GPT-6 Pricing Got 50% Cheaper. DeepSeek Got 4x Pricier. What AI Founders Should Charge

OpenAI released GPT-6 Sol and Luna on September 22, below the flagship GPT-6 Astra it launched earlier this month. Sol is for complex work like coding and debugging. Luna is for high-volume, clearly defined jobs such as summarizing, extracting data and answering quick questions. OpenAI says Sol makes about half as many mistakes as GPT-5.6 Sol and credits better inference and caching for the lower prices.

The Hacker News launch thread passed 1,700 points. The tone was mostly "how do they make money on this?" One commenter called Luna "at the pareto for most tasks." Another posted a token-by-token comparison in which the same monthly workload cost about $40 on DeepSeek and $184 on Luna. That is one anecdote, but it shows that a price per token and a monthly bill are different numbers.

On August 16, DeepSeek introduced peak and off-peak pricing that raised V4 rates sharply. Its stated reason was "to allocate computing resources more reasonably." Peak hours are 01:00–04:00 and 06:00–10:00 UTC. Off-peak rates are half of peak, which is still above the old flat price.

Revenue rose after the hike. According to The Information, CEO Liang Wenfeng told investors the annualized run rate is now above $1 billion, up from under $500 million a few months ago. He credited the 2.3x–4.5x price increases and demand that held up after them. Reports say the hikes did not cause customer churn. DeepSeek is now seeking about $7.5 billion in new funding.

Some caution: these are company numbers relayed through investor reporting, not audited figures. DeepSeek also had a lot of room to raise prices. Even at peak, V4-Pro costs less than GPT-6 Sol.

Take a simple example. A document-summary SaaS on a $49/month plan uses 30 million input tokens and 3 million output tokens per customer each month, with no caching. At list prices, the model cost per customer looks like this:

In five weeks, the cheapest option for this workload changed places. Before August 16, DeepSeek Flash was about half the price of GPT-5.6 Luna. Today GPT-6 Luna is the cheapest, and DeepSeek Flash at peak costs almost 4x as much. A founder who set prices around either provider's rate card saw model cost swing between 9% and 35% of revenue without changing anything in the product.

That is the real point of this week. Model prices are volatile inputs, not a basis for your prices. OpenAI can cut because it has scale and is competing for share. DeepSeek could hike because customers had built on it and found real value there. Neither decision had anything to do with what your customers are willing to pay you.

1. Don't pass the cut straight to customers. A lower model bill is a margin gain. Keep it, or put it into quality: more retries, a stronger model on hard cases, and better review. Cutting your own price starts a race you can't win against a lab that just cut 50%.
2. Check whether you're underpriced. DeepSeek raised prices up to 4.5x and didn't lose customers because the product was worth more than it charged. If your churn is low and customers depend on the output, test a higher price on new signups before assuming they are price-sensitive.
3. Keep the model swappable. Put providers behind one interface, keep a short evaluation set of real tasks, and re-run it whenever a rate card changes. This week proved that the cheapest provider can change within a month.
4. Move batch work to cheaper time windows. If you use DeepSeek, run nightly jobs, backfills and bulk extraction off-peak at half the peak rate. Use cached reads wherever prompts repeat. OpenAI discounts them 90%.
5. Charge for outcomes, not tokens. If your plan is a resold token meter, every lab price change moves your price too. Charging per report, per resolved ticket or per workflow gives you a price customers understand and a margin you control. Our guide to outcome-based pricing covers the models that work.

Before you switch defaults, measure cost per successful task on your own workload. Our Opus 5.5 cost-per-task test walks through the method, and it applies to Sol and Luna as well.

Expect more pressure from both directions. Anthropic says Sonnet 5.5 and Haiku 5.5 are coming, Google is post-training Gemini 4, and OpenAI clearly wants Luna to be the default for high-volume work. At the same time, labs that spend heavily on training, like DeepSeek with over 70% of its compute on new models, need revenue and have shown they will raise prices to get it.

The likely result is that low-end prices keep falling while peak-hour pricing, capacity pricing and tier changes become more common. Founders who treat pricing as an infrastructure decision will absorb these shocks. Founders who resell tokens at a fixed markup will keep repricing every month.