OpenAI Adds GPT-6 Sol and Luna Models to ChatGPT

I recently spoke with a senior engineer managing automated scripts at a large financial institution. His main complaint did not concern the intelligence of language models. He complained about the price. Every time his software asked a question, he paid a fraction of a cent. Those fractions add up quickly when you process a million requests a day. Today, OpenAI responded directly to that financial pressure. The San Francisco firm officially released two new variations of its artificial intelligence architecture, naming them GPT-6 Sol and GPT-6 Luna. These additions join the massive GPT-6 Astra system introduced earlier this month, bringing heavy computational power down to a much cheaper price bracket.

The pricing mechanics are the most interesting part of this announcement. Programmers and corporate clients who rely on application programming interfaces constantly monitor their daily expenses. By releasing Sol and Luna, OpenAI cut the cost of access by fifty percent compared to the promotional pricing of the older GPT-5.6 equivalents. The company claims they achieved this massive price reduction by improving their internal caching systems and making the inference process require less raw electricity.

The breakdown of the new lineup reveals a distinct strategy regarding size and capability. The company apparently abandoned the medium-tier model, previously known internally as Terra. They opted for a three-tiered structure instead. Luna serves as the smallest, fastest option. It is built for high-volume, rapid-fire tasks where speed matters more than deep reasoning. If a developer needs to quickly sort thousands of customer reviews or extract dates from messy text files, Luna is the correct choice. It requires very few computing tokens to return an answer.

Sol sits directly above Luna. This model acts as the workhorse for the professional class. The developers trained Sol exactly for complex programming duties and agentic workflows. When an intelligent software agent needs to string together ten different commands to compile a piece of code or debug a broken website, Sol provides the necessary logic without costing a fortune. Early testing shows that Sol matches the performance of top-tier competing models while costing nearly eighty percent less per task.

We recently covered how OpenAI unveils Astra multimodal artificial intelligence model to handle massive visual and audio inputs. While Astra remains the absolute top tier choice for highly complicated medical or scientific research, it is simply too expensive for everyday office work. Sol and Luna fill that exact gap. They provide access to the same underlying intelligence without requiring a massive budget.

The integration into existing software products is happening immediately. Paying subscribers using the Plus, Pro, Business, Enterprise, and Education tiers will see the new models appear inside ChatGPT Work and Codex starting today. Free tier users are not completely excluded. They can access the smaller Luna model exclusively through the dedicated desktop application. Interestingly, the company decided to keep these models out of the standard consumer Chat interface for now, restricting them to professional workspaces. Corporate administrators must actively flip a switch in their settings menu to grant their employees access.

This release arrives during a brutal price war across the artificial intelligence sector. Rival firms constantly try to undercut each other by offering cheaper access to their respective application programming interfaces. When a developer builds a new software product, they usually choose one single provider and stick with them. Switching providers later requires rewriting massive amounts of code. By lowering the price of Sol and Luna so aggressively, OpenAI wants to lock developers into their ecosystem before they consider trying alternative services. The strategy is entirely about volume. They prefer to make a tiny amount of money on billions of daily requests rather than a large amount of money on a few requests.

The pressure to maintain safety while dropping prices is immense. After a recent public mistake, OpenAI confirms wiki incident promises disclosure rules, proving that faster output sometimes leads to unpredictable errors. When you speed up the reasoning process and reduce the size of the model, you risk increasing the hallucination rate. The engineering team claims they used the exact same training methods for Sol and Luna as they did for the massive Astra model. This shared lineage supposedly ensures that the smaller models maintain high levels of factual accuracy, even when operating at maximum speeds.

We must look at the physical infrastructure required to support this launch. Every time a new model drops, millions of users rush to test it simultaneously. This sudden spike in traffic places an incredible strain on the remote server farms hosting the software. If the servers crash, the corporate clients who rely on Codex to write their daily software code will experience massive delays. By creating smaller models that require less memory, the company effectively increases their total server capacity without actually buying more computer chips. Luna and Sol can process more user requests simultaneously than the heavier Astra system.

Codex is perhaps the biggest beneficiary of this update. Software engineers rely heavily on Codex to autocomplete their programming scripts and find hidden bugs. When a programmer asks the software to fix a broken line of code, they expect an answer in milliseconds. If the model hesitates, the programmer loses their train of thought. Sol is designed strictly to solve this exact delay. It gives developers the confidence to ask much more complicated programming questions because they know the answer will arrive almost instantly. The reduced cost also means a small startup can afford to give every single one of their engineers full access to the tool.

The artificial intelligence industry is maturing rapidly. Last year, the narrative focused entirely on building the biggest, smartest models possible. Companies bragged about parameter counts and massive training clusters. Today, the conversation is entirely focused on margins and practical utility. A brilliant model is completely useless if a small business cannot afford to turn it on. This change mirrors the early days of personal computing, where massive mainframes eventually gave way to smaller, cheaper desktop machines that anyone could buy.

Some industry leaders remain highly cautious about this rapid expansion. We noted recently that the Anthropic CEO urges industry slowdown advanced AI models to ensure proper safety guardrails are established. OpenAI clearly disagrees with that cautious approach. They are flooding the market with cheaper, highly capable tools to ensure their architecture becomes the default standard for global business operations.

One area where this new lineup truly shines involves direct computer use. In recent offline testing using the OSWorld 2.0 benchmark, the Sol model achieved an incredibly high success rate when tasked with navigating operating systems and executing commands. The test results showed Sol matching the performance of much older, heavier competing models while using a fraction of the computing power. This means an automated agent can now open a spreadsheet, read the data, copy the relevant numbers, and paste them into an email without drawing massive amounts of electricity.

This capability is terrifying for some and highly appealing for others. When a cheap language model can operate a mouse cursor and read a screen autonomously, a massive portion of administrative office work becomes instantly obsolete. The Luna model pushes this reality even further. While it might not possess the deep reasoning required to handle complex visual tasks, it can still process simple text commands at a speed that a human typist cannot possibly match. This allows companies to automate their simple customer service routing entirely, saving millions of dollars in payroll expenses.

The absence of the medium Terra model tells a compelling story about consumer behavior. When presented with three choices, buyers typically avoid the middle option. They either want the absolute cheapest tool available, or they want the absolute smartest tool available. By eliminating the middle tier entirely, the company simplifies the purchasing decision for chief information officers. You either buy speed with Luna, or you buy deep capability with Sol and Astra. This streamlined product menu reduces confusion and makes the sales pitch much easier for enterprise account managers.

As the year progresses, we expect to see even deeper integration of these models into everyday applications. Microsoft will likely push Sol and Luna into their enterprise office suites, allowing word processors and presentation software to generate content without stalling the entire computer. The financial gravity of this price drop cannot be ignored. When the cost of intelligence approaches zero, the entire internet changes. We will see a massive explosion of automated websites, fully generated applications, and synthetic media. The guardrails currently protecting the internet from automated spam will face a severe stress test as Luna becomes widely available to the public.

Read More on TechRobust:

• OpenAI cracks millennium math problem Navier Stokes
• Sam Altman calls 2026 OpenAI IPO plan ill-advised
• OpenAI partners Samsung next gen AI silicon