OpenAI GPT-6 pricing falls by half with new Sol and Luna models

OpenAI has quietly rewritten its pricing playbook for artificial intelligence models, and the numbers are hard to ignore. The company released two new additions to its GPT-6 family, GPT-6 Sol and GPT-6 Luna, cutting API costs by roughly half compared with GPT-5.6 promotional rates. The move puts OpenAI GPT-6 pricing at the center of a broader industry shake-up, one that arrived on the same day Anthropic rolled out its own cost-cutting release, according to CNBC.

Key takeaways

• GPT-6 Sol is priced at $2 per million input tokens and $10 per million output tokens, roughly 50% cheaper than GPT-5.6 Sol.
• GPT-6 Luna costs $0.10 input and $0.50 output per million tokens, a drop of about 58% on the output side.
• Sol reportedly outperforms Claude Opus 5 on AutomationBench while costing about 9% as much per completed task.
• Luna scores 66.6% on the DeepSWE v1.1 coding benchmark at roughly 93% lower cost than Opus 5, per OpenAI’s figures.
• Both models are live in the OpenAI API, ChatGPT Work and Codex, with Luna also reaching the ChatGPT desktop app for Free and Go users; neither is yet available in standard Chat mode.

Three Tiers, One Pricing Strategy

OpenAI now sells the GPT-6 lineup as a three-tier system built on the same training approach. GPT-6 Astra sits at the top, reserved for the hardest reasoning and multi-step tasks since its launch earlier this month. GPT-6 Sol slots underneath, aimed at complex coding and professional workloads that developers and knowledge workers repeat constantly, like building features, reviewing code and analyzing data. GPT-6 Luna occupies the bottom tier, built for high-volume, tightly scoped jobs such as summarization, extraction and answering routine questions, according to VentureBeat.

That segmentation matters for a simple reason: agent-driven workflows depend heavily on how many model calls a task requires and whether every step actually needs a frontier-level model. Routing simpler jobs to a cheaper tier like Luna, while reserving Astra for genuinely hard problems, changes the underlying economics of running AI agents at scale.

How Much Cheaper Are Sol and Luna?

OpenAI says the two new models are 50% cheaper than their GPT-5.6 predecessors across the board, though the actual numbers show some nuance. GPT-6 Sol now costs $2 per million input tokens and $10 per million output tokens, down from $4 and $20 — an exact 50% cut in both directions. GPT-6 Luna dropped from $0.20 to $0.10 on input, also a 50% reduction, but its output price fell from $1.20 to $0.50, a steeper cut of roughly 58%.

An OpenAI spokesperson confirmed to VentureBeat that these rates are permanent, not limited-time promotional pricing. For context, Sol’s $2/$10 pricing lands exactly at the level Anthropic set for Claude Sonnet 5, while sitting at half the token price of Anthropic’s newly announced Opus 5.5.

Benchmarks: Cost Per Completed Task

Rather than leaning purely on raw benchmark scores, OpenAI is framing its pitch around cost per completed task. On AutomationBench 1.0.6, OpenAI reports GPT-6 Sol at its highest effort setting scoring 33.2% at $0.27 per task. By comparison, Claude Opus 5 at maximum effort scored 26.9% while costing 11.1 times as much per task, according to OpenAI’s figures. Low-effort GPT-6 Astra scored 30.3% on the same benchmark, below Sol’s result, despite costing 3.9 times more per task.

On Agents’ Last Exam, GPT-6 Sol at maximum effort reached 56.4%, edging past Opus 5’s best recorded score at what OpenAI describes as 60% lower cost per task.

Coding and Computer-Use Performance

According to OpenAI, GPT-6 Sol achieved 68.8% on the DeepSWE v1.1 coding benchmark when pushed to maximum effort, falling 1.1 points short of Claude Fable 5’s top-setting result yet doing so at roughly 80% less cost per task. GPT-6 Luna, tested on the same benchmark, hit 66.6%, a figure OpenAI describes as being on par with Opus 5 and Fable 5 running at medium effort, despite costing about 93% less per task than Opus 5 and 96% less than Fable 5.

When evaluated on FrontierCode 1.1 Main, a benchmark assessing whether generated code is genuinely merge-ready, Sol equaled Claude Fable 5.1’s top-effort performance for a substantially lower price. OpenAI’s internal testing also shows Sol scoring 60.5% on the OSWorld 2.0 computer-use benchmark, edging out Opus 5’s medium-effort result of 60.3%, while costing roughly 80% less per task.

Worth flagging: these comparisons benchmark Sol against Claude Opus 5, not Anthropic’s newer Opus 5.5, which launched hours before OpenAI’s own release and which Anthropic says runs about 40% cheaper than Opus 5 on typical workloads. There isn’t yet a public, same-harness test pitting Sol directly against Opus 5.5.

Factuality and Reliability Gains

Beyond raw benchmark scores, OpenAI is also making an accuracy argument. Using an internal evaluation built from de-identified ChatGPT conversations where users had previously flagged factual errors, the company says GPT-6 Sol makes about half as many mistakes as GPT-5.6 Sol, describing it as approaching Astra-level reliability at a fraction of the cost. Luna shows a different kind of gain: at higher effort settings, OpenAI says it can match GPT-5.6 Sol’s factuality at roughly one-hundredth of that model’s task cost. OpenAI notes the evaluation deliberately selects error-prone conversations and isn’t representative of typical everyday usage.

Prompt Caching Cuts Costs Further

Token pricing tells only part of the cost story for agents that run long, repeated conversations. Coding and business agents routinely resend the same system prompts, files, tool definitions and conversation history on every turn, and OpenAI says GPT-6 improves default cache hit rates while offering discounts of up to 90% on cached input-token reads. Cached prefixes reused within a 30-minute window now qualify for these savings.

New developer tools accompany the update, including a Prompt Caching Dashboard to track hit rates over time, a diagnostics tool that explains cache misses, explicit breakpoints to control where a cached prefix ends, and the ability to change reasoning effort mid-conversation without invalidating the cache. GitHub reported that these caching improvements reduced the share of prompt tokens requiring fresh processing by more than 50%, helping Copilot respond faster.

Where GPT-6 Sol and Luna Are Available

Both GPT-6 Sol and GPT-6 Luna are live now in the OpenAI API under the names gpt-6-sol and gpt-6-luna. They’re also accessible through ChatGPT Work and Codex for Plus, Pro, Business, Enterprise and Edu users. Luna has additionally reached the ChatGPT desktop app for Free and Go users. Neither model has rolled out to standard ChatGPT Chat mode yet — OpenAI describes that rollout as gradual through launch day.

A Pricing War Across the AI Industry

The timing of OpenAI’s release wasn’t accidental, given the broader market moment. Anthropic unveiled Claude Opus 5.5 on the same day, pricing it at $4 input and $20 output per million tokens and describing it as roughly 40% cheaper to run than Opus 5 on typical workloads, according to CNBC. Anthropic’s Dianne Penn, head of product management, research and labs, told CNBC the company is focused on making its models’ reasoning more token-efficient depending on effort settings.

Both releases landed just weeks after Anthropic CEO Dario Amodei called for an industrywide slowdown on advanced AI development, a debate that intensified after a former Anthropic researcher publicly resigned, warning that leading labs were “gambling with our lives,” CNBC reported. OpenAI CEO Sam Altman and Tesla and SpaceX CEO Elon Musk also joined that call, per CNBC’s reporting.

Yet the cost-cutting pressure isn’t coming only from within the frontier-lab rivalry. VentureBeat’s pricing comparison shows Luna’s $0.10/$0.50 rate sitting close to open-weight competitors like Xiaomi’s MiMo-V2.6-Flash at $0.14/$0.28, while options from DeepSeek and MiniMax are cheaper still. Chinese firms including Alibaba, Moonshot AI and DeepSeek continue offering low-cost, open-weight alternatives that both OpenAI and Anthropic must now price against, CNBC noted. That dynamic reframes GPT-6 Sol’s value proposition: its case rests less on being the absolute cheapest option on the market and more on whether its task-level reliability and coding performance justify a premium over increasingly capable open-weight rivals.

What happens next in ChatGPT’s standard Chat mode remains the open question. OpenAI has confirmed the rollout there is gradual, leaving millions of everyday users waiting to see whether Sol and Luna’s cost advantages eventually reach the interface most people actually use.

Article produced with the assistance of artificial intelligence and reviewed by the editorial team.