OpenAI just made the biggest pricing move of 2026. On July 30, the company cut API prices for GPT-5.6 Luna by a massive 80% — from $1.00 down to just $0.20 per million input tokens — while GPT-5.6 Terra dropped 20% to $2.00 per million. If you build with AI, pay for AI subscriptions, or run a business on AI APIs, this changes your math. Here is exactly what changed, how the new prices compare to Claude and Gemini, and what you should do about it.
What Exactly Changed on July 30, 2026
OpenAI adjusted pricing across its GPT-5.6 family, which comes in three tiers:
| Model | Old Input / Output | New Input / Output | Change |
|---|---|---|---|
| GPT-5.6 Sol (flagship) | $5.00 / $30.00 | $5.00 / $30.00 | No change |
| GPT-5.6 Terra (mid-tier) | $2.50 / $15.00 | $2.00 / $12.00 | -20% |
| GPT-5.6 Luna (fast) | $1.00 / $6.00 | $0.20 / $1.20 | -80% |
All prices are per million tokens. The headline number is Luna: at $0.20 per million input tokens, OpenAI's fastest model now costs less than one-tenth of what GPT-4 cost at launch in 2023 — for a model that is dramatically more capable.
The timing was not random. This came days after OpenAI moved GPT-5.6 Luna to the default free-tier model in ChatGPT and made free text chats effectively unlimited, and it landed in the middle of an intensifying price war with Google and Anthropic.
Why OpenAI Is Doing This
Three forces are driving prices down across the industry:
1. Inference costs are collapsing. Newer chips like Nvidia's Vera Rubin generation and custom silicon deals (including Anthropic's reported $10B compute agreement with Volta) have pushed the cost of serving each token down sharply. Vendors can cut prices and keep margins.
2. The agent economy needs cheap tokens. AI agents burn tokens at a rate humans never did — a single agentic coding session can consume millions of tokens. Companies like Rippling learned this the hard way, reportedly burning 40% of an R&D budget on tokens. For agents to be economically viable at scale, per-token prices had to fall. OpenAI wants agent builders locked into its stack.
3. Competition from cheap challengers. Meta's Muse Code launched at $0.10 per million input tokens in its Contributor tier, undercutting everyone. Open-weight models running locally cost nothing per token. OpenAI's 80% cut on Luna is a direct response: it removes "price" as a reason to leave.
How the New Prices Compare (August 2026)
Here is the current landscape for the models most professionals actually use:
| Model | Input / Output (per 1M tokens) | Best For |
|---|---|---|
| GPT-5.6 Luna | $0.20 / $1.20 | High-volume, speed, agents |
| Claude Haiku 4.5 | $1.00 / $5.00 | Fast tasks, light coding |
| Gemini 3.1 Flash | ~$0.30 / ~$1.50 | Multimodal at volume |
| GPT-5.6 Terra | $2.00 / $12.00 | Balanced production work |
| Claude Sonnet 5 | $2.00 / $10.00 (intro until Aug 31) | Coding, analysis |
| Gemini 3.1 Pro | $2.00 / $12.00 | Long context, multimodal |
| GPT-5.6 Sol | $5.00 / $30.00 | Hardest reasoning tasks |
| Claude Opus 5 | $5.00 / $25.00 | Frontier coding, agents |
Two things stand out. First, Luna is now the cheapest frontier-lab model on the market by a wide margin. Second, note the Claude Sonnet 5 fine print: its $2.00 introductory input price ends August 31, 2026, after which it rises to $3.00 — so the effective gap between OpenAI and Anthropic mid-tiers is about to widen further.
What This Means If You Use ChatGPT (Not the API)
You might think API pricing does not affect you. It does:
- Free tier got real. Unlimited text chats on GPT-5.6 Luna means the free ChatGPT is now genuinely capable for everyday tasks — drafting, summarizing, brainstorming, basic coding help.
- The Plus question. At $20/month, ChatGPT Plus now has to justify itself through Sol access, higher limits on advanced features, and tools like agent mode — not through basic chat quality. If you only use text chat, the free tier may honestly be enough now.
- Expect cheaper AI inside your apps. Every SaaS product that embeds AI — email tools, CRMs, writing assistants — just saw its costs drop. Watch for either better free AI features or price cuts in tools you already pay for. If your vendors keep prices flat while their costs fell 80%, that is worth asking about at renewal time.
What This Means If You Build With the API
Re-route your traffic. The classic pattern — route easy tasks to a cheap model, hard tasks to a flagship — just got more attractive. At $0.20/$1.20, Luna can handle classification, extraction, summarization, and most chat workloads for pennies. A workload that cost $1,000/month on Luna in June costs about $200 now.
Recalculate your unit economics. If you priced your product when input tokens cost $1.00, your margins just improved — or your competitors' prices are about to drop. Model costs are now moving fast enough that quarterly pricing reviews are a business necessity.
Do not sleep on caching. OpenAI's cached-input discounts stack with the new prices, and for agent workloads with long repeated system prompts, effective input costs can fall well below the headline $0.20.
Benchmark before you switch. Cheap tokens are worthless if quality drops for your use case. Run your own eval set. In our testing across article drafting and data-extraction tasks, Luna holds up remarkably well against mid-tier models for structured work, but still trails Terra and Sonnet 5 on multi-step reasoning.
The Bigger Picture: Intelligence Is Getting Too Cheap to Meter
Zoom out and the trend is unmistakable. Frontier-quality AI that cost $30+ per million input tokens in 2023 now starts at $0.20. That is a 99%+ decline in three years — faster than any comparable technology cost curve, including early cloud computing.
Reddit skeptics reasonably ask whether anyone makes money at these prices, and it is true that price wars compress margins. But the strategic logic is clear: whoever owns the developer relationship when agents go mainstream wins the next decade. Cheap tokens are customer acquisition.
For everyone else, the takeaway is simpler: the cost of adding intelligence to any product, workflow, or side hustle just collapsed again. The bottleneck is no longer the price of AI — it is knowing what to build with it.
Key Takeaways
- GPT-5.6 Luna now costs $0.20/$1.20 per million input/output tokens (down 80%); Terra costs $2.00/$12.00 (down 20%); Sol is unchanged.
- Luna is now the cheapest model from a major frontier lab, undercutting Claude Haiku 4.5 and roughly matching Gemini Flash.
- Claude Sonnet 5's intro pricing ends August 31 — if you are choosing a mid-tier model this month, factor that in.
- API builders should re-route high-volume workloads to Luna and re-check unit economics now.
- ChatGPT free tier is significantly more capable; casual users may no longer need Plus.