OpenAI drops GPT-6 Sol and Luna with 50% price cuts

If you have been running coding agents lately, your cloud bill probably hurts.

OpenAI gave us a glimpse into its own labs recently: its median researcher consumes over $600 worth of tokens every day, while the top 10% burn through more than $7,000 daily. Cutting-edge intelligence is amazing, but scaling it across continuous workflows is bankrupting teams.

OpenAI tackled that wall head-on, rolling out two new additions to its frontier family: GPT-6 Sol and GPT-6 Luna.

Halving the price tag and the errors

While the flagship GPT-6 Astra introduced this generation of intelligence earlier this month, it remains an expensive heavyweight. Sol and Luna take the exact same core training improvements and package them into faster, dramatically cheaper tiers.

Both models drop API pricing by 50% compared to their GPT-5.6 predecessors. Sol sits at $2 per million input tokens and $10 for output, while the lightweight Luna costs just $0.10 for input and $0.50 for output.

OpenAI also overhauled its prompt caching system, offering a 90% discount on cached input reads. GitHub reported that these enhancements have already cut newly processed prompt tokens by more than 50% across billions of Copilot queries.

More importantly, OpenAI claims Sol makes about half as many factual mistakes as its predecessor in internal evaluations, nearing Astra-grade reliability without the price tag. Luna, when given higher reasoning effort, matches GPT-5.6 Sol's factuality at roughly one-hundredth of the cost.

Squeezing Anthropic on cost per task

Just 90 minutes before OpenAI published the release, Anthropic pushed out an updated Opus 5.5. The rivalry could not be more explicit, and OpenAI spent much of its benchmark deck aiming straight at Claude.

On AutomationBench, which grades agents navigating multi-app business workflows, GPT-6 Sol beat Claude Opus 5 while chewing up only 9% of Opus 5's task cost.

On the software engineering benchmark DeepSWE v1.1, max-effort Sol reached 68.8%, landing within 1.1 percentage points of Claude Fable 5 while saving roughly 80% on cost per task.

Luna held its own too, matching mid-effort Opus 5 and Fable 5 scores on DeepSWE while running 93% to 96% cheaper per task.

Both models are now rolling out in ChatGPT Work and Codex for paid tiers, and Amazon Bedrock announced general availability on day one. Free and Go users can try Luna in the desktop app.

Why it matters

We think this is one of the smartest tactical moves OpenAI has made all year. Instead of flexing unreplicable ceiling intelligence, it is attacking the operational cost curve that keeps businesses from actually deploying autonomous agents.

For developers and teams building multi-step automations, Sol changes the math. Agentic workflows require loops, retries, and constant context re-reads. When token costs drop by half and caching gives you a 90% break, you can afford to let an agent iterate three times to get the right answer.

For the industry, it forces Anthropic into a brutal margin squeeze. Matching performance on coding benchmarks is one thing, but undercutting Claude by 80% to 90% on task-level costs creates immediate procurement pressure for enterprise buyers.

The catch is that Sol is still not Astra. For the highest-stakes architecture decisions or deep multi-hour research, the flagship model remains unmatched. But for the daily grind of debugging, document processing, and customer routing, Sol and Luna make frontier AI economical enough to leave running in the background.