1. OpenAI releases GPT-6.1 Terra at half the API price
OpenAI released GPT-6.1 Terra on October 26, an update to its middle-tier model that the company says costs developers 50% less through standard API pricing than GPT-5.6 Terra while completing computer-based tasks faster and with fewer mistakes. The model is rolling out to paying ChatGPT users and is available to developers under the name gpt-6.1-terra.
Terra occupies the middle of OpenAI's lineup. It is meant to be the everyday workhorse for routine office work: reading what is on a screen, moving through applications, and drafting replies. Two terms matter for understanding the announcement. The first is computer use, which describes an AI system that looks at a screen and then clicks and types much as a person would. The second is standard API pricing, the per-use rates that developers pay when they build the model into their own products. It is this second category that OpenAI has cut in half.
In its announcement, OpenAI described GPT-6.1 Terra as a significant step up from GPT-5.6 Terra. The company paired that claim with the lower price, a combination that is unusual in a market where newer models have often arrived at equal or higher cost. A cheaper model that also performs better changes the economics for any business running large volumes of automated tasks.
Taken together, the announcement rests on three claims: lower cost, higher accuracy and better performance on agent-style work. Each is supported by figures that OpenAI supplied, and each is examined in the sections below. The model is aimed at the large share of work that does not require a flagship system, which is where most commercial usage of AI is concentrated.
2. Who can use GPT-6.1 Terra, and where?
Access begins with OpenAI's paid ChatGPT tiers. The model is rolling out now to Plus, Pro, Enterprise and Edu subscribers, and it is available inside ChatGPT Work and Codex. Developers who want to call it directly can do so through the API using the model name gpt-6.1-terra.
ChatGPT Work is OpenAI's agent mode within ChatGPT. It operates across applications, can browse websites and can carry out multi-step tasks on a user's behalf. Codex is the company's coding tool. Both are settings in which a model must act rather than merely answer, which is why improvements in computer use and command-line performance are central to this release.
Users on the free tier are not part of the rollout. They continue to receive GPT-5.6 Luna, a lower-tier model in the same family. The split follows a pattern in how OpenAI organizes its products, with the cheapest model reaching the widest audience and the more capable ones reserved for subscribers and developers who pay.
Together, these access points show that GPT-6.1 Terra is aimed at working professionals and developers first, with the broader free audience served by a different, smaller model for now.
3. How much did the factual error rate fall?
The most striking number in the announcement concerns accuracy. OpenAI reports that, in an internal test built from prompts designed to trick models into making mistakes, the factual error rate at low reasoning effort dropped from 53.6% for GPT-5.6 Terra to 16.7% for the new model. Reasoning effort is a setting that controls how long a model spends thinking before it answers, so low effort represents the fastest and cheapest way to run it.
At maximum reasoning effort, the error rate fell by 18 percentage points, to 4.3%. The announcement does not state the earlier baseline at that setting, although the size of the drop implies a starting point of roughly 22%. Because the baseline is not given, the maximum-effort comparison is harder to evaluate than the low-effort one.
These figures come from OpenAI's own testing, and the evaluation is internal. Independent researchers have not yet reproduced the results, so the numbers are best read as the company's claim rather than as verified measurements. Adversarial test sets also differ widely in difficulty, which means an error rate on one set says little about performance on another.
To put the low-effort result in plain terms, the earlier model produced a factual error on more than half of the deliberately tricky prompts, while the new one does so on roughly one in six. That is a large change, but it describes a hard test set. Performance on everyday questions, which are typically easier, would be expected to show much smaller absolute differences.
4. What do the Terminal-Bench 4.0 results show?
On coding and command-line work, OpenAI points to Terminal-Bench 4.0. The benchmark tests coding agents on tasks carried out in a command-line environment and scores the share of tasks they complete successfully. At maximum reasoning effort, GPT-6.1 Terra solved 50.3% of the tasks, compared with 21.5% for GPT-5.6 Terra. That is more than double the earlier model's completion rate.
OpenAI also makes a speed claim. According to the company, Terra finishes tasks on average more than three times faster than GPT-6 Astra and GPT-6.1 Sol, two models positioned above it in the lineup. Speed matters for agents because each step of a task, such as reading a screen or running a command, adds delay, and a faster model can complete long chains of steps in a practical amount of time.
As with the error-rate test, these results are reported by OpenAI itself. Benchmarks give a useful common yardstick, but results published by a model's developer reflect the conditions the developer chose, including the reasoning settings and the tools made available to the agent.
The jump from 21.5% to 50.3% is notable because command-line tasks leave little room for approximation. A command either runs correctly or it fails, and an agent that misreads one step can derail an entire job. A higher completion rate therefore points to better planning and fewer small errors along the way, which is the quality that agent users care about most.
5. How does Terra fit into OpenAI's recent releases?
The model has a short but eventful history. OpenAI first introduced Terra on June 26 as part of the GPT-5.6 family, initially as a limited preview that the Trump administration had requested. The company opened Terra to the public on July 9, describing it at the time as competitive with GPT-5.5 at about half the cost. The new release repeats that logic of delivering more capability at a lower price.
In September, OpenAI launched GPT-6 Sol and Luna within minutes of Anthropic's release of Claude Opus 5.5. Terra was not updated at that point, leaving the middle tier on the older generation while the top and bottom tiers moved ahead. GPT-6.1 Terra closes that gap.
The release also shows how OpenAI divides labor among its models. For in-depth research or complex coding, the company still directs users to GPT-6.1 Sol or GPT-6 Astra. Terra is intended for the large middle ground of everyday tasks where cost and speed count as much as peak intelligence.
6. Why does a 50% price cut matter for developers?
For businesses that build products on top of AI models, the price per use is a direct input cost. A company running an automated support desk or an agent that processes documents all day may send millions of requests a month, and halving the rate reduces that bill in proportion. A cheaper rate can also make applications viable that were previously too expensive to operate at scale.
Computer-use agents are especially sensitive to cost. Because they work through a task one screen at a time, a single job can involve dozens of model calls. Lower prices combined with fewer errors mean fewer failed attempts that must be rerun, so the savings can compound beyond the headline 50% figure, although that depends on whether the reported accuracy gains hold in real deployments.
The competitive context is also relevant. With OpenAI and Anthropic releasing models in close succession, pricing has become one of the main ways to win developer adoption. A mid-tier model that is both cheaper and better puts pressure on rivals to respond at the same price point.
Finally, price changes of this size can shift how teams choose among models. Organizations often route simple tasks to a cheaper model and reserve the most capable and expensive one for difficult problems. A stronger mid-tier model widens the set of tasks that can be handled at the lower price, leaving fewer jobs that justify the top tier.
7. What should users watch next?
The main open question is whether independent testing confirms OpenAI's numbers. The error-rate figures come from an internal evaluation, and the benchmark scores were published by the company, so third-party results will show whether the improvements carry over to ordinary work.
Users should also watch how the model behaves in agent settings such as ChatGPT Work and Codex, where reliability across long sequences of actions matters more than any single benchmark score. For tasks that demand deep research or complex software engineering, OpenAI itself recommends stepping up to GPT-6.1 Sol or GPT-6 Astra, so Terra is best understood as a fast, low-cost option for routine work rather than a replacement for the flagship models.
Ultimately, the practical test will be how the model performs for the people who use it daily. Subscribers on the Plus, Pro, Enterprise and Edu plans will be the first to see whether faster, more accurate computer use shows up in their routine workflows, and developers will see the cost effect directly on their invoices.