On July 30, 2026, OpenAI cut the price of GPT-5.6 Luna by 80% and GPT-5.6 Terra by 20%. Three weeks earlier, these models launched at $1 and $2.50 per million input tokens respectively. Luna now costs $0.20 per million input tokens and $1.20 per million output tokens. Terra dropped to $2 per million input and $12 per million output. Sol, the flagship, kept its $5/$30 pricing unchanged.

If you blinked, you missed the window where Luna cost a dollar. The model that was already the cheapest option in OpenAI's lineup just got five times cheaper in under a month.

The numbers that matter

Here is the before-and-after for API pricing, straight from OpenAI's announcement:

Luna: $1.00 → $0.20 per million input tokens. $6.00 → $1.20 per million output tokens. That is an 80% cut on both sides.

Terra: $2.50 → $2.00 per million input tokens. $15.00 → $12.00 per million output tokens. A 20% cut.

Sol: $5.00/$30.00. No change.

The Luna numbers are the story. At $0.20 per million input tokens, Luna is now cheaper than DeepSeek V4 Flash's $0.14 per million — wait, no, DeepSeek is still cheaper on raw token price. But Luna outperforms Claude Opus 4.8 on coding benchmarks and costs roughly one-sixth the price of comparable models from Anthropic. The gap between "cheap model" and "capable model" just collapsed.

For context, a year ago you were paying $5 per million input tokens for GPT-4-class intelligence. Luna delivers comparable or better performance at 25 times less cost. That is not a gradual price reduction. That is a reset.

Why three weeks?

OpenAI's blog post frames this as passing efficiency gains on to customers. The GPT-5.6 architecture is genuinely more efficient to serve — Sol helped rewrite production kernels that cut serving costs by 20%, and token-generation efficiency improved by more than 15%. Those are real engineering wins.

But the timing is not accidental. DeepSeek V4 Flash shipped on July 31 at $0.14/$0.28 per million tokens and immediately started scoring higher than models costing 50x more. Google's Gemini 2.5 Flash is free for low-volume usage. The market shifted under OpenAI's feet while they were still celebrating the GPT-5.6 launch.

OpenAI cut prices on July 30. DeepSeek promoted V4 Flash to general availability on July 31. These announcements were one day apart. Nobody in this market is waiting for quarterly reviews to adjust pricing.

What this changes for builders

If you are running AI-powered workflows — lead classification, customer support triage, content generation, code review, document analysis — the cost equation just changed materially. Here is what the new pricing looks like in practice:

A workflow that processes 10,000 requests per day with an average of 1,000 input tokens and 500 output tokens per request costs roughly $12 per day on Luna. That is $360 per month for a process that would have cost $1,800 per month at launch pricing and $5,000+ per month on GPT-4 a year ago.

The more interesting shift is architectural. At $0.20 per million tokens, the economic case for using a cheaper model for "good enough" tasks gets much stronger. A coding workflow might use Sol to plan and resolve ambiguity, then hand off well-specified implementation tasks to Luna at one-twenty-fifth the token cost. Customer service bots can run Luna for initial triage and escalate to Terra or Sol only when the query is genuinely complex.

This is the multi-model routing pattern that everyone talked about in 2025 but few actually implemented, because the price gap between tiers was not big enough to justify the engineering effort. At these prices, it is.

The Fast mode addition

OpenAI also replaced Priority Processing with a new Fast mode for Sol. Fast mode delivers up to 2.5x faster inference at twice the standard price. For Sol, that means $10 per million input tokens and $60 per million output tokens in Fast mode.

This is not cheap. But for latency-sensitive applications — real-time coding assistants, live customer interactions, interactive agents — the speed premium may be worth it. The old Priority Processing was a blunt instrument. Fast mode is more targeted and backward compatible: existing API requests tagged with "priority" automatically route to Fast mode.

The bigger picture

OpenAI's pricing trajectory is aggressive. GPT-5 launched in August 2025 at $1.25 per million input tokens. Terra, which matches or exceeds GPT-5's capability, now costs $2.00 — but Luna, which outperforms last year's frontier models, costs $0.20. The floor keeps dropping.

The competitive pressure is real and it is coming from multiple directions. DeepSeek's open-weight models are functionally free if you self-host. Google is subsidizing Gemini usage to build market share. Anthropic is holding prices steady while improving quality. OpenAI's move to cut prices this hard, this fast, signals that they are prioritizing volume and developer lock-in over margin.

For anyone building on AI APIs, this is unambiguously good news. The cost of experimentation just dropped by another order of magnitude. A side project that would have cost $500/month in API calls six months ago now costs $100. A production workflow that was borderline viable is now comfortably profitable.

The question is not whether prices will keep falling. They will. The question is whether the quality floor holds. Luna at $0.20 per million tokens is a remarkable value today. If OpenAI keeps cutting, and the next generation of budget models maintains this quality level, the era of "expensive AI" is functionally over for most use cases.

Check your current API spend, model by model. If you are still running Terra or Sol for tasks that Luna handles well, you are leaving money on the table.

Sources: OpenAI — Advancing the price-performance frontier with GPT-5.6, DeepSeek V4 Flash release