OpenAI just made cheap AI features a lot more affordable — and it wants you to believe they’re also more trustworthy. With GPT-6 Sol priced at $2/$10 per million input/output tokens (roughly 50% cheaper than GPT-5.6 Sol) and GPT-6 Luna coming in at $0.10/$0.50, the cost floor for shipping AI-powered features just dropped again. OpenAI is pairing that price cut with claims of roughly half as many factual errors and a deception rate that fell from 10.4% to 1.3% on an internal coding-deception test.
Why it matters
For most product teams, model capability stopped being the bottleneck a while ago — cost, latency, and reliability are what decide whether a feature ships or dies in review. This release moves the needle on all three at once, at least on paper. But OpenAI is also explicitly marketing reliability as a product feature, not just a side effect of scale, which means builders need to verify those claims rather than inherit them.
The numbers that matter for your budget
| Model | Input / Output (per 1M tokens) | Notable claim |
|---|---|---|
| GPT-6 Sol | $2 / $10 | ~50% cheaper than GPT-5.6 Sol; ~half the factual errors |
| GPT-6 Luna | $0.10 / $0.50 | Ultra-low-cost tier for high-volume tasks |
OpenAI also reported GPT-6 Sol scoring 33.2% on AutomationBench 1.0.6 at high effort, at a cost of $0.27 per task — a metric that ties quality directly to task economics rather than raw token price, which is the number builders should actually be tracking.
What this means if you ship product this quarter
- Re-run your cost model. Features you shelved as too expensive at GPT-5.6 pricing — high-volume support triage, document summarization, low-stakes classification — may now clear margin at Sol or Luna pricing. Recalculate cost-per-resolved-ticket or cost-per-completed-task, not just cost-per-token.
- Don’t skip your own evals. “Half as many mistakes” and “1.3% deception” are OpenAI’s internal numbers on OpenAI’s internal tests. Run your own eval suite against your actual prompts and data before swapping models in production — reliability claims from a vendor selling reliability as a feature deserve extra scrutiny, not less.
- Watch AutomationBench-style cost-per-task metrics, not just per-token price. A cheaper token price with more retries or longer outputs can cost more per completed task than a pricier model that gets it right the first time.
- Factor in Anthropic’s new blocked-request pricing if you’re in regulated or high-risk categories. If your workflow touches sensitive content — health, legal, financial, safety-adjacent — check whether rejected requests now carry a cost on the vendor side you’re using, since that changes unit economics for anything with a meaningful refusal rate.
- Segment by tier, not just by model family. Luna’s pricing suggests a real opening for high-volume, low-stakes automation (tagging, routing, first-pass drafts) where occasional errors are cheap to fix, while Sol’s reliability claims target the mid-tier use cases — support responses, internal tools — where mistakes are more expensive to unwind.
The honest counter-argument
It’s worth saying plainly: these are vendor-reported numbers from OpenAI’s own launch materials, not independent third-party audits. A 33.2% score on AutomationBench 1.0.6 sounds precise, but benchmark performance and production performance on your specific data and prompts are often different stories. Price cuts are easy to verify; deception and reliability claims are not — and a company that just launched two new SKUs has every incentive to frame its own model favorably. Treat the pricing as fact and the reliability claims as a hypothesis to test.
Bottom line
The API pricing trend is now firmly downward and shows no sign of stopping, which shifts competitive pressure toward margin, latency, and task-level economics rather than raw capability. If you’re planning Q1 or Q2 roadmap decisions, this is a good moment to re-price shelved AI features, build (or refresh) an eval set tied to cost-per-task, and treat every reliability claim from a model vendor as something to verify against your own data before it shapes your shipping decision.