OpenAI releases GPT-6.1 Terra: API price halved, focused on computer operation capabilities

Seeing rival Anthropic steal all the spotlight with a single prospectus, OpenAI fired back with GPT-6.1 Terra—

Want to hog the headlines? First ask the new model if it agrees.

GPT-6.1 Terra's pitch is "half the price, stronger capabilities."

Compared with GPT-5.6 Terra, its standard API price is cut in half outright:

$1 per million input tokens, $0.05 per million cached input tokens, and $5 per million output tokens.

And while the price comes down, the capabilities go up?!

In Agents' Last Exam, a practical test of workplace software, Terra not only clearly outperforms GPT-5.6 Terra, but also matches GPT-6 Astra's performance at one-tenth of its token price.

Compared with Claude Opus 5.5, Terra's token price is only a quarter of it, its token efficiency is more than 3 times higher, and its per-task cost is only about one-ninth of the other's.

In terms of use cases, GPT-6.1 Terra's core selling point is also extremely clear—faster, stronger, while drastically lowering the cost of computer control.

It natively supports screen understanding, cross-software collaborative operation, and natural-language writing, aiming squarely at the worker's entire workday.

OpenAI co-founder Greg Brockman even said that Terra's computer-operation capability has crossed a "qualitative threshold" and that using it is faster than operating things himself.

Currently, Plus, Pro, Enterprise, and Edu users can already try it directly in ChatGPT Work and Codex, and developers can call it anytime via the standard API (gpt-6.1-terra).

Even more explosive: on the same day, OpenAI also announced three scientific results produced by its internal AI system, spanning mathematics, computer science, and physics.

What's going on—going crazy before Thursday even arrives??

GPT-6.1 Terra officially launches, price cut in half

As for how good it actually is, let's first look at the official report card OpenAI provided.

First, the overall conclusion:

GPT-6.1 Terra is not OpenAI's most capable model, but it has already reached the capability range of flagship models at a lower cost.

In Agents' Last Exam, GPT-6.1 Terra's medium tier scores 15 percentage points higher than GPT-5.6 Terra, while its per-task cost is only about one-third of the latter's.

This test specifically examines whether AI can use professional software to complete complex work, such as turning architectural drawings into 3D models, or reviewing animation shot by shot and generating repair logs with time codes.

Even more striking, OpenAI says Terra's performance is already on par with GPT-6 Astra, but its per-token price is only one-tenth of the latter's, and it completes tasks more than 3 times faster than both GPT-6 Astra and GPT-6.1 Sol.

Now look at visual capabilities.

In the MMMU-Pro benchmark, GPT-6.1 Terra surpasses GPT-5.6 Terra at all reasoning intensities and approaches GPT-6 Astra and GPT-6.1 Sol at lower cost.

On the computer-operation benchmark OSWorld 2.1, its medium tier also exceeds GPT-5.6 Terra's previous best level.

In other words, the new Terra is not just better at "seeing," but also better at actually getting work done based on screen content.

Coding capability has likewise seen a big jump.

In Terminal-Bench 4.0, GPT-6.1 Terra's max tier solved 50.3% of tasks, while GPT-5.6 Terra's previously published score was only 21.5%.

More than double.

However, OpenAI also drew a line on its own initiative:

For the most complex programming, scientific research, and deep reasoning tasks, GPT-6.1 Sol and GPT-6 Astra remain the more suitable choices.

Another obvious upgrade is saying less nonsense.

In OpenAI's internal high-difficulty factual stress test, the error rate at low reasoning intensity dropped from 53.6% for GPT-5.6 Terra to 16.7%.

The max tier further drops to 4.3%, already close to GPT-6 Astra and GPT-6.1 Sol.

At least judging from this round of official results, Terra's positioning is already clear:

Use near-flagship capability to drive down the cost of a large number of computer tasks.

In the comprehensive intelligence index of third-party organization Artificial Analysis, GPT-6.1 Terra's max tier scores 48, ranking behind Claude Opus 5.5, Kimi K3.1 Pro, Google Gemini 4 Argon, Meta Muse Wonder, and GPT-6.1 Sol, but ahead of Grok 4.7 and Xiaomi MiMo-V2.6-Pro.

DesignArena gave GPT-6.1 Terra and GPT-5.6 Terra the same prompt:

Recreating Lisbon's Belém Tower at golden hour.

Not only did it have to draw this 16th-century limestone fortress, it also had to reproduce the carved stone ropes, open balconies, battlements, and domed corner turrets, and the image also had to include seagulls, sailboats, and the 25 de Abril Bridge in the mist.

The results are interesting.

GPT-6.1 Terra took nearly 4 times as long as the previous generation and used 48% more tokens.

But because the unit price is less than half that of the previous generation, the final total cost was actually 38% lower (the new Terra cost $0.047, while the old Terra cost $0.077).

Fortunately, the extra time was indeed reflected in the image.

Which one is more refined and more complete is clear at a glance.

Another developer, Victor Nunez, brought out the "long form."

He had both GPT-6.1 Terra and GPT-6.1 Sol find and fill out a relatively long general-purpose form in Fast mode, then compared the two side by side in terms of the computer-operation process.

According to Nunez, in the first test, Terra completed the form 53.5 seconds earlier than Sol.

The OpenAI developer account then directly reposted this set of demonstrations.

This also hits exactly on Terra's official positioning:

It is not necessarily responsible for thinking the hardest problems all the way through, but it is very suitable for recognizing information in interfaces, clicking continuously, filling in content, and quickly running through a long series of steps.

Another netizen, Leo, put Terra and Claude side by side.

Judging by the composite intelligence score, Terra got 48 points, lower than Sonnet 5.5's 56 points and Opus 5.5's 58 points.

But when it came to speed and cost, the results flipped directly:

Terra's generation speed reached 283 tokens/s, 3.4 times that of Opus 5.5, and the cost per task was only $0.4, about one-fifteenth that of Opus 5.5.

So Leo's summary was also very direct:

If you want intelligence and capability, choose Sonnet 5.5; if you want speed and cost-effectiveness, choose Terra. As for Opus 5.5, Terra is not yet in the same league.

And another netizen's actual test results also confirmed Leo's conclusion.

Comparing Sonnet 5.5 and 6.1 Terra with the same prompt, there is no need to say more about which one comes out ahead.

Compared with Opus 5.5, it is indeed much cheaper, but considering the generated results, one can only say:

You get what you pay for (doge).

On the same day, OpenAI also released 3 new AI scientific discoveries

Besides Terra, OpenAI also dropped a bigger "academic bombshell" on the same day.

And it was not just a piece of natural-language derivation and that was the end of it.

The natural-language drafts and Lean formal proofs of the mathematics and computer science results have already been made public.

Among the physics results, the Lean formalization of the existence of pure Yang-Mills theory has also been completed, but the related materials will first be handed to an external expert panel for review.

I will not go into the details here; I have briefly summarized the progress OpenAI made this time:

1. Cracking a conjecture that has troubled the mathematics community for more than 30 years

OpenAI claims that its AI system proved the local smoothing conjecture in all dimensions, and thereby further solved the Kakeya conjecture and the Stein restriction conjecture.

Simply put, this research discusses how waves focus during propagation.

It was proposed by mathematician Christopher Sogge in 1991, and previously only some dimensions had been solved.

Now, AI has not only provided a natural-language proof, but also written it into the Lean formal proof system.

2. A foundation of computer science theory that has been used for decades may be wrong

The second result points to the Exponential Time Hypothesis (ETH).

This hypothesis roughly holds that for complex problems like 3-SAT, it is difficult to find algorithms much faster than brute force.

Over the past decades, many computer science conclusions have been built on the assumption that it holds.

But OpenAI claims that its AI system proved that ETH is wrong, and has completed the Lean formalization.

However, this does not mean that the P≠NP problem has been solved, nor does it mean that practical new algorithms will immediately appear; it only shows that some complex problems may not be as hard to compute as previously imagined.

3. Filling in a mathematical proof for four-dimensional quantum field theory in physics

The third result went directly into quantum field theory.

OpenAI claims that its system provided existence proofs for certain four-dimensional quantum field theories, including quantum chromodynamics (QCD).

Simply put, it attempts to prove from a rigorous mathematical perspective that:

These physical theories used to describe quarks, gluons, and the strong interaction can indeed hold.

This also includes a proof of the existence of pure Yang-Mills theory—

This is part of the famous "Yang-Mills existence and mass gap" Millennium Prize Problem.

But note, only a part.

The mass gap problem remains unsolved, so it cannot yet be said that AI has conquered this $1 million world problem.

At present, the related results are also being reviewed by external experts such as Nobel laureate David Gross.