computer use. The company's pitch is about speed and price rather than a capability leap: it says the standard API prices are lower than those of GPT-5.6 Terra and GPT-6.1 Sol.

Announcing the model on X, CEO Sam Altman wrote: "6.1 Terra is here! It is good at a lot of things, but we are especially excited for this new era of computer use." The positioning is unusually specific: this is not sold as the strongest model OpenAI has, but as the one most teams can afford to run at volume.

The price cut is the headline

Standard API prices are 50% lower than those of GPT-5.6 Terra and GPT-6.1 Sol, according to OpenAI. Chinese tech outlet 36Kr reported the figures as $1 per million input tokens, $0.05 per million cached input tokens and $5 per million output tokens, which it described as about 50% and 58% below the previous generation's input and output prices respectively.

That gap is the commercial argument. Tasks such as drafting proposals, interpreting technical drawings, filling out forms, editing videos and processing invoices into accounting software are repetitive and high-volume, so a halved input price changes the economics of automating them far more than a small benchmark gain would.

OpenAI's numbers and the independent index disagree

OpenAI reports improvements on computer-use and visual reasoning benchmarks. It said Terra scored 15 percentage points higher than GPT-5.6 Terra on Agents' Last Exam at medium reasoning, and 36Kr reported a score of 60.0% there, close to Claude Opus 5.5's 63.2%. On Terminal-Bench 4.0, the task completion rate rose from 21.5% to 50.3%.

The independent picture is flatter. Artificial Analysis placed Terra's intelligence index at 48 at its highest reasoning setting, six points above the previous generation but below GPT-6.1 Sol's 52 and well below Claude Opus 5.5's 58, according to 36Kr's summary. OpenAI's cost claim — roughly one-ninth of Opus 5.5's task cost at the top computer-use setting — is a different axis from capability, and the two should not be read as the same result.

Users were waiting for a frontier model

The reaction in the comments 36Kr reviewed was blunt. One commenter questioned OpenAI directly, saying what users really need is a cutting-edge model stronger than GPT-6 Astra; another asked why the company had not published more benchmarks. 36Kr itself summed the model up as roughly mediocre. That is a sample of early commenters, not a measurement of adoption, but it identifies the expectation the launch ran into.

Anthropic has taken much of the attention in coding with Claude Opus 5.5, and users waiting for OpenAI to answer head-on received a model positioned on cost instead. The disappointment is about category, not necessarily about quality: a cheaper workhorse is a different product from a frontier competitor.

The launch landed next to Anthropic's IPO filing

The release came the same day Anthropic formally made its IPO prospectus public, and 36Kr put the gap since OpenAI released GPT-6.1 Sol on September 29 at 27 days. That timing puts a mid-tier price cut in the middle of a week when investors were reading Anthropic's financial disclosure, which is a comparison OpenAI did not control.

For buyers, the practical question is narrower than the rivalry. Terra is priced to be the default for high-volume work, and the release notes the company itself emphasises computer use and office tasks rather than long-horizon reasoning.

What would show whether the price argument holds

Three things would settle it: independent reproductions of the computer-use gains, the cost per completed task rather than per token on real workflows, and whether developers move volume onto Terra rather than keeping frontier models for everything. None of those is visible on launch day.

Related coverage on this site includes GPT-6's ability to turn an answer into a small interactive tool, Claude Sonnet 5.5's trade of scope for everyday speed and Mistral opening the weights of its 1T-parameter flagship. All three describe the same market pressure: models are being separated by price and workload fit, not only by the top of the benchmark table.

What the launch material does not answer

The reviewed sources describe positioning, prices and selected benchmarks, and they leave basic specifications out. Neither report gives the context window, the release date for the API beyond availability, or a latency figure, which are the numbers an engineering team checks before moving production traffic onto a cheaper model.

The benchmark picture is also incomplete in a specific way. A score on Agents' Last Exam and a completion rate on Terminal-Bench 4.0 say something about professional tasks and coding agents, and nothing about long-context retrieval, tool reliability over many steps, or behaviour under rate limits. Those are the conditions where a mid-tier model usually shows its limits first, and they are not in the launch material.

One number in the launch material is easy to misread. A 50% cut applies to standard API prices, and the larger discount quoted for output tokens is a comparison with the previous generation rather than with every model on the market. Teams that run cached prompts, batch jobs or long outputs will see a different blended cost from the headline figure, which is why the pricing table matters more than the percentage in the announcement.