Watching its rival Anthropic steal all the spotlight with an IPO prospectus, OpenAI countered by throwing out GPT-6.1 Terra —
Want the headlines all to yourself? First ask whether the new model agrees.
GPT-6.1 Terra's main pitch:"Half the price, stronger capability"。
Compared with GPT-5.6 Terra, its standard API price is cut in half:
1 dollar per million tokens for input, 0.05 dollars per million tokens for cached input, and 5 dollars per million tokens for output.
And while the price went down, the capability actually got stronger?!
In the workplace software practical test Agents’ Last Exam, Terra not only clearly surpassed GPT-5.6 Terra,it also matched 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, its token efficiency is over 3 times higher, and the cost per task is only about one-ninth of its rival's.
Seriously, has AI cost-cutting and efficiency-boosting already gotten this competitive?
In terms of use cases, GPT-6.1 Terra's core selling point is extremely clear —faster, stronger, while greatly lowering the cost of computer control。
It natively supports screen understanding, cross-software coordinated operation, and natural language writing, firmly targeting office workers' full workday.
OpenAI co-founder Greg Brockman even stated that Terra's computer operation capability has crossed the "qualitative threshold", running faster than he could do it 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 systems, spanning mathematics, computer science, and physics.
On one hand driving down the cost of AI operating computers, on the other raising the imagination of AI participating in frontier scientific research.
What, going crazy before Thursday even arrives??
GPT-6.1 Terra officially launches, price cut in half
As for the specific capabilities, let's first look at the official scorecard released by OpenAI.
First, the overall conclusion:
GPT-6.1 Terra is not OpenAI's most capable model, but at a lower cost it has already reached the capability range of flagship models.
In Agents’ Last Exam, GPT-6.1 Terra at the medium tier scored 15 percentage points higher than GPT-5.6 Terra, while its cost per task was 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 timecoded fix logs.
Even more striking, the company says Terra's performance is on par with GPT-6 Astra, but its per-token price is only one-tenth, and it completes tasks more than 3 times faster than both GPT-6 Astra and GPT-6.1 Sol.
Now let's look at vision capabilities.
In the MMMU-Pro benchmark, GPT-6.1 Terra surpasses GPT-5.6 Terra at all reasoning effort levels, and approaches GPT-6 Astra and GPT-6.1 Sol at a lower cost.
On the computer-use benchmark OSWorld 2.1, its medium-effort score also exceeds the previous best of GPT-5.6 Terra.
In other words, the new Terra is not only better at "seeing", but also better at actually getting tasks done based on screen content.
Coding ability also saw a significant leap.
In Terminal-Bench 4.0, GPT-6.1 Terra's max-effort setting solved 50.3% of tasks, while GPT-5.6 Terra's previously published score was only 21.5%.
More than doubled outright.
However, OpenAI also drew a line itself:
For the most complex coding, scientific research, and deep reasoning tasks, GPT-6.1 Sol and GPT-6 Astra remain the more suitable choices.
Another clear upgrade is talking less nonsense.
In OpenAI's internal high-difficulty factuality stress tests, the error rate at low reasoning effort dropped from 53.6% for GPT-5.6 Terra to 16.7%.
At the Max tier it fell further to 4.3%, already approaching 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 volume of computer tasks.
In the composite intelligence index from third-party organization Artificial Analysis, GPT-6.1 Terra's Max tier scored 48 points, 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.
Overall it sits in the middle of the pack among the world's top-tier models.
Netizen hands-on tests: finer rendering, with costs cut by 38%
So what about the model's real-world capabilities?
As usual, let's look at hands-on results shared by netizens.
DesignArena gave GPT-6.1 Terra and GPT-5.6 Terra the same task:
Recreate the Belém Tower in Lisbon at golden hour.
It required not only drawing this 16th-century limestone fortress, but also restoring the carved stone ropes, open balconies, battlements, and domed turrets, and the scene also had to include seagulls, sailboats, and the April 25 Bridge in the mist.
The results were quite 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 was less than half that of the previous generation, the final total cost was actually 38% lower(the new Terra spent $0.047, while the old Terra spent $0.077).
Fortunately, the extra time spent was indeed reflected in the image.
Which one is more refined and more complete is plain to see at a glance.
Another developer, Victor Nunez, brought out a "long form."
He had both GPT-6.1 Terra and GPT-6.1 Sol find and fill out a fairly long general form in Fast mode, then compared their computer operation processes side by side.
According to Nunez,in the first test, Terra completed the form 53.5 seconds ahead of Sol.
OpenAI's developer account subsequently reposted this demonstration directly.
This fits Terra's official positioning perfectly:
it may not be the one to think the hardest problems through to the end, but it is well suited to recognizing information in interfaces, clicking repeatedly, filling in content, and rapidly running through a long list of steps.
Another netizen, Leo, put Terra and Claude side by side.
In terms of composite intelligence scores, Terra scored 48 points, below Sonnet 5.5's 56 points and Opus 5.5's 58 points.
But when it came to speed and cost, the result was completely reversed:
Terra generates at a speed of 283 Token/s, which is 3.4 times that of Opus 5.5, with a per-task cost of only $0.4, roughly one fifteenth of Opus 5.5.
So Leo's summary is also very direct:
For intelligence and capability, choose Sonnet 5.5; for speed and value for money, choose Terra. As for Opus 5.5, Terra is not yet in the same league for now.
Moreover, another netizen's hands-on test results also confirmed Leo's conclusion.
Using the same prompt to compare Sonnet 5.5 and 6.1 Terra, no words are needed on which one comes out on top.
Compared with Opus 5.5, the price is indeed much cheaper, but considering the generated results, one can only say:
Expensive things are expensive for a reason (doge).
On the same day, OpenAI also dropped 3 new AI scientific discoveries
Besides Terra, on the same day OpenAI also dropped an even bigger "academic bombshell".
The company officially announced thatits AI system obtained three new results in mathematics, computer science and physics, concerning harmonic analysis, complexity theory and quantum field theory respectively。
And it wasn't just a matter of handing over a passage of natural-language derivation and calling it done.
The natural-language manuscripts and Lean formalized proofs for the mathematics and computer science results have been made public.
For the physics result, the Lean formalization of the existence of pure Yang–Mills theory has also been completed, but the relevant materials will first be reviewed by an external panel of experts.
I won't go into the specifics here, but I have briefly summarized OpenAI's advances this time:
1. Conquering a conjecture that has plagued the mathematics community for more than 30 years
OpenAI stated thatits 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 as they propagate.
It was proposed by mathematician Christopher Sogge in 1991, and previously only some dimensions had been resolved.
Now, the AI not only provided a natural-language proof but also wrote it into the Lean formalized proof system.
If it is subsequently confirmed by the mathematics community, it will mean that AI has genuinely entered a problem that humans have been unable to fully conquer for decades.
2. A foundation of computer science theory 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 very hard to find algorithms much faster than brute force.
Over the past decades, many conclusions in computer science have been built on the assumption that it holds.
But OpenAI stated thatits AI system proved that ETH is false, and the Lean formalization has already been completed。
However, this does not mean the P≠NP problem has been solved, nor does it mean 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 the mathematical proof for four-dimensional quantum field theory in physics
The third result goes straight into quantum field theory.
OpenAI stated thatits system provided existence proofs for certain four-dimensional quantum field theories, including quantum chromodynamics (QCD)。
Simply put, it is an attempt to prove, from a rigorous mathematical perspective, that:
These physical theories used to describe quarks, gluons, and the strong interaction do indeed hold up.
This also includes a proof of the existence of pure Yang-Mills theory—
which is part of the famous 'Yang-Mills Existence and Mass Gap' Millennium Prize Problem.
But note: only a part of it.
The mass gap problem remains unsolved, so it cannot yet be said that AI has conquered this $1 million world-class challenge.
Currently, the relevant results are also being reviewed by external experts such as Nobel laureate David Gross.
Releasing a new model on one hand, while casually solving a problem that has troubled humanity for decades on the other.
OpenAI:
"Can't IPO before Company A? I have plenty of ways to grab attention."
Reference links:
[1]
https://openai.com/index/introducing-gpt-6-1-terra/
[2]
https://developers.openai.com/api/docs/models/gpt-6.1-terra
[3]
https://developers.openai.com/api/docs/pricing
[4]
https://x.com/OpenAI/status/2114780711664398349
[5]
https://x.com/DesignArena/status/2114800086391878104