OpenAI Expands GPT-6 With Cheaper Sol And Luna Models — And Cuts API Prices By 50%

OpenAI Just Cut The Price Of GPT-6 In Half — And The AI Cost War Is Accelerating

The AI Price War Intensifies As OpenAI Launches Cheaper GPT-6 Models

OpenAI has expanded its GPT-6 family with two new models, GPT-6 Sol and GPT-6 Luna, cutting their API prices by 50% compared with the promotional pricing of their GPT-5.6 predecessors while bringing across advances developed for the more powerful GPT-6 Astra.

The launch on September 22, 2026, gives OpenAI a three-level GPT-6 line: Astra remains the company's most capable option, while Sol targets demanding professional work at a lower cost and Luna is designed for cheaper, high-volume workloads.

The pricing change is substantial.

GPT-6 Sol costs $2 per million input tokens and $10 per million output tokens, down from $4 and $20 respectively for GPT-5.6 Sol's promotional pricing.

GPT-6 Luna costs just $0.10 per million input tokens and $0.50 per million output tokens, compared with $0.20 and $1.20 for GPT-5.6 Luna. OpenAI describes both models as 50% cheaper overall than the promotional rates of the previous generation.

That makes this more than another routine model release.

OpenAI is increasingly competing not only over which company can produce the most capable artificial intelligence, but over how cheaply that intelligence can be deployed across software development, AI agents, business automation and everyday digital services.

GPT-6 Is Becoming A Family Rather Than A Single Flagship Model

GPT-6 began earlier in September with Astra, which OpenAI continues to describe as its strongest model for the hardest tasks.

Sol and Luna now fill out the cheaper end of the range.

OpenAI says it trained both using methods similar to those behind Astra, transferring improvements in professional work, factual reliability, coding, computer use and alignment into models designed to operate at significantly lower cost.

The distinction matters because most real-world AI applications do not require the most powerful model available for every individual step.

A business might use Astra for a difficult strategic analysis, Sol for repeated coding or agent tasks and Luna for huge volumes of classification, extraction, summarisation or relatively straightforward responses.

That kind of model routing can radically change the economics of an AI product.

Instead of sending every request through an expensive flagship system, companies can reserve premium intelligence for the tasks where it materially improves the outcome.

GPT-6 Sol Costs Half As Much As GPT-5.6 Sol

OpenAI's previous promotional price for GPT-5.6 Sol was $4 per million input tokens and $20 per million output tokens.

GPT-6 Sol reduces those figures to $2 and $10.

That is particularly significant for agentic applications, where a single user request can trigger multiple model calls, tool operations and reasoning steps.

A relatively small reduction in token cost can therefore multiply into a much larger saving when millions or billions of tokens are being processed.

OpenAI is pitching Sol as a serious professional model rather than a stripped-down budget alternative.

On AutomationBench, which evaluates AI agents completing business workflows using dozens of tools, OpenAI reported that GPT-6 Sol at extra-high reasoning effort scored 33.2%.

GPT-6 Astra at low effort scored 30.3%, while Claude Opus 5 at maximum effort scored 26.9% in the comparison published by OpenAI.

OpenAI calculated Sol's cost at $0.27 per task and said Opus 5 cost 11.1 times as much in that evaluation.

Those figures should be treated in context: benchmark performance varies according to prompts, reasoning settings, tooling and evaluation methodology, and OpenAI notes that some competitor results came from publicly reported figures rather than identical in-house runs.

Even with that qualification, OpenAI's strategic message is obvious.

The company wants Sol to compete not merely by being cheaper than Astra, but by making capable agentic AI economical enough to use repeatedly.

GPT-6 Luna Pushes AI Prices Even Lower

Luna is where the numbers become particularly striking.

At $0.10 per million input tokens and $0.50 per million output tokens, it is priced for workloads where volume matters more than extracting the absolute maximum intelligence from every request.

That could include document processing, classification, customer-service triage, summarisation, structured extraction and large fleets of AI workers handling clearly defined tasks.

Yet OpenAI says the new Luna also delivers meaningful capability gains.

On its internal factuality evaluation, OpenAI said GPT-6 Sol makes roughly half as many mistakes as GPT-5.6 Sol.

It also reported that GPT-6 Luna at higher reasoning settings can reach factual reliability comparable with GPT-5.6 Sol at roughly one hundredth of the cost in that particular evaluation.

OpenAI cautions that the factuality test was deliberately constructed from conversations where users had previously flagged errors, meaning it is not representative of the average ChatGPT conversation.

Nevertheless, the direction is important.

The cheapest models are becoming capable enough to perform work that recently required substantially more expensive systems.

Coding Could Be One Of The Biggest Battlegrounds

Coding is a major focus of the GPT-6 Sol and Luna launch.

OpenAI says the cost of sustained coding-agent usage is becoming increasingly important as developers allow AI systems to work on longer tasks and entire codebases rather than individual snippets.

On the DeepSWE 1.1 software-engineering benchmark, OpenAI reported GPT-6 Sol at maximum reasoning effort scoring 68.8%.

Claude Fable 5 reached 69.9% at its highest reported setting, but OpenAI estimated that GPT-6 Sol completed tasks at roughly 80% lower cost.

GPT-6 Luna scored 66.6% at maximum effort.

OpenAI said that was comparable with Claude Opus 5 and Fable 5 at medium reasoning levels while costing 93% less per task than Opus 5 and 96% less than Fable 5 in its comparison.

Again, benchmark comparisons between competing companies deserve caution.

But the wider trend is hard to miss.

The frontier of AI competition is shifting from simply asking which model scores highest to asking how much useful work a model produces for every dollar spent.

Computer-Using AI Is Getting Cheaper Too

The same pattern appears in computer-use benchmarks.

OpenAI says GPT-6 Astra remains its strongest model for controlling computers, but Sol and Luna provide lower-cost alternatives.

On OSWorld 2.0 offline, GPT-6 Sol at extra-high effort scored 60.5%, compared with 60.3% for Claude Opus 5 at medium effort.

OpenAI estimated Sol's cost per task to be around 80% lower in that comparison.

Luna at maximum effort was also reported to outperform GPT-5.6 Sol at medium effort while costing roughly one tenth as much.

That matters because computer-use agents can consume substantial numbers of tokens while repeatedly examining screens, planning actions and deciding what to do next.

Lower model prices could make such systems economically viable in situations where using a flagship model for every interaction would be too expensive.

OpenAI Is Also Cutting Costs Through Better Caching

The headline price cuts are only part of the story.

OpenAI has also improved prompt caching across GPT-6.

Caching allows applications to reuse sections of context that have already been processed rather than paying the full price to process the same information repeatedly.

OpenAI says cached input-token reads can receive discounts of 90%.

It has also introduced tools that allow developers to alter reasoning effort or enabled tools without necessarily invalidating previously cached context.

For long-running agents, that could matter enormously.

An AI agent working with a large repository, policy manual or conversation history may repeatedly send the same background information to a model.

If much of that context can remain cached, the effective cost of each additional interaction falls.

OpenAI said GitHub reported that caching improvements had reduced the proportion of prompt tokens requiring fresh processing by more than 50% across billions of requests to OpenAI models over recent months.

OpenAI And Anthropic Are Fighting On Price

The timing is notable.

Anthropic also launched Claude Opus 5.5 on September 22, with the company presenting its own release around stronger cost efficiency.

The Next Web reported that Anthropic's announcement came roughly 90 minutes before OpenAI published GPT-6 Sol and Luna.

That does not prove either release was a direct response to the other.

Model launches are prepared well in advance.

But together they illustrate the increasing importance of cost in the AI market.

For developers and businesses, a model that is slightly more capable but several times more expensive may not always be the best fit.

At sufficient scale, token pricing directly affects whether an AI product is profitable.

This creates pressure on laboratories to deliver more intelligence per dollar rather than simply pushing benchmark scores higher regardless of expense.

The AI Industry Is Moving Towards Model Routing

There is also a deeper architectural trend behind the launch.

AI systems increasingly do not need to depend on one model.

An application can decide dynamically which level of intelligence a particular task requires.

Easy requests can go to Luna.

Harder professional work can go to Sol.

The most complex or consequential problems can be escalated to Astra.

That approach resembles the way computing infrastructure already allocates resources according to workload.

Running every task through the most expensive available AI model can be wasteful.

Running complex tasks through an excessively weak model can create errors and require more retries.

The economic goal is therefore to match model capability to task difficulty as efficiently as possible.

GPT-6 Sol and Luna make that strategy more attractive because the price gap between the tiers is enormous.

Where Can People Use GPT-6 Sol And Luna?

OpenAI says GPT-6 Sol and GPT-6 Luna are available through the API under the model identifiers gpt-6-sol and gpt-6-luna.

They are also rolling out in ChatGPT Work and Codex for Plus, Pro, Business, Enterprise and Edu customers.

Free and Go users can access GPT-6 Luna through the desktop app.

At launch, OpenAI said the models were not yet generally available in ordinary Chat, with the rollout taking place gradually to maintain service stability.

Availability may therefore differ depending on the product, subscription and timing.

Astra Still Sits At The Top

The arrival of cheaper models does not replace GPT-6 Astra.

OpenAI explicitly recommends Astra when customers want its strongest overall performance without prioritising cost.

That leaves the company with an increasingly clear hierarchy.

Astra is the premium option.

Sol targets complex professional work with considerably lower operating costs.

Luna is intended for speed, scale and cost-sensitive workloads.

The more interesting question is how often users will genuinely need the premium tier as the cheaper models improve.

If Sol can perform many professional tasks close to Astra's level, and Luna can perform work that previously required Sol-class models, then capability gradually migrates down the price ladder.

That has happened repeatedly across computing.

Yesterday's premium capability eventually becomes today's standard feature.

Why The 50% Price Cut Matters

AI model launches are usually marketed around higher benchmark scores, better reasoning or new capabilities.

GPT-6 Sol and Luna illustrate another increasingly important metric: economic usefulness.

For individual ChatGPT users, token prices can appear abstract.

For companies processing billions of tokens, they are not.

A 50% reduction can determine whether an automated workflow is used occasionally or continuously.

It can decide whether an AI agent supports a few employees or thousands.

And when caching, model routing and lower inference costs are combined, the reduction in the cost of completing an entire workflow can potentially be larger than the headline token-price cut alone.

OpenAI's GPT-6 expansion therefore says something important about where the AI race may be heading.

The competition is no longer solely about building the most intelligent model.

It is increasingly about making high-quality intelligence cheap enough to use everywhere.