The short version: Three frontier AI models landed inside four days, and not one of them raised its price. Anthropic’s Claude Opus 5 shipped on July 24 at exactly the same rate as the model it replaces. Moonshot AI is giving away its largest model ever as free weights. If you run a small business, none of this touches your invoice this month. It changes what the AI software for small business you already pay for can afford to hand you next year, and that is worth knowing before you sign anything twelve months long.
What actually happened in four days?
Start with the facts, because the headlines this week were mostly about a security incident and buried the more consequential story underneath.
On July 24, Anthropic released Claude Opus 5. The company priced it at $5 per million input tokens and $25 per million output tokens, which it notes is the same as Opus 4.8. It added an adjustable effort setting that customers can use, in Anthropic’s words, to optimize for intelligence or conserve tokens for faster and cheaper results.
Days earlier, Moonshot AI put Kimi K3 into service through its API, with the full weights scheduled to go public on July 27. As Tom’s Hardware reported, it is a 2.8-trillion-parameter mixture-of-experts model, the largest open-weight release yet attempted, and it took the top spot in the Frontend Code Arena benchmark at 1,679 points, ahead of Claude Fable 5 at 1,631 and GPT-5.6 Sol at 1,618.
And DeepSeek pushed V4 out in mid-July with a one-million-token context window and, per TechNode, a new pricing structure that charges twice the off-peak rate during two daily peak windows.
Why is every headline a cost claim in disguise?
Here is the part almost nobody said out loud. Read Anthropic’s own list of what makes Opus 5 better and count how many of the claims are about capability alone.
On the OSWorld 2.0 benchmark it outperforms Fable 5 “at just over a third of the cost.” On CursorBench 3.2 it lands within 0.5% of Fable 5’s peak score “but at half the cost per task.” On Frontier-Bench v0.1 it “more than doubles Opus 4.8’s performance at a lower cost.”
Every one of those is a price-performance claim. The pitch is not that the model does something it could not do before. The pitch is that it does the same work for meaningfully less money. That is not a small rhetorical shift. It is the signature of a market moving out of its land-grab phase and into the phase where vendors compete on unit economics, which is the phase where buyers start winning.
What does this change for AI software for small business?
Almost nothing this month, and possibly quite a lot next year. The reason is a gap most owners never think about: you buy seats, your software vendor buys tokens.
When you pay a flat monthly fee for a booking tool, a CRM, or an invoicing app with an AI feature bolted on, that vendor is paying per-token costs behind the scenes every time you use it. For the last two years, many of those AI features were loss leaders. They existed to keep the product competitive, and finance tolerated them.
When the cost of the intelligence underneath drops by half, that math changes. The realistic effects to watch for are features migrating out of premium tiers into base plans, usage caps quietly loosening, and monthly “AI credit” allowances getting more generous without an announcement.
Be careful how much you bank on that. Falling input costs create room for vendors to pass savings along; they do not obligate anyone. Some will simply keep the margin. The point is not that your software will definitely get cheaper. The point is that your negotiating position just improved, and most owners will not notice.
Should you sign an annual AI software contract right now?
Our honest read: prefer monthly for anything AI-heavy, at least through the rest of this year, unless the annual discount is genuinely large.
Locking in twelve months at today’s price is a bet that the cost curve flattens out. This week is evidence it has not. If a vendor pushes an annual commitment, two questions are worth asking before you sign. First, what happens to my usage limits if your underlying model costs fall during the term. Second, can I move down a tier mid-contract without a penalty. The answers tell you a lot about who is capturing the savings.
This is the same underlying pressure driving the owners we covered who are firing Salesforce and HubSpot to build their own AI CRMs. Cheap intelligence keeps changing what counts as a fair price for software.
Where cheap intelligence is not actually cheap
Two honest caveats, because the “AI is getting free” story gets oversold constantly.
DeepSeek’s peak-hour structure is the first. Charging double during two daily windows means the advertised low rate is a floor, not the number you will actually average if your usage clusters during business hours. Cheap is conditional.
Kimi K3 is the second. Free weights are not free deployment. The model runs roughly 1.4 terabytes in four-bit precision, which puts self-hosting well outside what any small business would attempt; in practice you would rent it from a hosting provider anyway. And when the API went live, Moonshot had not yet published its license terms, so whether the weights are usable commercially was still an open question. “Open” and “free to use in your business” are not synonyms until someone reads the license.
What to do this month
Three moves, none of which take more than an hour.
Pull up your card statement and list every software subscription with an AI component, along with its renewal date. Most owners are surprised by the total. Then, for anything renewing in the next ninety days, check whether monthly costs meaningfully more than annual; where the gap is small, take the flexibility. Finally, pick one AI tool you rejected six months ago because the price did not justify it and test it again. Price-performance moved. Your old no may have expired.
If you want a starting point for that last one, our roundup of the best AI tools for small business, organized by the job you need done, is built for exactly this kind of re-evaluation. And once you have picked something, OpenAI’s four-part ROI scorecard for AI spend is a lightweight way to check whether it is actually paying for itself.
Frequently Asked Questions
Does the AI price war mean my ChatGPT or Claude subscription will get cheaper?
Not directly, and probably not soon. Consumer and business subscription tiers have stayed remarkably stable while per-token API prices have fallen. What tends to change instead is what you get for the same money: higher limits, better default models, and features that used to sit behind a higher tier.
Should small businesses care about open-weight models like Kimi K3?
Not as something you would run yourself. A 1.4-terabyte model needs infrastructure no small business should be buying. Care about it for its effect on prices: every capable free model puts downward pressure on what the commercial vendors can charge, and you benefit from that indirectly through your software.
Is a cheaper AI model a worse AI model?
Not anymore, which is the genuinely new part. This week’s releases claimed better results at lower cost simultaneously, and an open-weight model topped a coding benchmark ahead of two frontier commercial models. Judge tools on whether they do your specific job well, not on price as a proxy for quality.
How do I know if my software vendor is passing savings on to me?
Watch your plan’s fine print at renewal rather than your price. If your usage caps, AI credits, or included features have not improved in a year while the cost of models has halved, the savings are being kept upstream. That is a reasonable thing to raise in a renewal conversation.
Here is what we are genuinely curious about: when you look at your own software bill, has any vendor actually given you more AI for the same money this year, or has the price just quietly stayed the same? Tell us what you are seeing.
