DeepSeek’s new prices go live at 16:00 UTC this Sunday, and they arrive with something the AI industry has not tried at scale before: peak hours. Run a job during the busy window and you pay one rate. Run the identical job a few hours later and you pay half.
The short version: over the past two weeks the four largest model providers moved their prices in four different directions. OpenAI cut. Anthropic cancelled a planned increase. Google halved the price of its workhorse model, but only until December 31. DeepSeek, for two years the cheapest serious option on the market, is raising some rates by more than 350%. If you are trying to work out what AI software for small business actually costs in 2026, the number matters less than the date printed beside it. Nearly every price on this list has an expiry attached, and most of them moved because a competitor moved first.
What actually changed in the past two weeks?
Four moves, in order.
OpenAI, July 30. The company cut GPT-5.6 Luna from $1.00 to $0.20 per million input tokens, and from $6.00 to $1.20 per million output tokens, an 80% reduction, while trimming the mid-tier Terra by 20% and leaving flagship Sol untouched. OpenAI framed it as passing along efficiency gains, including work the model did helping rewrite its own production inference code. These were permanent list changes with no expiry attached, which as you will see is now the unusual part. VentureBeat covered the cut and the competitive pressure behind it.
Anthropic, August 10. Claude Sonnet 5 launched at an introductory $2 and $10 per million input and output tokens, with a scheduled 50% rise to $3 and $15 on September 1. Anthropic cancelled that increase and made the introductory rate permanent, in an edit dated August 10. This one is worth flagging for a second reason: a number of pricing roundups published after August 10 still list the September increase as though it were coming. If you are budgeting from a comparison article, check when it was written.
Google, August 13. Gemini 3.7 Flash arrived barely three weeks after 3.6 Flash, priced at $0.75 and $3.75 per million tokens, half what its predecessor cost, and aimed squarely at coding and automated business workflows. The catch is in the fine print: that rate is introductory and runs through December 31, 2026. On January 1 it doubles to $1.50 and $7.50.
DeepSeek, August 16. Both V4-Flash and V4-Pro go up sharply. Output tokens on V4-Pro move from $0.87 per million to $3.96 at peak and $1.98 off peak. V4-Flash output moves from $0.28 to $1.32 at peak and $0.66 off peak, according to Fortune’s reporting on the change. DeepSeek said the goal is to allocate resources more reasonably and encourage developers to shift work toward less congested hours. Even after the increase, its rates remain among the lowest available, which is precisely what makes the move interesting.
Why is AI suddenly priced like a utility?
Because it has started to behave like one. Peak and off-peak billing is not a software pricing idea, it is an electricity pricing idea, borrowed later by shipping, airlines and hotels. It shows up whenever a provider owns a fixed amount of capacity, demand arrives in a lumpy daily curve, and building more capacity is slow and expensive. That is an exact description of GPU serving in 2026.
Here is the part almost nobody has drawn out yet. If time-of-day pricing spreads, then for the first time the cost of a piece of AI work depends on when you choose to run it. And most of what a small business actually wants from AI is not urgent. Drafting next month’s newsletters, cleaning up a customer list, summarising a quarter of reviews, generating product descriptions for 400 items, rewriting old service pages: none of that has to happen at 9am on a Tuesday. It happens then only because that is when someone sat down to do it.
Interactive work, the chatbot answering a customer at 11pm, genuinely cannot wait. Batch work can. Under flat pricing those two things cost the same, so the distinction never mattered and nobody built the habit of separating them. Under time-of-day pricing they stop costing the same. The owner who can say “this job is not urgent, run it overnight” gets a discount for a decision that costs nothing. That is a real lever, and it is available to a two-person shop exactly as much as to an enterprise.
What does this mean for AI software for small business?
Directly, for most owners, very little this week. You are almost certainly not buying tokens. You are buying a subscription to something built on top of them, and the two prices are only loosely connected.
That gap is the thing to understand. When the underlying model gets 80% cheaper, your software vendor’s costs fall. Your invoice does not automatically follow. Sometimes the saving is passed on, sometimes it funds features you did not ask for, and sometimes it is simply kept. We have watched all three happen this year: Microsoft folded Copilot into the Microsoft 365 plans small businesses already buy and raised the base plans for everyone else, OpenAI added a $125 Premium seat above its standard business tier, and free ChatGPT text chats became unlimited with the company openly calling the pricing a phase.
The pattern across all of it is that the era of the stable software price is over for this category. Not necessarily higher, which is the fear, and not necessarily lower, which is the sales pitch. Variable. Re-priced on a cadence set by someone else’s competitive position, and increasingly metered rather than flat.
What should you actually do about it?
Four things, none of which require you to become a procurement department.
Find the expiry date. For each AI tool you would genuinely miss, spend ten minutes finding out whether the price you pay is a standard rate or an introductory one, and write the renewal date in the calendar you actually look at. Google’s current Flash pricing has a published end date. Most vendors bury theirs in a footnote. This is the single highest-value habit in this article, and almost no comparison guide gives you the number.
Sort your AI work by urgency, once. Make two short lists: work that must happen while a customer is waiting, and work that could run overnight. You do not have to act on it today. You will want it the first time a tool you use offers a cheaper off-peak tier, and the sorting is easier done calmly than under a billing change.
Keep the inputs portable. Your prompts, your templates, your cleaned customer data and your standard operating procedures are the assets. The model underneath is a commodity that just proved it can reprice overnight in either direction. Anything you can export and carry to another tool protects you from a price move; anything locked inside a vendor’s interface does not. Owners who run smaller models on their own hardware arrive at the same conclusion from a different road.
Do not rebuild anything for a price. Switching costs are real, and they are usually paid in your team’s attention, which is the scarcest thing you have. A 20% saving on a tool that costs $40 a month is $96 a year, and is not worth a fortnight of retraining. Watch the prices. Move only when the gap is large enough to notice without a spreadsheet.
The reassuring read on all of this is that four providers competing hard on price is a good problem for a buyer to have. Three of the four moves in the past fortnight made AI cheaper or held it steady. The fourth came from the provider that had been cheapest by such a wide margin that it could raise rates 350% and remain competitive. That is what a market looks like when it is working, and it is a considerably better position than the one small businesses occupied two years ago, when the question was whether they would be priced out of this technology at all.
Frequently Asked Questions
Does a cheaper API price mean my AI subscription gets cheaper?
Usually not, and rarely straight away. The API price is what your software vendor pays for the model; your subscription is what the vendor charges you for the product wrapped around it. Those move independently, and vendors have several reasonable options when their costs drop, including passing the saving on, adding features, or improving their own margins. Falling model prices make a lower subscription possible, not automatic.
Should I switch to whichever provider is cheapest?
Not on price alone, and this fortnight is a good illustration of why. The provider that was cheapest by a wide margin is the one that just raised rates by more than 350%, so a decision made purely on today’s number could have been wrong within a month. Choose on whether the tool does your job well, whether your data can leave it, and whether the price is standard rather than introductory. Treat cost as a tiebreaker.
What are peak and off-peak hours, and do they affect me?
They are a billing structure that charges more when the provider’s servers are busiest and less when they are quiet, the same idea as off-peak electricity tariffs. Right now DeepSeek is the notable provider using it for AI, with peak windows defined in UTC, so it affects you today only if you or your developer use that API directly. It is worth knowing about because it is a leading indicator, and because it rewards a habit worth forming early: deciding which of your AI jobs genuinely need to run right now.
How do I find out whether my AI pricing is introductory?
Check the vendor’s own pricing page rather than a comparison article, and look specifically for words like introductory, promotional, launch pricing, or a date. If nothing is stated, a short email to support asking whether your current rate is standard or time-limited will usually get a clear answer in writing, which is useful to keep. Be aware that third-party roundups go stale quickly; several published this week still list an Anthropic price increase that was cancelled on August 10.
One question we keep coming back to: if a tool you rely on offered you a genuine discount for running your non-urgent work overnight, would you actually change how you work to take it, or is the saving simply not worth the rearranging? We would like to hear where your line sits.
