Key Notes
- OpenAI launched GPT-6 Sol and Luna for ChatGPT Work, Codex and the API, with regular Chat access not yet available.
- Standard API rates are $2 input and $10 output for Sol, and $0.10 input and $0.50 output for Luna, per million tokens.
- Sol scored 33.2% on AutomationBench at xhigh effort with a reported cost of $0.27 per task.
OpenAI launched GPT-6 Sol and GPT-6 Luna on September 22, expanding its latest model family with lower-cost options for coding, agent workflows and routine processing. Sol’s standard API prices are half the promotional rates of GPT-5.6 Sol, while Luna’s input and output prices fall by 50% and about 58%, respectively.
OpenAI positions Sol for complex coding and agent work, including feature development, debugging and analysis. Luna targets focused tasks performed at scale, such as extracting information and summarizing material. The two models expand the choices below GPT-6 Astra, which remains the company’s highest-capability offering.
The business significance is the cost of sustained use. A model used once to answer a question and an agent making repeated calls over a long assignment can create very different bills. Lower token prices make more experimentation possible, although the cost of a successful result still depends on accuracy and retries.
How the New Prices Compare
In its announcement, OpenAI lists Sol at $2 per million input tokens and $10 per million output tokens. The previous promotional prices were $4 and $20. Luna costs $0.10 for input and $0.50 for output, compared with $0.20 and $1.20 for its predecessor.
The Luna figures deserve a precise reading: cutting $1.20 to $0.50 is a reduction of roughly 58.3%, rather than exactly 50%. At the new base rates, Luna’s input and output tokens each cost one-twentieth as much as Sol’s. That price relationship does not imply equivalent capabilities.
For illustration, one million uncached input tokens plus one million output tokens would cost $12 with Sol or $0.60 with Luna under standard short-context pricing. These are token-only calculations, not estimates for a complete agent job. Tool charges, cached content and the amount of reasoning generated can change the final bill.
OpenAI’s Sol documentation also specifies higher rates when a prompt exceeds 272,000 input tokens. For those requests, input and cache rates double and output rates rise by 50%, applying to the full request. A long context window therefore should not be confused with a single flat price at every context length.
Batch and Flex processing cost half the standard rates, while Fast mode costs twice the applicable rates. Regional processing adds a 10% premium where offered. These distinctions give developers several ways to trade turnaround time and processing requirements against cost.
AutomationBench Shows the Trade-Off
GPT-6 Sol scored 33.2% on AutomationBench at xhigh reasoning effort, with a reported cost of $0.27 per task. OpenAI compares that with 26.9% for Claude Opus 5 at maximum effort and approximately 11.1 times Sol’s cost per task.
The independent Zapier leaderboard confirms Sol’s 33.2% result and $0.27 cost. It also shows why this is a price-performance argument rather than an overall lead: Astra reaches 41.4% at maximum effort, and Claude Opus 5.5 reaches 40.0%, at higher reported costs.
The 31.4% Fable 5.1 comparison includes an Opus 5 fallback. Zapier says that fallback handled about 40% of the tasks, and its listed cost excludes those fallback tokens. The benchmark measures strict completion of simulated business workflows, not the percentage of ordinary office jobs a model can replace.
OpenAI’s comparison with Astra uses Astra at low effort. It does not show Sol beating Astra at every setting. Reasoning effort is part of the product choice: allocating more computation may improve a result, increase its cost or sometimes fail to produce a better score.
For a business deploying agents, a useful evaluation would include the expense of failed attempts and human correction. A lower-priced model that completes routine work consistently may be economical, while a harder task could justify paying for a stronger model. The leaderboard supplies evidence for those decisions without settling them for every workflow.
Better Factuality Is a Measured Claim
OpenAI says Sol produces about half as many factual mistakes as GPT-5.6 Sol on its internal evaluation, approaching Astra’s reliability at a lower price. The evaluation uses conversations selected because users had flagged earlier factual errors. OpenAI explicitly says that sample is not representative of typical usage.
The distinction prevents a misleading interpretation. The result does not mean that every customer will experience exactly half as many errors, or that an unverified answer can now be treated as authoritative. It measures improvement on a particular collection of difficult examples.
That progress still matters for work involving long documents or several connected steps. An incorrect assumption introduced early can influence later output even when the subsequent reasoning is coherent. Customers need evaluations that inspect the evidence behind an answer as well as whether its final wording sounds convincing.
OpenAI also reports alignment improvements, including fewer misleading claims about coding work in its tests. Its recent misalignment reporting framework provides context for how the company discusses such failures. Evaluation gains should be read alongside the situations being tested and the behavior observed in deployment.
What Developers Can Build With Sol and Luna
Both models support text and image input with text output. Their documentation lists a 1,050,000-token context window and up to 128,000 output tokens. Sol’s stated knowledge cutoff is April 20, 2026; Luna’s specifications give May 18, 2026. Current information still requires supplied material or appropriate tools.
Luna supports function calling and structured output, making its lower price relevant to applications that need predictable fields rather than a long conversational answer. Examples include extracting dates from documents or organizing incoming records. These are potential deployment patterns, not guarantees that every extraction will be correct.
Both models offer several reasoning levels. Built-in tools are available through the Responses API; function calling through Chat Completions requires reasoning effort to be set to none. EU data residency is limited to Standard processing.
Cached input reads receive a 90% discount relative to uncached input. Reusing stable instructions or document context can therefore change the economics of repeated calls. Developers still need to account for cache writes and measure how often their application actually reuses cached material.
Availability Starts With Work, Codex and the API
GPT-6 Sol and Luna are rolling out in ChatGPT Work and Codex for Plus, Pro, Business, Enterprise and Edu users. Free and Go users can access Luna in the desktop app. OpenAI says the models are not yet available in regular Chat.
API customers can request them as gpt-6-sol and gpt-6-luna. OpenAI says the ChatGPT rollout will proceed gradually through launch day, so access may not appear simultaneously for every eligible account.
The combination of lower rates and stronger measured performance broadens the tasks worth testing with AI agents. For customers, the practical next comparison is whether Sol or Luna can deliver the required result within a defined budget, with enough consistency that the savings survive review and correction.
Disclaimer: AIstify is an independent media brand owned and operated by NuvexMedia LLC, publishing news, research, and insights on artificial intelligence, emerging technologies, automation, and related industries. NuvexMedia LLC invests in and collaborates with companies across the AI, technology, software, and digital innovation sectors. These relationships do not influence AIstify’s editorial coverage, and the publication maintains full editorial independence to provide accurate, timely, and objective information. © 2026 NuvexMedia LLC. All rights reserved. This content is for informational purposes only and should not be considered legal, tax, investment, financial, or other professional advice.
