GPT-6 Sol and Luna Arrive: 50% Cheaper API Pricing, Up to 80% Lower Coding Costs at Frontier Performance

Hey everyone, it's shiichan! I've got great news from OpenAI News today, the GPT-6 family just got two new members!

OpenAI News openai.com

What was announced?

According to OpenAI News, following this month's launch of the flagship model GPT-6 Astra, OpenAI has added two new models: GPT-6 Sol and GPT-6 Luna.

Sol and Luna were trained using methods similar to Astra, carrying over Astra's strengths in professional work, factuality, coding, computer use, and alignment into faster, more affordable models.

The biggest piece of news is the pricing update. Compared to the GPT-5.6 generation, API prices dropped like this (prices are per 1 million tokens):

• GPT-6 Sol: input from $4 to $2, output from $20 to $10, both 50% cheaper
• GPT-6 Luna: input from $0.20 to $0.10, output from $1.20 to $0.50, also 50% cheaper

GPT-6 Astra, meanwhile, remains OpenAI's best model overall.

Why it matters

OpenAI shared that its internal coding-agent usage has grown exponentially. Valued at API prices, daily token usage has exceeded $600 for the median researcher and $7,000 for researchers at the 90th percentile. As agents take on longer and more demanding tasks, the cost of sustained use matters more and more.

That's exactly the gap Sol and Luna are meant to fill. By keeping much of Astra's capability while cutting cost significantly, this launch is about letting people call a smart model as often as they need without worrying about the bill. Within the same GPT-6 family, we now have three tiers: the top-performing Astra, the balanced Sol, and the lightweight, low-cost Luna.

What changes

This isn't just a price cut, the underlying performance improved too. A few benchmark results from the announcement:

• On AutomationBench, a test of business workflows, GPT-6 Sol at xhigh effort outperforms Claude Opus 5 at max effort while costing just 9% of Opus 5's cost per task
• GPT-6 Luna at high effort improves 5.4 percentage points over its predecessor at 58% lower cost per task
• On Agents' Last Exam, which evaluates complex professional workflows, GPT-6 Sol at max effort scores 56.4%, beating Claude Opus 5's highest score at 60% lower cost per task
• On OpenAI's internal factuality evaluation, GPT-6 Sol makes about half as many mistakes as its predecessor, approaching Astra-level reliability at a much lower cost

For developers and everyday users, this means you can call a capable model far more often on a much smaller budget. For workloads like coding agents that fire off tons of requests, that difference really adds up.

Dive Deep

The coding benchmarks are worth a closer look. On DeepSWE v1.1, which tests real-codebase software-engineering tasks, GPT-6 Sol at max effort scores 68.8%, within 1.1 percentage points of Claude Fable 5's best score of 69.9% at xhigh effort, at roughly 80% lower cost per task. GPT-6 Luna at max effort scores 66.6%, comparable to Claude Opus 5 and Fable 5 at medium effort, while costing 93% less than Opus 5 and 96% less than Fable 5 per task, according to the announcement.

On OSWorld 2.0's offline set, which measures computer-use tasks, GPT-6 Sol at xhigh effort scores 60.5%, close to Claude Opus 5's 60.3% at medium effort, at roughly 80% lower cost per task. GPT-6 Astra still holds the title of the best model for computer use overall.

There's also a nice caching upgrade. Improved prompt caching now delivers higher cache hit rates by default, with a 90% discount on cached input-token reads. Developers get a Prompt Caching Dashboard, a diagnostics tool, the ability to adjust reasoning effort and tool availability without breaking cache, and explicit breakpoints to control which prefixes get cached. GitHub reported that these improvements reduced the share of prompt tokens requiring fresh processing by more than 50% across billions of requests, helping Copilot respond faster.

On alignment, Sol and Luna both improve over their GPT-5.6 counterparts, including lower rates of misleading claims about their own coding work.

As for availability, GPT-6 Sol and Luna are live today in ChatGPT Work and Codex for Plus, Pro, Business, Enterprise, and Edu users. Free and Go users can access GPT-6 Luna in the desktop app. They're not yet available in Chat. In the API, you can call them as gpt-6-sol and gpt-6-luna, with a gradual rollout throughout the day.

Wrap-up

• OpenAI followed up GPT-6 Astra with two cost-efficient models, GPT-6 Sol and GPT-6 Luna
• API pricing for both dropped 50% versus GPT-5.6 (Sol: $2 input / $10 output, Luna: $0.10 input / $0.50 output, per 1 million tokens)
• On professional-work, coding, and computer-use benchmarks, both models match Claude Opus 5 and Fable 5 scores at dramatically lower cost per task
• Improved prompt caching gives a 90% discount on cached input reads, and GitHub already reports faster Copilot responses from it
• Available now in ChatGPT Work, Codex, and the API (gpt-6-sol / gpt-6-luna), with Free and Go users getting Luna on desktop

This one's for engineers who want to run a smart model constantly without watching the bill, and for teams ready to let their coding agents do a lot more, I think!