What does AI cost in July 2026? A frontier model costs $5 to $10 per million input tokens and $25 to $50 per million output tokens, a strong mid-tier model costs $2 to $3 in and $10 to $15 out, and the cheapest usable models cost well under a dollar in either direction. Those are list prices from Anthropic, OpenAI, Google, and DeepSeek as of this week.
The interesting part is not the numbers. It is the shape. In the last six weeks, Anthropic, OpenAI, and DeepSeek all shipped new models, and the price at the top of the market did not move. The price one step down did. If you run a business and you have been waiting for AI to get cheaper, it did, but not where you were looking.
What did the providers change?
Three announcements matter. Anthropic released Claude Sonnet 5 on June 30 at $2 per million input tokens and $10 per million output tokens, an introductory price that runs through August 31 and then returns to the standard Sonnet rate of $3 and $15. OpenAI released GPT-5.6 on July 9 in three sizes: Sol at $5 in and $30 out, Terra at $2.50 and $15, and Luna at $1 and $6. And Anthropic’s top model, Claude Fable 5, launched in June at $10 and $50, which is double the price of Opus.
Set those next to what the same tiers cost at the start of the year and a pattern appears.
| Tier | Early 2026 | July 2026 | Change |
|---|---|---|---|
| Anthropic premium (Fable 5, June) | none | $10 / $50 | New tier above Opus |
| Anthropic flagship (Opus 4.6 to Opus 4.8) | $5 / $25 | $5 / $25 | Flat |
| Anthropic mid (Sonnet 4.6 to Sonnet 5) | $3 / $15 | $2 / $10 | Down a third |
| OpenAI flagship (GPT-5.5 to GPT-5.6 Sol) | $5 / $30 | $5 / $30 | Flat |
| OpenAI mid (GPT-5.4 to GPT-5.6 Terra) | $2.50 / $15 | $2.50 / $15 | Flat, much smarter |
| OpenAI small (GPT-5.6 Luna) | none | $1 / $6 | New |
| DeepSeek V4 Flash | $0.14 / $0.28 | $0.14 / $0.28 | Flat |
Prices are per million tokens, input then output. Sources: Anthropic pricing, OpenAI, and DeepSeek’s published API rates.
So the flagship price has been $5 in and $25 to $30 out for a year and a half now, across two companies and five model generations. What you get for that $5 has improved with every generation, but the sticker has not budged. Meanwhile the model most businesses should actually be running dropped a third in price and, by Anthropic’s own framing, sits close to the previous flagship in capability.
Why is the top price flat while the middle falls?
Because the top price is not really a price. It is a position. The labs price their flagship at whatever the market will bear for the best available thing, and that number has settled at $5 in, $25 to $30 out, with a premium tier above it for people who need the very edge. When a new flagship ships, it takes the old flagship’s price and the old flagship gets quietly retired.
The middle tier is where competition actually happens. Sonnet 5 at $2 is aimed straight at Terra at $2.50, and Luna at $1 is aimed at Haiku. That is the part of the market where a small business lives, so that is the part where the price war is being fought on your behalf.
One caveat, and it is a real one. Anthropic changed its tokenizer with Sonnet 5, and its own announcement says the same text can now count as anywhere from the same number of tokens to 35% more, depending on content. Code and structured data land near the high end. So a 33% price cut on a model that uses up to 35% more tokens is, for some workloads, a wash. For plain English, which is what most business documents are, the cut is real. Check your own bill rather than the press release.
What does this do to a small business bill?
For most small business workloads, the monthly bill falls somewhere between a little and a lot, and the difference comes down to whether you were paying for the flagship when you did not need it.
Take the example we used in The Business Owner’s Guide to AI Token Economics: a business handling 200 support tickets a day, drafting a response with context for each one. At Sonnet 4.6 rates that was in the range of $4 to $10 a day in tokens. At Sonnet 5 introductory rates it is $3 to $7. Not life-changing, and not nothing across a year.
The bigger lever is the one that has not changed since we wrote The Real Cost of an AI Employee in April. An autonomous agent running all day still costs $20 to $200 in API fees depending on the model, and the price of the model at the top of that range is exactly what it was. If you are running agents on the flagship out of habit, moving the routine 80% of the work to the mid tier is worth more than any price cut the labs will give you.
Ballpark, for a small business with a couple of automations and a few people using AI daily: $50 to $400 a month in API spend, plus subscriptions. The bottom of that range dropped this summer. The top did not.
Which model should a small business use?
Use the mid tier by default, the flagship for the hard 10%, and the small tier for anything that is basically sorting.
That is the whole answer, and it has been the answer for a year. What changed is that the mid tier is now good enough that the hard 10% got smaller. The old rule of thumb was that the mid model was for classification and drafting and the flagship was for anything involving judgment. Sonnet 5 and Terra both handle a lot of judgment now. The flagship is for long, multi-step agent runs where a single wrong turn early compounds, and for the kind of analysis where you would otherwise pay a specialist.
The small tier (Luna, Haiku, DeepSeek Flash) is for routing, tagging, extracting a date from an email, and other jobs where the answer is short and the volume is high. At $0.14 per million tokens, DeepSeek Flash will classify a million emails for about the cost of lunch. It is also a Chinese model with its own terms of service, which matters for some businesses and not for others. Decide that on purpose.
Is the price going to keep falling?
At the top, no, not in the next year, and I would plan on the premium tier above the flagship becoming normal. At the middle and bottom, yes, and the labs have told you so: Sonnet 5’s introductory price has an end date, Gemini and DeepSeek both publish promotional windows, and every one of those windows tends to get extended when a competitor undercuts it.
The practical move is the same one we keep making on this blog. Budget for AI the way you budget for electricity, an ongoing line item that goes up and down, and build your automations so that swapping the model underneath is a one-line change. The company that can move from one provider to another in an afternoon is the company that gets the benefit of every price cut. The one that hard-wired a single model in February is paying February prices.
Frequently Asked Questions
How much does Claude Sonnet 5 cost?
Claude Sonnet 5 costs $2 per million input tokens and $10 per million output tokens at its introductory price, which Anthropic has said runs through August 31, 2026. After that the standard Sonnet price is $3 and $15. Anthropic also updated its tokenizer with Sonnet 5, so some content uses up to 35% more tokens than before.
How much does GPT-5.6 cost?
GPT-5.6 comes in three sizes. Sol, the flagship, costs $5 per million input tokens and $30 per million output tokens. Terra costs $2.50 and $15. Luna costs $1 and $6. All three were released on July 9, 2026.
Has AI gotten cheaper in 2026?
Yes at the middle and bottom of the market, no at the top. Flagship models from Anthropic and OpenAI have held at $5 per million input tokens for over a year, and Anthropic added a premium tier above that at $10. Mid-tier models like Claude Sonnet 5 and GPT-5.6 Terra dropped in price or held price while getting much more capable, and small models now cost $0.14 to $1 per million input tokens.
Which AI model is cheapest for a small business?
For high-volume simple tasks like sorting, tagging, and extraction, the cheapest options in July 2026 are DeepSeek V4 Flash at $0.14 per million input tokens and GPT-5.6 Luna at $1. For general business writing and analysis, Claude Sonnet 5 at $2 per million input tokens is the best value. Reserve flagship models for complex multi-step work.
Why do output tokens cost more than input tokens?
Generating text is more computationally expensive than reading it. Every provider charges 3 to 6 times more for output tokens than for input tokens, so a task that produces a long answer costs more than one that reads a long document and returns a short one.