OpenAI Wants GPT-6 Sol and Luna to Do More Work for Less

 

Image: Digiopedia / Illustration

OpenAI’s latest GPT-6 models aren’t trying to replace Astra at the top. Instead, Sol and Luna are aimed at making capable AI agents, coding tools and everyday automation considerably cheaper to run.

Less than three weeks after introducing GPT-6 Astra, OpenAI is expanding its newest model family with two models focused less on pushing the absolute performance frontier and more on the economics of putting AI to work.

GPT-6 Sol and GPT-6 Luna, released on September 22, are designed to bring some of the advances developed for Astra into faster and substantially cheaper models. OpenAI says the models improve across professional work, factual reliability, coding, computer use and alignment while cutting API prices compared with their GPT-5.6 predecessors.

The positioning is straightforward. Astra remains OpenAI’s highest-capability GPT-6 model. Sol sits below it as a more economical model for demanding coding and agentic work, while Luna targets high-volume tasks where efficiency matters most.

That makes Sol and Luna less of a dramatic new frontier-model moment than Astra was. Their significance is different: OpenAI is trying to make increasingly capable AI economical enough to use repeatedly, at scale.

And as AI agents begin consuming far more tokens than conventional chatbot interactions, that may matter as much as another few points on a benchmark.

GPT-6 is becoming a model family

OpenAI launched GPT-6 Astra on September 3 as the highest-end member of its new generation. Sol and Luna extend that generation downward rather than attempting to displace Astra.

The three models now occupy distinct positions:

GPT-6 Astra is intended for OpenAI’s hardest end-to-end work and most demanding reasoning.

GPT-6 Sol is designed for complex coding, professional work and agentic workflows where capability still matters but Astra’s cost may not be justified.

GPT-6 Luna is optimized for focused, repeatable and high-volume tasks where low cost and efficiency are the priority.

OpenAI says Sol and Luna were trained using methods similar to those behind Astra, carrying improvements from the higher-end model into less expensive tiers.

That structure increasingly resembles the way cloud computing is sold: users are not expected to choose the most powerful option every time. Instead, different workloads can be routed to different levels of intelligence depending on how difficult, urgent or valuable the task is.

The biggest change is the price

The headline feature of Sol and Luna is not a spectacular new capability. It is cost.

For standard API usage on shorter contexts, GPT-6 Sol costs $2 per million input tokens and $10 per million output tokens.

GPT-6 Luna costs just $0.10 per million input tokens and $0.50 per million output tokens. Cached input costs $0.20 per million tokens for Sol and $0.01 for Luna.

OpenAI describes the new models as 50% cheaper than the promotional pricing of the equivalent GPT-5.6 models. The comparison is important: it is specifically against GPT-5.6’s promotional rates, not necessarily every historical price OpenAI has charged for those models.

For comparison, Astra's standard API rates are substantially higher at $10 per million input tokens and $50 per million output tokens.

Model	Input per 1M tokens	Output per 1M tokens	Position
GPT-6 Astra	$10	$50	Highest capability
GPT-6 Sol	$2	$10	Capability/cost balance
GPT-6 Luna	$0.10	$0.50	High-volume efficiency

Those differences become particularly significant in agentic systems.

A conventional chatbot might generate one answer and stop. An AI agent can repeatedly inspect files, reason, invoke tools, write code, evaluate results and continue working. A single assignment can therefore involve many model calls and large amounts of context.

At that scale, the cost per token stops being an abstract API specification and becomes a central product constraint.

Why cheaper AI matters more as agents get busier

OpenAI’s launch announcement gives an unusually revealing example of how quickly agent costs can accumulate.

The company says that, when valued at API prices, internal coding-agent usage has exceeded $600 per day for the median OpenAI researcher and $7,000 per day for researchers at the 90th percentile.

Those figures describe OpenAI’s own environment and should not be treated as representative of ordinary developers. But they illustrate why model efficiency is becoming important as AI systems take on longer jobs.

If an agent works for minutes or hours instead of responding once, even relatively small efficiency gains can compound.

That is the real proposition behind Sol and Luna: not simply cheaper conversations, but cheaper workflows.

Independent coverage has reached a similar conclusion. Ars Technica characterized the latest OpenAI and Anthropic releases as part of a shift toward models delivering incremental capability gains while competing increasingly on operating cost.

Sol is aimed at serious work without Astra pricing

GPT-6 Sol is the more capable of the two new models.

OpenAI describes it as a model built for complex coding and agentic workflows, and its API supports reasoning settings ranging from no additional reasoning through low, medium, high, extra-high and maximum effort.

In OpenAI’s AutomationBench results, which test agents completing business workflows across multiple applications, GPT-6 Sol at its highest tested reasoning level scored 33.2%.

OpenAI reported GPT-6 Astra at low effort at 30.3% in the same comparison, although comparing different models at different reasoning settings does not establish that Sol is generally more capable than Astra. OpenAI continues to describe Astra as its strongest model overall.

The more relevant number may be cost per completed task.

OpenAI says Sol achieved its AutomationBench result at $0.27 per task, while the tested low-effort Astra configuration cost 3.9 times as much. It also reported substantially lower task costs than the Anthropic models included in its evaluation.

Those comparisons come from OpenAI and depend on specific model settings, benchmark methodology and token prices. They are useful indicators, but they should not be read as universal rankings of model quality.

Luna makes the more aggressive bet on efficiency

If Sol is intended to be a workhorse, Luna pushes much further toward low-cost scale.

At $0.10 per million input tokens and $0.50 per million output tokens, Luna costs a fraction of Sol and Astra.

OpenAI positions it for focused tasks that may be performed frequently: triage, repeatable processing, extraction, structured work and high-volume automations.

What makes Luna interesting is that OpenAI is not presenting it simply as a lightweight model for trivial requests.

On the company's DeepSWE 1.1 software-engineering evaluation, GPT-6 Luna at maximum reasoning effort scored 66.6%. Sol scored 68.8% at its maximum setting. OpenAI says Luna also produced competitive results against considerably more expensive models in its comparison.

Again, those are vendor-reported benchmark results rather than an independent guarantee of real-world performance.

But they show how OpenAI wants developers to think about Luna: a model cheap enough for large-scale deployment without necessarily being limited to basic summarization.

OpenAI says factual reliability is improving, too

Lower prices would be less useful if they came with significantly less reliable answers.

OpenAI says GPT-6 Sol produced about half as many mistakes as GPT-5.6 Sol in an internal factuality evaluation based on de-identified conversations in which users had previously flagged factual errors.

The company says Luna also improved substantially and, at higher reasoning levels, can approach the factual performance of significantly more expensive earlier models.

There is an important limitation.

OpenAI explicitly says the conversations in this evaluation were selected because they had produced factual errors before, so the dataset is not representative of typical ChatGPT usage.

That caveat matters. A 50% reduction in errors on an error-heavy evaluation does not mean a user should expect exactly half as many hallucinations in every real-world task.

It does, however, suggest factual reliability remains a major target as OpenAI tries to push smaller models into more consequential workflows.

Coding is where the economics become particularly interesting

Coding agents are one of the clearest applications for the new pricing structure.

Modern coding systems do much more than autocomplete a line. They may inspect a repository, search through files, plan a change, edit multiple components, run tests, diagnose failures and repeat the process.

OpenAI says GPT-6 Sol substantially improves over GPT-5.6 Sol on FrontierCode, an evaluation that considers not only whether code works but whether a change is suitable to merge into a real codebase.

On DeepSWE, OpenAI reports Sol at 68.8%, compared with 69.9% for Anthropic's Claude Fable 5 at its highest reported configuration in OpenAI's comparison. OpenAI says Sol completed those tasks at approximately 80% lower cost per task.

The narrow score difference is less important than the direction of competition.

AI labs are increasingly competing not just over which model can solve a coding task, but over how much it costs to solve that task reliably.

That is a much more relevant metric for businesses running thousands or millions of model interactions.

Computer use is coming down the model stack

GPT-6 Astra put heavy emphasis on AI that can operate software and computers. Sol and Luna bring some of that capability into cheaper tiers.

On the offline portion of OSWorld 2.0, OpenAI reports GPT-6 Sol at extra-high reasoning effort scoring 60.5%, while the Claude Opus 5 configuration included in its comparison scored 60.3%. OpenAI estimates Sol's cost per task was about 80% lower in that test.

Luna, meanwhile, reportedly surpassed a medium-effort GPT-5.6 Sol configuration at around one-tenth of the cost.

OpenAI still describes Astra as its best model for computer use.

The point of Sol and Luna is therefore not to make Astra obsolete. It is to make computer-operating agents affordable enough that developers do not necessarily need the premium model for every step of a workflow.

All three GPT-6 models have unusually large context windows

Despite their large price differences, Astra, Sol and Luna share the same headline context capacity in OpenAI's API documentation.

Each supports a 1,050,000-token context window and up to 128,000 output tokens.

That gives Sol and Luna room to work with large documents, long-running agent histories and substantial codebases.

Pricing changes for especially large requests, however. OpenAI's model documentation says prompts exceeding 272,000 input tokens are charged at higher rates, so the base token prices do not apply unchanged across the model's entire maximum context window.

That distinction is worth noting for developers estimating costs from the headline API prices.

OpenAI is also making caching more important

Alongside lower token prices, OpenAI has upgraded prompt caching across GPT-6.

Caching allows an application to reuse portions of context instead of paying the full input price every time the same information is sent back to a model.

OpenAI says cached GPT-6 input reads can receive a 90% discount, while changes to reasoning effort or tool availability can now preserve earlier context for cache reuse.

For long-running agents that repeatedly carry the same instructions, documents or codebase context between steps, that can materially change total operating costs.

It also reinforces the larger message of the launch: OpenAI is optimizing not just the intelligence of individual responses, but the infrastructure required to keep an AI system working over time.

Where can you use GPT-6 Sol and Luna?

As of September 23, GPT-6 Sol and Luna are available through the OpenAI API, ChatGPT Work and Codex.

OpenAI's current ChatGPT documentation says these Work and Codex models are separate from those available in ordinary ChatGPT conversations. Availability within Work and Codex can also depend on plan and workspace settings.

OpenAI says Plus, Pro, Business, Enterprise and Edu users can access Sol and Luna through Work and Codex, while Free and Go users can access GPT-6 Luna through the desktop app where supported.

Developers can call the models directly using gpt-6-sol and gpt-6-luna.

The AI race is increasingly becoming a price war

There is a broader industry story behind this release.

OpenAI announced Sol and Luna on the same day Anthropic introduced a less expensive version of Opus, underscoring how quickly frontier-model competition is shifting from raw capability toward price, speed and efficiency.

For several years, the easiest way to market a new AI model was to show that it scored higher on benchmarks.

That is becoming less sufficient.

Developers care about latency. Businesses care about predictable costs. Agent builders care about how much a complete task costs rather than how much a single token costs. And many applications do not need the most intelligent model available for every operation.

The result is an emerging hierarchy in which premium models tackle the hardest reasoning steps while cheaper models handle routine portions of a workflow.

Sol and Luna fit directly into that strategy.

The real test will happen outside the benchmarks

OpenAI's numbers paint a favorable picture: stronger coding, better computer use, fewer factual errors and substantially lower prices.

But most of those performance claims come from OpenAI itself.

Real applications are messier than controlled benchmarks. They contain ambiguous instructions, unpredictable websites, unusual codebases, incomplete information and workflows that can fail in ways benchmark scores do not capture.

The key question is therefore not whether GPT-6 Sol can beat a particular competitor by a few percentage points on one test.

It is whether developers can move meaningful workloads from expensive frontier models to Sol or Luna without losing enough reliability to erase the savings.

That will take broader use and independent evaluation to establish.

The bottom line

GPT-6 Sol and Luna are not designed to make Astra irrelevant.

They are designed to make using GPT-6 routine.

Astra remains OpenAI's model for the most demanding work. Sol offers much of the new generation's capability at one-fifth of Astra's standard input and output token prices. Luna pushes the economics much further, targeting workloads where AI may need to run constantly rather than occasionally.

That makes this release less about another spectacular leap in what AI can theoretically accomplish and more about something potentially just as consequential: how cheaply useful intelligence can be deployed repeatedly.

The frontier AI race is no longer only about building the smartest model.

Increasingly, it is about making capable models inexpensive enough that people can afford to keep them working.