OpenAI Halves Terra's Price, Undercuts Claude Opus Ninefold

By Daniel Rao, TechPulse USA — October 27, 2026

OpenAI has finally done the unglamorous thing that actually changes who can afford AI agents: it made the workhorse model cheap. GPT-6.1 Terra, released on October 26, 2026, does not top any leaderboard that matters, and OpenAI does not pretend otherwise. It costs half what its predecessor did, finishes computer-use tasks more than three times faster on OpenAI's own timing runs, and — on the company's published chart — completes a top-end agent task for about a ninth of what Anthropic's Claude Opus 5.5 charges. That last number, not any benchmark crown, is the story.

For a year, the model race has been scored like a horse race: who is smartest, who writes the best code, who hallucinates least. Terra reframes it as a procurement question. If an agent is going to click through a thousand invoices, fill a thousand forms, or triage a thousand support tickets, the question is not whether it can do one of them brilliantly. It is what a thousand of them cost, how long they take, and how often a human has to clean up afterwards. On all three, OpenAI's pitch is that Terra is now good enough — and that good enough, at this price, is the product.

9to5Mac, which has tracked every rung of OpenAI's model ladder, opened its report on the launch with a grin: "Earth is so back." The joke lands because Terra was the odd one out last month, when OpenAI moved Sol and Luna onto the GPT-6 family and left its middle tier running a generation behind. This release closes that gap, and it does so with a price cut rather than a capability boast.

File photo: hands on a keyboard and mouse at an office desk — the kind of computer work GPT-6.1 Terra is built to automate. Image: Pixel.la Free Stock Photos, via Wikimedia Commons, CC0.

What shipped, and where you can use it

GPT-6.1 Terra is the successor to GPT-5.6 Terra, and OpenAI describes it, in its own announcement, as "a fast, efficient model that combines detailed visual reasoning and precise computer use alongside clearer, more natural writing." In plain terms: it looks at what is on screen, takes action across apps, and writes like a colleague rather than a autocomplete demo. It is available now in ChatGPT Work and Codex for Plus, Pro, Enterprise and Edu users, and to developers through the API as gpt-6.1-terra.

The rollout did not stop at OpenAI's own products. Amazon Web Services says the model is generally available on Amazon Bedrock, where teams can call it through Bedrock's APIs or point ChatGPT Work and Codex at it, and GitHub has brought it to Copilot across its paid tiers, including Copilot Pro, with a gradual rollout and administrator controls for business customers. That breadth matters: a model that only lives inside one company's chat app is a demo, but a model that lands the same day in the API, in Bedrock and in Copilot is infrastructure, ready to be wired into real workflows.

We checked the launch announcement in OpenAI's developer community ourselves before writing this — the "50% lower" pricing claim and the gpt-6.1-terra model identifier are both there, in OpenAI's own words, not just in second-hand coverage. The documented context window is 1,050,000 tokens, with maximum output of 128,000 tokens and a knowledge cutoff of June 1, 2026, according to OpenAI's developer documentation as reported by blockchain.news.

Do the maths: a dollar a task changes who can automate

Start with the price list, because OpenAI wants you to. Standard API pricing is $1 per million input tokens, $0.05 per million cached input tokens, and $5 per million output tokens. Against GPT-6.1 Sol, that is exactly half on every line: Sol charges $2 and $10. Against GPT-5.6 Terra, the input price halves from $2 to $1, the cached input price falls from $0.20 to $0.05, and the output price drops from $12 to $5 — further than half, in fact, though OpenAI headlines the overall cut as 50% lower.

But token prices are the wrong number to stare at, and the launch proves it. What a finished task costs depends on how many tokens a model burns getting there, and on which effort setting it runs at. OpenAI's own chart for Agents' Last Exam, a benchmark of professional computer tasks, puts Terra at medium reasoning effort at 57.6%, against 42.6% for GPT-5.6 Terra — 15 points higher — at $0.29 per task instead of $0.84. That is roughly a third of the cost for a materially better result, on the vendor's own figures.

At maximum effort, the comparison OpenAI is clearly proudest of is the one against Anthropic. Its published chart shows Terra scoring 60.0% on Agents' Last Exam at $1.01 per task, against Claude Opus 5.5 at 63.2% for $8.82. Read that honestly: Opus still scores higher, by 3.2 points. It just costs about 8.7 times as much per task. Run the arithmetic out to 10,000 such tasks and the gap becomes a budget line: about $10,100 for Terra against about $88,200 for Opus, on OpenAI's chart figures. A finance director does not need a benchmark to understand that sentence.

The independent numbers tell a similar story at a different scale. Artificial Analysis, the independent benchmarking firm, measured Terra at an Intelligence Index of 48.1 at maximum effort for $0.40 per task on average across its own evaluation tasks, against $0.72 for GPT-6.1 Sol and $5.98 for Claude Opus 5.5. Its summary, as reported by Thorsten Meyer AI, is that Terra gains 6 points over GPT-5.6 Terra "at less than a third of the cost per task ($0.40 vs $1.40)." Artificial Analysis also posted on X that Terra is "the first Terra-class model since GPT-5.6, and is more intelligent for less than a third of the cost." Two different benchmarks, two different cost bases — but the direction is the same, and that is rarer than it should be in this industry.

The benchmark story, honestly told

Every number in the previous section needs a label, because vendor benchmarks are marketing with error bars. Here is the honest version. On Terminal-Bench 4.0, a coding-agent test built around command-line problems, OpenAI reports that Terra solves 50.3% of tasks at maximum reasoning effort, more than double GPT-5.6 Terra's published 21.5%. Artificial Analysis, measuring independently, puts the gain at roughly 20 points and Terra close to Sol's level — a different absolute figure, because the harness and settings differ, but the same story of a large jump. For the hardest coding work, OpenAI itself still points customers to Sol and Astra, and that caveat should be believed: it costs OpenAI nothing to say it, and it saves buyers an expensive mistake.

On factual reliability, OpenAI's internal test of prompts designed to induce errors shows Terra's error rate falling from 53.6% for GPT-5.6 Terra to 16.7% at low reasoning effort, and to 4.3% at maximum effort — nearly matching Astra and Sol, on the company's own test. Artificial Analysis measures something related but different, and the distinction matters: on its AA-Omniscience test, which counts how often a model answers anyway when it does not know, Terra's guessing rate fell from 88% to 38%. That is a colossal improvement and still a warning label. A model that bluffs 38% of the time when it is out of its depth cannot be left unsupervised on anything where an error is expensive, and OpenAI's own safety classification underlines the point: the company says Terra meets the Critical threshold for cybersecurity under its Preparedness Framework, and retains the safeguards developed for GPT-6 Astra. Critical is OpenAI's own label, under its own framework — useful as a signal of capability, not as an independent safety verdict.

Then there is speed. OpenAI says Terra completes tasks more than three times faster than GPT-6 Astra and GPT-6.1 Sol, and that it matches Astra on Agents' Last Exam at one-tenth the price per token. Both claims rest on OpenAI's own timing runs and chart data. Artificial Analysis had not published independent speed measurements for Terra when Thorsten Meyer AI checked on launch day, so the "three times faster" claim is, for now, unverified by anyone except the seller. Treat it as a promising claim, not a fact, until an independent lab times it.

One independent check cuts the other way, and it is worth hearing. Thorsten Meyer, author at Thorsten Meyer AI, ran Terra at all six effort settings through his own newsroom benchmark on launch day and found the dial behaving oddly: "On a task that rewards discipline, more thinking made Terra worse." At maximum effort, his test found Terra pulling in outside sources and adding facts its source material did not contain. His conclusion generalises well beyond his newsroom: "the setting that wins a general benchmark does not automatically win your task." That is the single most useful sentence in the launch coverage, because it tells buyers to test Terra on their own work at low and medium effort first — where Artificial Analysis finds it on the cost-efficiency frontier — before paying for maximum effort at all.

What it is actually for: the unglamorous workday

Strip away the benchmarks and Terra's pitch is deliberately boring, which is why it might work. OpenAI's own example of high-volume work is reading incoming invoices and entering them into accounting software. Its broader list runs through drafting proposals, interpreting technical drawings, filling out forms, and editing video. Interesting Engineering's coverage adds the concrete benchmark tasks behind the jargon: converting architectural drawings into 3D models, and documenting individual film shots for restoration teams. None of this is the stuff of keynote gasps. All of it is the stuff that eats an afternoon in every mid-sized company in America.

That is also why the form-factor matters as much as the model. Computer use is not a chatbot feature; it is a model that can be trusted, within limits, to operate software the way a junior employee would — reading a screen, clicking the right field, moving between apps without losing the thread. On OSWorld 2.1, a desktop-agent benchmark of multistep application workflows, OpenAI reports that Terra beat GPT-5.6 Terra's best score at medium effort while costing substantially less per task. And on MMMU-Pro, a visual-reasoning test, it outperformed its predecessor at every reasoning setting and came close to Astra and Sol. Again, these are OpenAI's figures, but the shape of the claim is consistent: a middle-tier model that can see, read and act, rather than merely answer.

For US buyers, the practical consequence is mundane and large. A small firm that could never justify running its invoice backlog through a flagship model at flagship prices can now plausibly automate it, check a sample, and keep a human in the loop for the exceptions. A large firm running tens of thousands of agent tasks a month can reprice its whole automation budget. The risk has not gone away — the 38% guessing rate and the Critical cybersecurity classification are reminders of that — but the price of trying has collapsed, and in enterprise software, the price of trying is often the whole ballgame.

Where Terra sits in the GPT-6 lineup

OpenAI's current ladder, from cheapest to most capable, runs Luna, Terra, Sol and Astra. Luna is the small, fast tier; Sol is the serious coding and professional-work model; Astra is the flagship. Terra is the middle rung that most workplaces will actually live on, and its absence from last month's GPT-6 update is what made this release feel overdue rather than surprising. When OpenAI shelved Astra and shipped Sol instead at its September DevDay, Terra was left behind — a gap this launch finally fills, giving the middle tier a GPT-6.1 model of its own.

The pricing ladder now tells its own story. Terra at $1/$5 sits at half of Sol's $2/$10 and a tenth of Astra's $10/$50 token prices, according to Artificial Analysis's listings. But the per-task ladder is flatter and more interesting: Terra at maximum effort costs $0.40 per task on Artificial Analysis's average, against $0.72 for Sol and $3.26 for Astra. The flagship is still better — Astra scores 52.7 on the same index against Terra's 48.1 — but the buyer is paying eight times as much per task for 4.6 more index points. For most office work, that is a botched trade, and OpenAI knows it: its own guidance is that in-depth research and complex coding still belong with Sol or Astra, while everything else in the workday belongs with Terra.

There is a wider context here that regular TechPulse USA readers will recognise. OpenAI has spent the autumn pushing its models into professional workflows from every direction, including Astra for Law, its bid for legal work, and the comparison that matters most for Terra's pricing story is the one against Claude Opus 5.5 — the model that Kimi K3.1 Pro nearly matched in independent testing just days earlier. The frontier is crowded, and every vendor in it is now cutting prices to hold its place. Terra is OpenAI's sharpest cut yet in the tier where most customers actually buy.

Our take: price per task is now the real model war

Here is the stance, stated plainly: the benchmark crown no longer decides this market, and anyone still ranking models by a single score is measuring the wrong thing. What decides it now is cost per finished task, speed per task, and error rate per task — the three numbers that determine whether an agent is a toy, a tool, or a liability. Terra does not win all three outright. It wins the combination at a price nobody else in its class is matching, and it does so while being honest, for once, about where it loses: Opus scores higher, Sol and Astra code better, and maximum effort can make a disciplined task worse rather than better.

That honesty is the most encouraging part of this launch, and also its biggest risk. A model priced to be run thousands of times a day will be run thousands of times a day, including by people who skip the human review step that the 38% guessing rate demands. The Critical cybersecurity rating is not decoration: a cheap, capable agent that can operate a computer is also a cheap, capable agent that can be pointed at the wrong computer. OpenAI's safeguards, prompt-injection resistance and Astra-derived protections are real mitigations, but they are OpenAI's mitigations, described by OpenAI, and the history of this product category says the first serious misuse will teach us more than any launch chart.

Still, the direction of travel is unmistakable, and it favours buyers. A year ago, running an agent across a company's daily paperwork at flagship prices was a board-level decision. Today, at roughly $0.29 a task on OpenAI's computer-use figures — or $0.055 a task at Terra's low effort setting on Artificial Analysis's average — it is a line item a department head can approve. That is how technologies actually spread: not when they become brilliant, but when they become cheap enough to try, boring enough to trust, and fast enough not to annoy. Terra is not brilliant. It is cheap, fast and good enough, and in the workday market, that combination has a habit of winning.

What to watch next is simple. First, whether independent labs confirm the speed claim — if Terra is not actually three times faster outside OpenAI's lab, the value case weakens fast. Second, whether Copilot and Bedrock customers report the same per-task costs in production, or whether real invoices and real drawings are messier than benchmark tasks. And third, how Anthropic answers: a model that costs nearly nine times as much per task cannot hold its price while scoring only 3.2 points higher, not for long, and not in a market where the buyer is now a finance director rather than a researcher. The next move in this war will be a price cut, and Terra just fired the starting gun.

GPT-6.1 Terra FAQ: what the independent tests add

What did independent testing find that OpenAI did not say?

Artificial Analysis, testing Terra on 26 October, found it Pareto-optimal for intelligence against cost at low and medium effort — in plain terms, no other model it tested gave you more intelligence for the money at those settings. At maximum effort, Terra gained 6 Intelligence Index points over GPT-5.6 Terra at less than a third of the cost per task: $0.40 against $1.40. Cached input, the cheap repeat-token rate, fell to $0.05 per million tokens from $0.20. Those are Artificial Analysis's figures, not OpenAI's, and they are the reason the price story is bigger than a headline rate cut.

Does Terra still guess when it does not know?

Yes, but far less. On Artificial Analysis's AA-Omniscience test, Terra guessed 38% of the time when it did not know an answer, down from 88% for GPT-5.6 Terra, with a 2-point accuracy gain alongside. That is a concrete quality gain, not just a cheaper way to be wrong — and it matters most for the unglamorous workday tasks Terra is pitched at, where a confident wrong answer costs more than a slow one.

Is price per token the same as price per task?

No, and the same independent testing proves it. Luna at maximum effort costs $0.068 per task against Terra at low effort's $0.055, despite Luna's token price being roughly a tenth of Terra's — because Luna writes far more tokens to finish each task. The "ninth of Opus" claim in this post is a per-task figure at stated effort settings, and effort settings are doing real work inside it. Anyone budgeting on token prices alone will misprice the job.

Anthropic's own IPO filing shows why price pressure cuts both ways: the lab selling the expensive model is also the one reporting staggering costs of its own. And the insiders warning that the AI buildout is outrunning returns are worth reading before assuming cheaper models end the spending war — they may just move it.

Sources: OpenAI, "GPT-6.1 Terra: delivering fast and capable computer use at a lower price" (developer community announcement, October 26, 2026); 9to5Mac, "OpenAI upgrades ChatGPT and Codex with new GPT-6.1 Terra model" (October 26, 2026); Interesting Engineering, "OpenAI's GPT-6.1 Terra matches GPT-6 Astra at 10% of the token cost" (October 26, 2026); Seeking Alpha, "OpenAI unveils new AI model, GPT-6.1 Terra" (October 26, 2026); AI Affairs, "OpenAI releases GPT-6.1 Terra for computer use at lower API prices" (October 26, 2026); blockchain.news, "OpenAI launches GPT-6.1 Terra for ChatGPT Work and Codex" (October 26, 2026); Artificial Analysis launch coverage and model pages, as reported by Thorsten Meyer AI, "GPT-6.1 Terra: Cheap per Task, Worst at Max in Our Test" (October 26, 2026); Amazon Web Services Bedrock announcement, as reported by Oossa (October 26, 2026); GitHub Copilot availability, as reported by Superpower Daily (October 26, 2026). Benchmark figures are OpenAI's own unless identified as Artificial Analysis measurements.

By Daniel Rao — Daniel Rao covers AI models and the business of compute for TechPulse USA — the benchmarks, the bills, and the bots doing the office work.