Updated July 2026. This post originally tracked our own API bills falling a hundredfold between 2023 and 2025. The trend has not slowed, so we’ve refreshed it with the current AI API pricing landscape — and what those numbers unlock for businesses that aren’t tech companies.
The downward shift in AI API prices isn’t a minor discount — it’s a game changer. What was prohibitively expensive is now within reach for businesses of every size, and business models that were abandoned as too costly are now routine.
Our own receipts: $60 to $0.60 per million tokens
When InfiniteUp first shipped AI features in spring 2023, we paid OpenAI’s GPT-4 rate of $60 per million tokens. It felt fair at the time. By late 2024 GPT-4 Turbo had halved that to $30. Then we moved our DOER Business agents to a small model at $0.60 per million tokens — one hundred times cheaper than where we started, eighteen months earlier, with better speed and a more natural voice.
If your daily coffee budget suddenly bought a hundred cups instead of one, you would not just drink more coffee — you would invent new things to do with coffee. That is what has happened to AI product design.
What AI APIs cost in July 2026
Prices per million tokens, standard list rates as of July 2026. These change often — treat this as a snapshot and confirm on the providers’ pricing pages before budgeting.
| Model | Input / 1M | Output / 1M | Tier |
|---|---|---|---|
| OpenAI GPT-5.6 Sol | $5.00 | $30.00 | Frontier |
| Anthropic Claude Opus 5 | $5.00 | $25.00 | Frontier |
| Anthropic Claude Sonnet 5 | $3.00 | $15.00 | Mid |
| OpenAI GPT-5.6 Terra | $2.50 | $15.00 | Mid |
| Google Gemini 3.6 Flash | $1.50 | $7.50 | Mid |
| Anthropic Claude Haiku 4.5 | $1.00 | $5.00 | Small |
| OpenAI GPT-5.6 Luna | $1.00 | $6.00 | Small |
| DeepSeek V3 | $0.27 | $1.10 | Small |
| Google Gemini 2.5 Flash-Lite | $0.10 | $0.40 | Small |
Read that table against the 2023 baseline and the pattern is stark. Today’s frontier models — dramatically more capable than GPT-4 ever was — cost a twelfth of what GPT-4 did. And yesterday’s frontier capability is now available at commodity prices: a DeepSeek or Flash-Lite call costs so little that the API bill stops being a line item worth discussing for most small-business workloads.
The sticker prices are also only the ceiling. Every major provider now offers batch processing at roughly half price for anything that can wait an hour, and prompt caching that cuts the cost of repeated context by about 90%. A well-engineered AI feature in 2026 routinely runs at a quarter of its naive cost.
Why cheap AI APIs change what gets built
Unit economics are destiny for product design. At $60 per million tokens, an AI feature had to be the product — something users paid for directly, invoked deliberately, and used sparingly. At fifty cents, AI can be plumbing: invisible, always on, running on every interaction without anyone doing arithmetic about it.
- Voice as a first-class interface. Speech-to-text plus a language model plus text-to-speech now costs pennies per interaction. A tradesperson dictating inventory into a phone on the walk to the truck — the kind of system we described in our Main Street post — was a research budget in 2016 and is an API call today.
- AI for businesses that were never worth serving. Custom software with AI inside used to need enterprise budgets to recoup model costs. Now an eight-person firm’s entire monthly AI usage can cost less than one staff lunch, which makes custom business software for small companies economically sane for the first time.
- Agents that check their own work. When a call costs a hundredth of what it did, you can afford to spend five calls on one task: draft, critique, verify, retry, confirm. Reliability patterns that were cost-prohibitive are now standard practice.
- Free tiers that are actually free. Products can absorb AI costs for unpaid users to drive growth — impossible when each active user cost dollars a month in inference.
How we choose a model tier (and when cheapest is wrong)
The mistake we see most is treating the price table as a leaderboard and picking one model for everything. The pattern that actually works is routing: match each job in your product to the cheapest tier that does it well.
- Small models for classification, extraction, formatting, and short conversational turns — the bulk of most products’ volume.
- Mid-tier models for the main user-facing intelligence: summaries, drafting, multi-step reasoning with guardrails.
- Frontier models only where the hardest thinking happens — complex analysis, code generation, the one feature your product’s reputation rides on. One expensive call in a pipeline of cheap ones barely moves the bill.
And the counterweight: the cheapest model that almost works is the most expensive choice you can make. A model one tier too small produces failures your team pays for in support tickets, retries, and lost trust. We price models against the cost of being wrong, not against each other.
The lesson from three years of falling prices
Every architecture decision should assume today’s prices are the most you will ever pay. The feature that is marginal at today’s rates will be comfortable in a year — so build it now and let the curve catch up. That is not optimism; it is the only consistent pattern this industry has produced: 2023’s hundred-dollar workload is 2026’s rounding error, and there is no sign the curve is done.
Wondering what an AI feature would cost to run in your product — or whether your business has a workload this economics unlocks? A free 30-minute call is the fastest way to find out, or see how we build custom AI systems for operating businesses.