TL;DR: On June 10, 2025, OpenAI dropped o3 pricing by 80%—from $10/$40 to $2/$8 per million tokens. Cursor and Windsurf adjusted their credit systems the same day. Six months later, most AI startups still haven't recalculated their margins or adjusted their pricing. If you're running o3 workloads, you're either sitting on unexpected margin or missing a competitive opportunity. Here's how to figure out which.
What Changed
OpenAI announced an 80% price reduction on their o3 reasoning model:
| Metric | Before (Jan 2025) | After (June 2025) | Change |
|---|---|---|---|
| Input tokens | $10 per million | $2 per million | -80% |
| Output tokens | $40 per million | $8 per million | -80% |
| Cached inputs | $5 per million | $0.50 per million | -90% |
Sam Altman's announcement was direct: "We dropped the price of o3 by 80%!! Excited to see what people will do with it now."
What they did was optimize their inference stack. What you should do is recalculate your unit economics.
The Market Moved Immediately
The companies paying attention adjusted the same day:
Cursor now counts one o3 request the same as a GPT-4o call—previously it cost multiple credits.
Windsurf lowered their "o3-reasoning" tier to a single credit.
These aren't small players experimenting. They're well-funded AI coding assistants with millions of users and tight margins. When costs drop 80%, they move fast because cost structure visibility is part of their operations.
The question is: did you?
Why Most Startups Didn't Adjust
We've talked to dozens of AI founders since June. Here's what we hear:
"We don't use o3." Fair—but did you evaluate whether you should now that it's 80% cheaper? Tasks that were cost-prohibitive at $40/million output tokens might make sense at $8.
"Our pricing is locked in." Locked in by who? If your costs dropped and you're pocketing the margin, great. If you're still pricing based on old cost assumptions, you might be uncompetitive.
"We don't know our per-model costs." This is the real answer for most teams. They know their aggregate OpenAI bill went down, but they can't attribute it to specific features, customers, or workflows.
If you can't answer "what does o3 cost us per customer per month," you can't make informed decisions when that cost changes.
How to Recalculate Your Margins
If you're using o3 in production, here's the exercise:
Step 1: Get Your Usage Data
Pull your o3 consumption for the last 30 days:
- Total input tokens
- Total output tokens
- Cached input tokens (if applicable)
If you can break this down by customer or feature, even better.
Step 2: Calculate Old vs New Cost
Example: A document analysis feature
| Metric | Monthly Usage | Old Price | Old Cost | New Price | New Cost |
|---|---|---|---|---|---|
| Input tokens | 50M | $10/M | $500 | $2/M | $100 |
| Output tokens | 10M | $40/M | $400 | $8/M | $80 |
| Total | $900 | $180 |
That's a $720/month savings on a single feature. Across your entire o3 usage, the numbers compound fast.
Step 3: Calculate Margin Impact
If you were charging customers $1,500/month for this feature:
- Old margin: $1,500 - $900 = $600 (40%)
- New margin: $1,500 - $180 = $1,320 (88%)
You went from a healthy margin to an exceptional one—without changing anything.
But here's the uncomfortable question: should you keep it, or should you do something with it?
What to Do With the Savings
You have four options. Each makes sense in different situations.
Option 1: Pocket the Margin
When it makes sense:
- You were underwater or barely profitable before
- You need runway and can't afford to reduce prices
- Your competitive position is strong enough that customers won't leave
The risk: Competitors who pass savings to customers will look more attractive. If o3 is commoditizing, margin won't last.
Option 2: Lower Prices
When it makes sense:
- You're in a competitive market where price matters
- You want to accelerate adoption
- Your cost structure was previously limiting growth
The risk: You can't easily raise prices later. Make sure the volume increase justifies the per-unit revenue decrease.
Option 3: Add Features or Increase Limits
When it makes sense:
- Customers are hitting usage caps
- You have feature ideas that were cost-prohibitive
- You want to increase perceived value without touching price
Example: Your "Pro" plan had 100 o3 queries/month. Now you can offer 500 at the same cost to you. Customers feel they're getting more value; you maintain price integrity.
Option 4: Hybrid Approach
When it makes sense: Almost always.
Most sophisticated companies do a mix:
- Keep some margin improvement (strengthen the business)
- Pass some savings to customers (stay competitive)
- Invest some in features (increase value)
The ratio depends on your competitive position, runway, and growth goals.
The Bigger Picture: Price Volatility Is the New Normal
This won't be the last price change. In the past 12 months:
- OpenAI dropped o3 by 80%
- Anthropic launched Opus 4.5 at $5/$25 (making Opus-tier reasoning accessible)
- Google bundled more free features into Gemini
- DeepSeek released R1 at 15% of GPT-4 cost
The AI cost curve is collapsing faster than anyone predicted. Dr. Alexander Wissner-Gross captured it well: "The cost of superintelligence is crashing."
If your pricing model assumes stable costs, you're building on sand. Every few months, the ground shifts.
The Real Question: Do You Know Your Margins?
The o3 price cut is a forcing function. It reveals whether you actually understand your cost structure.
Can you answer these questions?
- What's your o3 spend per customer?
- Which features use o3 vs cheaper models?
- What's your margin per feature, not just per product?
- If costs dropped 50% tomorrow, would you know within a day or a quarter?
If you can't, you're not alone. Most AI startups track aggregate API spend, not per-customer or per-feature costs. They know their OpenAI bill; they don't know their unit economics.
We wrote a complete guide to this problem: Unit Economics for AI Products: A Complete Cost Framework. The core insight is that tokens aren't your unit—traces are. A complete workflow execution with all its LLM calls, tool costs, and reliability overhead.
Until you think in traces, price changes like this are surprises instead of opportunities.
What to Do This Week
- Pull your o3 usage from the OpenAI dashboard (or your observability tool)
- Calculate your cost reduction since June using the old vs new prices
- Decide what to do with the savings—pocket, pass through, or reinvest
- Set up per-feature cost tracking so the next price change isn't a research project
If you want help with step 4, that's what we built Bear Billing for. Per-customer margins, feature-level cost attribution, and pricing scenario modeling—so you can see the impact before you make changes.
The companies that thrive in AI aren't the ones with the best models. They're the ones who understand their economics well enough to move fast when the landscape shifts.
o3 dropped 80%. What did you do about it?