GPT-6.1 Terra: OpenAI halves API prices and its new model guesses far less

OpenAI has a new middle child. On Monday, October 26, the company introduced GPT-6.1 Terra, the successor to GPT-5.6 Terra and the first new model in the Terra line since the GPT-6 generation started. It sits below GPT-6.1 Sol and the frontier model GPT-6 Astra, and above the small GPT-6 Luna. On paper that sounds like a routine refresh. In practice it is one of the more interesting releases of the autumn, because it changes the price of everyday AI work more than it changes the top of any leaderboard.

The headline number is simple: standard API prices are 50 percent lower than those of GPT-5.6 Terra and GPT-6.1 Sol. At the same time the model scores higher than its predecessor on independent tests, makes far fewer factual errors and solves more than twice as many command line coding tasks. This article looks at what OpenAI shipped, what the independent numbers from Artificial Analysis say, where the model still falls short and what the price cut means for people who build with these models.

What OpenAI released

GPT-6.1 Terra is described by OpenAI as a fast, efficient model that combines detailed visual reasoning, precise computer use and clearer, more natural writing. The company frames it around the tasks that fill a normal workday: drafting proposals, interpreting technical drawings, filling out forms, reading invoices and entering their details into accounting software, or editing videos.

It is available in three places. In ChatGPT it appears in the Work and Codex experiences for Plus, Pro, Business, Enterprise and Edu plans, according to the release notes. It is not part of the regular Chat mode. In the API it is called gpt-6.1-terra and works with both the Responses API and Chat Completions, as listed in the developer documentation. Tool calling requires the Responses API.

The technical profile is generous for a mid range model. The context window is 1,050,000 tokens, with up to 922,000 input tokens and 128,000 output tokens. It accepts text and images, but not audio or video, and it outputs text. The knowledge cutoff is June 1, 2026. Reasoning effort can be set from none through low, medium, high and xhigh up to max, with medium as the default. On the tool side it supports web search, file search, code interpreter, a hosted shell, computer use, MCP servers and skills.

One practical detail for ChatGPT users: OpenAI says some Plus and Pro users may now see GPT-6.1 Terra at medium effort instead of GPT-6.1 Sol at light effort in the default Power presets of Work. Overall usage limits stay the same, but because Terra is cheaper to run, the same allowance covers more tasks than before.

The price cut in numbers

The new list prices are $1 per million input tokens, $5 per million output tokens and $0.05 per million cached input tokens. Writing to the cache costs $1.25 per million tokens with the standard 30 minute retention. For comparison, Artificial Analysis lists GPT-5.6 Terra at $2 for input and $12 for output, with cached input at $0.20. The Decoder puts Anthropic's Claude Sonnet 5.5, the most direct competitor in this class, at $2 and $10.

The cached input price stands out. It is a quarter of the old rate and only 5 percent of the uncached input price. For applications that send the same long instructions, documents or tool definitions again and again, such as customer support agents or coding assistants that keep a big repository in context, caching is often where most of the bill sits. A cut of this size there can matter more than the headline input price.

There are a few fine print rules worth knowing. Prompts longer than 272,000 input tokens are billed at twice the input and cache rates and 1.5 times the output rate for the whole request. Batch and Flex processing are 50 percent cheaper than standard. Regional processing adds 10 percent where it is offered, and the Fast mode costs twice the standard price.

Independent tests: more intelligence for less money

Benchmarks published by a model maker are always worth a second opinion, and in this case one arrived on launch day. Artificial Analysis ran GPT-6.1 Terra through its Intelligence Index, which combines ten evaluations covering knowledge work, coding, science and long context reasoning.

At maximum reasoning effort, Terra scores 6 points higher than GPT-5.6 Terra, while one index task costs $0.40 instead of $1.40. That is less than a third of the cost for a better result. According to the analysts, at lower effort levels the gain over the predecessor is even larger, between 7 and 13 points. The low and medium settings sit on what Artificial Analysis calls the Pareto frontier: no other model they measured delivers more intelligence at that cost.

In the overall ranking, Terra at max effort lands 4 points below GPT-6.1 Sol and 10 points above GPT-6 Luna. In other words, it fills the gap between OpenAI's big general model and its small one much better than its predecessor did. The individual model pages on Artificial Analysis also show how much the effort setting matters: at low effort Terra reaches 36 points for about five cents per task, at high effort 44 points for about 13 cents. That spread lets developers choose very precisely how much quality they want to pay for.

Fewer guesses, fewer errors

The most striking improvement is not about speed or price, it is about honesty under uncertainty. Artificial Analysis runs a private benchmark called AA Omniscience that checks both knowledge and hallucination. One of its measures is how often a model guesses when it does not actually know the answer. GPT-5.6 Terra guessed in 88 percent of those cases. GPT-6.1 Terra guesses in 38 percent. At the same time its accuracy on the benchmark went up by 2 points, so the drop in guessing did not come from simply refusing more questions.

OpenAI reports a similar trend on its own internal set of prompts designed to trigger mistakes. At low reasoning effort, the error rate fell from 53.6 percent with GPT-5.6 Terra to 16.7 percent. At maximum effort it fell by 18 percentage points to 4.3 percent, which OpenAI says is close to GPT-6 Astra and GPT-6.1 Sol.

For anyone who uses a mid range model for research summaries, support answers or data extraction, this is the number that matters most. A model that says it does not know is far easier to build around than one that invents a confident answer, because the uncertain cases can be routed to a human or a bigger model.

Computer use and visual reasoning

OpenAI positions Terra above all as a model for operating software. On Agents' Last Exam, a benchmark of professional tasks such as turning architectural drawings into a 3D model or reviewing an animated film shot by shot to produce a time coded log, Terra at medium effort scores 15 percentage points higher than GPT-5.6 Terra at about a third of the cost. According to OpenAI it matches GPT-6 Astra on that benchmark at a tenth of the price per token.

The comparison with Anthropic is bold. At maximum effort, OpenAI says Terra is more than three times as token efficient as Claude Opus 5.5 and costs a quarter as much per token, which works out to roughly nine times lower cost per task. It also finishes tasks more than three times faster than GPT-6 Astra and GPT-6.1 Sol. These are OpenAI's own measurements, and independent replication of Agents' Last Exam results is still pending, so they are best read as an indication rather than a verdict.

On MMMU Pro, a visual reasoning test, Terra beats its predecessor at every effort level and nearly reaches GPT-6 Astra and GPT-6.1 Sol. On OSWorld 2.1, which measures how well agents complete multistep tasks in desktop applications, Terra at medium effort already exceeds the best result GPT-5.6 Terra achieved at any setting.

The coding jump

The biggest single benchmark gain is in coding. On Terminal Bench 4.0, which tests whether an agent can solve problems from the command line, GPT-6.1 Terra solves 50.3 percent of tasks at maximum effort. GPT-5.6 Terra's published score was 21.5 percent. Artificial Analysis confirms a jump of about 20 points in its own run and says this puts Terra in the same tier as GPT-6.1 Sol on that test.

On the new Coding Agent Index v2.0 from Artificial Analysis, Terra scores 48. That is 6 points below GPT-6.1 Sol at 54 and 17 points above GPT-6 Luna at 31. On Terminal Bench Science 0.1, where agents use code for research tasks such as tracking cell divisions in a time lapse video, OpenAI says Terra beats GPT-6.1 Sol at lower cost and comes within five points of Claude Opus 5.5 at less than a fifth of the estimated cost.

OpenAI is unusually direct about the limits here. For the most complex coding work, the company still recommends GPT-6.1 Sol and GPT-6 Astra. Terra is meant to be the model that handles the large volume of ordinary tasks, not the one that untangles the hardest refactoring.

Safety: rated Critical for cybersecurity

A cheaper model that is better at operating computers and writing code is also more useful for attackers. OpenAI says GPT-6.1 Terra meets the Critical threshold for cybersecurity under its Preparedness Framework, the highest category, and inherits the safeguards built for GPT-6 Astra. The company published an addendum to the GPT-6 Astra system card rather than a full new system card.

According to the system card figures reported by The Decoder, Terra resists direct attacks on its instruction hierarchy about as well as GPT-6 Astra and GPT-6.1 Sol. Against indirect prompt injection, where malicious instructions are hidden in web pages, emails or documents the model reads, its defense rate rose from 96.2 percent with GPT-5.6 Terra to 98.8 percent. OpenAI also says Terra is more likely to stay within the scope of a task and to tell users honestly about limitations and blockers.

That 98.8 percent figure is good, but it is worth putting into perspective. An agent that reads thousands of pages a day will still meet hostile content that gets through now and then. Recent incidents in which coding agents followed instructions planted in files show that prompt injection remains an open problem for every vendor, and cheaper agents mean more agents running unattended.

What the price cut means in practice

To make the prices more concrete, here is a simple example. Imagine a company that processes 10,000 invoice style documents, each needing about 20,000 input tokens and 2,000 output tokens, with no caching. At list prices that job costs about $300 on GPT-6.1 Terra, about $600 on Claude Sonnet 5.5 and about $640 on GPT-5.6 Terra.

Real bills will differ. Reasoning tokens are billed as output, so higher effort settings raise the cost per task, and caching lowers it a lot. But the ratio holds: for the same work at list prices, Terra costs roughly half of what its predecessor and its closest competitor charge. Combined with the higher scores, the cost per useful result drops by more than half.

This continues a pattern that has shaped 2026. OpenAI has cut API prices several times this year, and reports based on payroll and expense data have suggested that cheaper OpenAI models are pulling coding and agent workloads away from Anthropic. A mid range model at half the price of Sonnet 5.5 puts more pressure on that market. It also puts pressure on open weight models, whose main selling point for many businesses is cost.

What to watch

Three things will decide how important GPT-6.1 Terra turns out to be. The first is independent replication of OpenAI's computer use claims, especially the comparison with Claude Opus 5.5 on Agents' Last Exam. The second is how Anthropic and Google respond on price for their own mid range models. The third is real world reliability: benchmark gains in hallucination rates are encouraging, but the true test is how often the model admits uncertainty in messy business documents rather than curated questions.

For now the picture is clear. GPT-6.1 Terra is not the smartest model OpenAI sells, and the company says so itself. It is the one that makes the most sense for high volume work, and at $1 and $5 per million tokens it resets what a capable everyday model is supposed to cost.

Sources

• OpenAI: Introducing GPT-6.1 Terra, October 26, 2026
• OpenAI: Release notes for ChatGPT Work, Codex and the API
• OpenAI Developers: GPT-6.1 Terra model page with pricing and API changelog
• OpenAI Deployment Safety Hub: Addendum to GPT-6 Astra System Card: GPT-6.1 Terra
• Artificial Analysis: GPT-6.1 Terra pushes the cost efficiency frontier and model comparison by effort level
• The Decoder: GPT-6.1 Terra halves OpenAI's API prices and still gains ground on its larger models