GPT-6 Sol and Luna Shock released!

It's the first time in the history of the world.

It was a night of intense fighting.
GPT-6 Sol and Luna have just been released!
A few hours ago, the front leg of Opus 5.5 was just starting.

This time, the entire GPT-6 barrel comes to the same end as the GPT-6 Sol and the GPT-6 Luna.

If Astra is at the peak of AI, Sol and Luna are OpenAI's handmade industrial amplifiers for hundreds of industries.
They perfectly inherit Astra's cutting-edge capabilities in professional reasoning, coding and alignment with humans, representing two-way practical scenarios.

Sol is a "super-brain" that is used to decipher extremely complex code logic and the workflow of super-long-range intelligences.

Luna is a "freak of efficiency", fast, low-resource, and focused on all types of, high-frequency, daily tasks.

Not only that, but they are 50% faster than the previous generation in API pricing! Since then, the consumption of millions of tokens is no longer surprising.

Ottman said that the GPT-6Sol and Luna models, both measured by mission pricing, have no competitors in the market.
As OpenAI puts it: "Build with Sol, expand with Luna, and move toward a productive environment and a broader future".

Standing on the shoulders of the giant Astra
This time, Sol and Luna were built using training methods similar to the GPT-6 Astra, drawing on Astra's breakthroughs in professional work, fact accuracy, coding, computer vision operations, and human alignment.
GPT-6 Sol: Complex workflows「Superbrains」
Sol's positioning is an extremely difficult task that requires deep thinking and logical reasoning.
It is specifically designed for complex code reconstruction, long logical chain business decisions, and multi-step planning of Agent workflows.
In short, it applies to that kind of "hard bone".
GPT-6 Luna: Focused, Godspeed「Efficiency Madness」
Luna was created for high-throughput, highly focused missions.
For example, extracting specific invoice data from hundreds of pages of PDFs instantly, or translating a large number of user chat records in real time.
Luna is extremely fast, low-cost and almost negligible, and is built specifically for enterprise-grade bulk processing tasks.
The price of the API is straightforward.Although the previous GPT-6 Astra was powerful, much of our daily work did not require this expensive model.
The birth of the GPT-6 Sol and Luna meant that Astra's cutting edge intelligence could be used daily by ordinary people through extreme cost controls.
Thanks to significant improvements in caching and reasoning underlying architecture, OpenAI has succeeded in lowering costs!
Compared to the GPT-5.6's promotional price, the GPT-6 Sol and Luna's overall API prices are down 50%!

The price per million tokens is as follows:

GPT-6 Sol (complex reasoning for signals)

Import prices: from $4 to $2!

Export prices: from $20 to $10!

GPT-6 Luna (spectral high frequency sputtering)

Import prices: from $0.20 to $0.10!

Export prices: from $1.20 to $0.50!

Now, with the price of white tea, you can buy top "cyberworkers", with higher quantities and lower costs!

Tiboti emphasized that this is a permanent price reduction, and the original limit time price is once again at a discount.

The runner put Claude on the ground and rubbed him.
In the official blog post, almost every run is contrasted with Claude, and the lateral axis of each chart is the cost of each task.
OpenAI emphasizes that "the same amount is spent".

Artificial Analysis Review: GPT-6 Sol and Luna are pushing the cost-effectiveness frontier to new heights

First, it's a real work environment.
It's not.AutomationBenchIn (commercial workflow across application testing), GPT-6 Sol (xhigh level) gets 33.2%, costing $0.27 per task, directly above Fable 5.1 (max level), and the reporting cost per task is down a staggering 88%!
Claude Opus 5's highest score was 26.9%, costing 11.1 times that of Sol.

OpenAI specifically pointed out that Fable 5.1 took 31.4%, but about 40% of the tasks were transferred to the bottom of Opus 5, and this part of the cost was not calculated, so it was underestimated.

It's not.Agents' Last ExamThe (Smart Body Final Exam) examines long-chain professional assignments in 55 segments of industry.
Sol achieved 56.4% of the best performance at the highest intensity, higher than Opus 5, with a 60% lower cost per mission.

It's not.Accuracy of factsUp there.The fact of the matter is that Sol's mistakes are about half the size of the previous generation, and close to Astra.
Luna is equivalent to the GPT-5.6 Sol in high-intensity ratings, costing about one percent of the latter.

Then the focus is on the programming scene.
FrontierCodeThe GPT-6 Sol made a huge leap forward from its predecessor. It matched the highest grades of Fable 5.1, but cost much less.
In addition to writing code, it also includes code style, test quality, and compliance with existing code libraries.

It's not.DeepSWEThe first is a complex software engineering task in a real code library.
Sol's max score is 68.8%! Fable 5's highest score on this one is 69.9%, and while the two are 1.1 percentage points behind, Sol's cost of completing each task is 80% lower than Claude's!
Luna scored 66.6% below the max level, directly on par with Opus 5 and Fable 5 (medium intensity), but the cost dropped 93% and 96%, respectively.

Next up is the computer operation.
In the OSWorld 2.0 offline test, the GPT-6 Sol was matched by Claude Opus 5 with only 20% of the cost.
At a cost of one-tenth, the Luna outperformed its predecessor, the GPT-5.6 Sol!

It's not.Human alignmentOn top of that, OpenAI specializes in testing whether models will lie when programmed.
GPT-5.6 Sol's fraud rate was 10.4%, while GPT-6 Sol's fraud rate dropped to 1.3%.

In addition, they also measured that when the search tool broke down, the model would not stick to the user's hard drive.
GPT-5.6 Sol was down to 4.9%, and Luna dropped from 78.3% to 28.7%.

The underlying details are greatly optimised: goodbye to AI, cache input tokens and take a break.
OpenAI has revealed that this time they have done a lot of optimization on the underlying details.
First, it's the two models that say goodbye to the taste and smell of AI.
GPT-6 Sol and Luna are very clear, no nonsense; they reject harsh professional terms and strange phrases; the answer is shorter, but full of dry goods.
For example, when asked about the structure of the website, GPT-6 Sol does not blink, but instead goes straight to the subject, using words that are precise and unambiguous.

The reason for faster generation and fewer tokens is because of the enhanced Prompt caching technology of the GPT-6.
In the past, AI has re-read the context every time, while in GPT-6, the caching rate has increased significantly.
What's even crazier is that the life-saving cache enters the Token and goes straight to 1!
Now, developers have super-high control.

You can adjust the "thinking intensity" of the model at any time, encountering difficulty problems, encountering simple problems, and this process will not interrupt caching!

It can be switched on or off at any time without interrupting the cache!

You can even manually set "breakdowns" to accurately control the location of the cache.

According to GitHub, of the billions of API requests monitored, the number of token requests that need to be reprocessed has dropped by more than 50 percent.

The first round of testing has arrived.The GPT-6 is hard Opus 5.5Since Opus 5.5 was hit on the same day, how could "Fight of the Century" have been played by a fanatic who was not a fan of Opus 5.5.
The first wave of testing is fresh!
Bridgebench was successful, inviting four new models to create a rocket launch pad together.
The GPT 6 Lun was completed in just 65 seconds at a cost of $0.01. The cost and timing of the other three models are as follows:
GPT 6 Sol: $0.11, 1 minute; Grok 4.7: $0.29, 11 minutes Claude Opus 5.5:1.52 dollars, 12 minutes

Another round of GPT 6 Sol and Opus 5.5 matches were played on the SVG battlefield.
Angel, the developer, said, "A word is better than a thousand words".

In the test below, it is clear that there are two models that differ significantly in the understanding of the scene and the details.

Another developer, Azazide, created a tougher shutdown and threw Opus 5.5 and GPT-6 Sol to Mars.
The rules are extremely cruel, 100 colonists live and die in 90 seconds, with zero human intervention.
They took over a colony on Mars on the brink of collapse, a leak of lithium-oxygen pipes, batteries set on fire, an extremely strong sandstorm, a complete power outage, and only three repair RVs in the district could move.
Both sides of the AI have to solve the "train puzzle" in milliseconds.
Unsurprisingly, they managed to save the lives of 100 colonists through their own strategies.

A trailer for a playful "Selda".
Flavio Adamo, the developer of Sol, said: "The model is extremely efficient, the speed and the use of tokens are amazing".
Introduced in The Legend of Zelda: Breath of the Wild: After playing a trailer, Sol directly handles a playable version of the game.

Netizens Chetan Ankola created an anime-style web page with GPT-6 unlocked.

GPT-6 unlocked the entire platformThe good news is that both models are available today!

ChatGPT platform: GPT-6 Sol and Luna will be rolled out in full to Plus, Pro, Business, Enterprise and Edu users in ChatGPT Work!

Codex platform: The two major models are synchronized online. Free wool channels: large Free and Go users can also pre-experience the GPT-6 Luna in the desktop app.

Developer API: The model is now open in the background of the API, named gpt-6-sol and gpt-6-luna respectively.

This evening, the ASI Duo match was too fierce.
Let's get the developers up, get up!