OpenAI updates GPT-6 Sol and Luna with lower prices and fewer errors

OpenAI has released updated versions of its smaller GPT-6 models, Sol and Luna, adding cheaper API pricing, better factual accuracy and lower coding error rates just weeks after launching GPT-6 Astra. The move matters because it broadens access to OpenAI’s newest model family while intensifying its competition with Anthropic across both enterprise and developer use cases.

The company says the refreshed models are designed to bring more of Astra’s capabilities into lighter, more efficient systems. Sol is aimed at demanding technical work such as coding, while Luna is positioned for high-volume administrative tasks like summarization, extraction and quick-response support.

What OpenAI announced and why it matters

OpenAI on Tuesday expanded its GPT-6 line with new versions of Sol and Luna, two smaller models that sit below its flagship Astra release in the company’s product hierarchy. The update is part of a broader effort to make the latest generation of OpenAI models cheaper to run, easier to deploy and more practical for routine business work.

The headline change is pricing. OpenAI says access to the GPT-6 series via API will cost half as much as the earlier 5.6 versions of Sol and Luna. The company attributes the savings to engineering changes in caching and inference, which reduce the amount of compute required to deliver results.

For customers, that kind of drop is significant. Lower pricing can make it easier for startups, software teams and enterprise buyers to use AI at scale without sacrificing too much quality. It also gives OpenAI a sharper commercial argument against rivals that are pushing similar claims about speed, cost and reliability.

How are Sol and Luna different?

Sol and Luna are not intended to do the same job. OpenAI describes them as separate tiers within the GPT-6 lineup, each tuned for different kinds of work.

Sol is built for complex technical work

Sol is the model OpenAI points to for tasks that require deeper reasoning and stronger code performance. The company says the new version makes fewer mistakes than its predecessor and is more reliable in coding scenarios.

In practical terms, that makes Sol the more appropriate choice for developers, analysts and product teams that need an AI system to help write, debug or transform code with less back-and-forth correction.

Luna is aimed at routine office tasks

Luna is positioned as the lighter, more clerical member of the family. OpenAI says it is better suited to “high-volume tasks with a clear goal,” including summarizing documents, pulling structured information from text and answering short, direct questions.

That framing suggests Luna is meant for workflow automation as much as conversation. It is the kind of model companies can place behind internal tools, customer-support systems or document-processing products where throughput and price matter as much as raw intelligence.

Why OpenAI is pushing efficiency now

OpenAI’s message is not just that the models are better. It is that they are better at a lower price point, which may be even more important commercially.

The company says the new versions owe much of their cost reduction to improved caching and inference. Those are the kinds of behind-the-scenes advances that can significantly reduce serving costs when a model is used repeatedly across many requests. For a company selling AI through APIs and software subscriptions, those savings can translate into more competitive pricing, improved margins or both.

Efficiency also matters because the market is moving from novelty to utility. Early attention around AI focused on how impressive these systems were in demos. Now the pressure is on vendors to prove that the technology can be dependable, affordable and easy to deploy across daily business operations.

What OpenAI says about reliability

OpenAI says the updated Sol and Luna models are more accurate and less likely to generate incorrect responses. The company specifically highlighted internal testing that compares model outputs against de-identified real-world conversations where users had flagged mistakes.

OpenAI said its internal factuality testing found that GPT-6 Sol makes roughly half as many mistakes as its predecessor, while reaching Astra-level reliability at a significantly lower cost.

That is a meaningful claim because factual accuracy remains one of the central weaknesses of large language models. Even as they have improved, users still encounter hallucinations, logical slips and coding errors. If OpenAI’s numbers hold up in production, the new models could be more attractive for business users that need consistency as much as creativity.

How does this affect developers and paid users?

OpenAI is making the refreshed models available immediately across several products, which should help adoption accelerate.

According to the company, GPT-6 Sol and Luna are now accessible in ChatGPT Work and Codex for most paid accounts, as well as in the ChatGPT API. Luna will also be available in the desktop app and for Free and Go users, while broader rollout to ChatGPT’s app and website will continue gradually over the course of the day.

That distribution strategy matters. By spreading the models across paid and free surfaces, OpenAI can gather usage data quickly, encourage experimentation and push the upgrades into everyday workflows. It also gives the company a way to showcase its latest generation not just to developers, but to the broad consumer base that uses ChatGPT as a general-purpose assistant.

Where users are likely to notice the difference

The most visible improvements are likely to show up in tasks where models previously required multiple attempts or manual checking. Examples include:

• generating code with fewer corrections
• summarizing long documents more cleanly
• extracting fields from text with fewer formatting errors
• answering direct questions with more confidence
• handling repetitive administrative workflows at lower cost

For businesses, even small quality gains can be valuable if they reduce human review time. That is especially true in high-volume environments where thousands of outputs are produced every day.

What does this mean for OpenAI’s rivalry with Anthropic?

OpenAI is framing the release as a clear competitive win, and it is once again pointing directly at Anthropic as its main benchmark.

In its announcement, the company said Sol and Luna outperform Anthropic’s leading models, including Fable and Opus, in several tasks. That kind of comparison has become a familiar feature of model launches, where vendors routinely position their newest systems as superior to rivals on coding, reasoning or factuality.

The timing of the announcement added extra drama. Anthropic had introduced a fresh version of Opus 5.5 just 90 minutes before OpenAI’s release, underscoring how closely the two companies are tracking each other’s product cycles.

OpenAI’s release suggested that the company sees the updated GPT-6 family as a way to press its advantage in performance and pricing against Anthropic’s newest models.

The broader implication is that the frontier-model race is no longer only about who can produce the most powerful AI system. It is also about who can package strong performance into something developers and businesses can actually afford to use repeatedly.

How the rollout fits into OpenAI’s product strategy

OpenAI’s current approach appears to be one of segmentation: a flagship model for maximum capability, plus lighter options optimized for specific workflows. Astra, Sol and Luna each serve different tiers of demand, from advanced technical assistance to everyday clerical tasks.

That structure allows OpenAI to cover more of the market without forcing every user onto the same high-cost model. It is a familiar software strategy, but in AI it is particularly important because model inference costs can rise quickly as usage scales.

It also gives OpenAI flexibility in how it markets the GPT-6 family. Astra can be the company’s showcase product, while Sol and Luna can serve as the practical workhorses that bring the benefits of the new generation into mainstream business settings.

Key release timeline

Why the factuality claim matters

One of the most important parts of the release is OpenAI’s emphasis on correctness. In the current AI market, raw benchmark scores matter, but reliability is becoming just as critical.

Businesses do not simply want impressive outputs. They want outputs that are repeatable, traceable and accurate enough to use in real operations. A model that is slightly less flashy but noticeably more dependable can be more valuable than a more powerful one that needs constant correction.

That is why OpenAI’s reference to fewer factual mistakes is strategically important. It signals that the company is trying to move beyond novelty and into trust. If enterprises believe the new Sol and Luna models reduce the burden of verification, they may be more willing to expand adoption across support, research, coding and internal automation.

What happens next?

The next test will be whether users experience the same gains that OpenAI describes in its internal evaluations. Model launches often arrive with strong claims, but the real verdict usually comes after developers and companies put the tools to work in live systems.

If Sol and Luna do prove cheaper and more reliable, OpenAI could strengthen its position across a wide swath of the market, from coding assistants to office productivity tools. If the models are only modestly better, the pricing cut may still help drive adoption, especially among customers with tight budgets.

Either way, the release shows that OpenAI is treating efficiency as a core competitive advantage rather than an afterthought. In a market where rivals move quickly and model costs remain central, that may be as important as any benchmark score.

For now, the company is signaling that GPT-6 is not just a single flagship model. It is becoming a family of products, each tuned for a different balance of power, price and precision.