Ninety percent. That’s the number that made me stop mid-scroll last week, coffee halfway to my mouth.
The news: OpenAI is offering discounts of up to 90 percent to Anthropic’s enterprise customers. Not coupons, not a Black Friday thing, a deliberate poaching operation. Rippling, Replit, and DataCamp have all switched, and all three say publicly that they cut their AI costs by up to 90 percent after the move. Anthropic’s growth has slowed as a result. It’s aggressive, a little ruthless, and judging by those numbers, it’s working.
If you run a small business, your situation looks nothing like Replit’s. I build admin automations for small companies for a living (there’s a longer version of who I am and how I work if you’re curious), and most of my clients’ AI bills are the size of a nice dinner, not a payroll line. If you’re new to the topic, our no-hype playbook on AI automation for small businesses is the best place to start. For everyone else: here’s what the discount war actually means for your API bill.
Short answer: it didn’t get cheaper. Nobody has announced SMB-specific pricing, and these are negotiated enterprise deals, the kind you get when your annual spend has six or seven digits in it. The news still matters for the rest of us, just not the way the headline suggests.
What OpenAI is actually offering, and to whom
The mechanics, as reported: OpenAI is going after Anthropic’s biggest enterprise customers directly, offering up to 90 percent off to switch sides. There’s no signup form for this; it’s the kind of deal you negotiate when your AI bill runs to millions a year. Rippling, Replit, and DataCamp took it, and each has said publicly that their AI costs dropped by up to 90 percent after switching. Anthropic’s growth has slowed since. Nobody gives away 90 percent out of generosity; you do it to bleed a rival’s revenue and lock customers in before they find their footing again.
What’s conspicuously missing for us: any SMB-specific pricing. Nothing in the reporting suggests a small business can call up and ask for its 90 percent. Temper that hope now.
Quick context: what OpenAI’s API does in small-business admin
If you haven’t touched an AI API directly, here’s the role it plays in the admin automations I build. OpenAI’s API drafts replies to customer emails, summarizes long PDFs, pulls line items out of invoices, sorts inbound messages into categories, and turns messy notes into clean records. Honestly, it’s the single most useful category of AI for a small admin team. You pay per token, roughly per chunk of text processed, so the bill scales with volume, not seats. Across my clients, that means anywhere from about $40 a month for a light email-triage flow to a few hundred for document-heavy processing. Those are numbers from my own projects, not official rate cards, so treat them as ballpark.
For a more systematic view, I wrote up what an SMB should actually budget for AI earlier this year, and the ranges there have held up.
Why Rippling, Replit, and DataCamp switched
I don’t blame them one bit. These are companies whose products run on enormous model usage, so a 90 percent cut isn’t a nice-to-have, it’s a board-level mandate. And the fact that all three put their names on the savings tells you the numbers were real enough to brag about.
What they didn’t disclose matters just as much. We don’t know their baselines. We don’t know whether “up to 90 percent” describes the whole bill or the best-case line item, whether the rate holds for a contract term or reverts after a year, or what happens the day the discount expires. “Up to” is doing heavy lifting in those announcements, the same way it does in furniture-store ads.
There’s a broader read here, and it’s mildly good news for the rest of us: when the two front-runners start a price war, every buyer’s position improves a little. If you’re an existing customer of either provider with a bill worth noticing, you suddenly have a legitimate conversation starter. Vendors hate losing accounts to a rival’s discount, and retention offers exist. It costs one email to ask.
The bill nobody puts in the press release
Switching model providers is almost never free, even when the new price is dramatically lower. The sales deck makes the APIs sound interchangeable. The models underneath them are not.
Prompts tuned on one model rarely transfer cleanly to another. Tone drifts, formatting breaks, and the quirks you quietly compensated for months ago come back with teeth. If your workflow depends on structured output (specific fields, fixed formats, JSON), you’ll be re-testing all of it, and your old “good output” examples guarantee nothing about the new model. When things break, more prompt tweaking is usually a band-aid, not a fix; I wrote about why prompt engineering can’t fix your AI mistakes a while back and I stand by it.
I learned the migration tax the expensive way. Last spring, a logistics client of mine (14 people, supplier invoices arriving as PDFs) ran an invoice-summary workflow on Anthropic’s API: read the document, extract the fields, draft a tight internal summary email. Around $180 a month, worked fine, nobody thought about it. I got it into my head that we could shave the bill by moving it to OpenAI, ran a quick test on a few invoices, and switched it over.
Within a week, their office manager noticed the summaries had grown chatty intros and bullet-point headers instead of the three-sentence format everyone was used to. Nothing crashed. It was just wrong in the way that makes people stop trusting a tool. Turns out a three-invoice spot check isn’t a migration test. I spent about six hours rewriting prompts and re-running 40 invoices by hand, and the new bill landed around $120 a month.
Sixty dollars saved. Roughly $450 of labor spent, two months of goodwill burned, break-even somewhere near the eight-month mark. On a small workflow, where I already knew both APIs cold.
The second story is worse, and it’s the one I actually regret. Different client, a support-email drafting flow, and I swapped models to chase about 40 percent off a bill that was, in hindsight, tiny. I eyeballed maybe ten outputs, thought “close enough”, and shipped it. The drafts came back subtly off: a little too formal, occasionally misreading which thread they belonged to. Their office manager caught it before customers did, which is the only reason this isn’t a worse story. We reverted after three weeks and I ate the hours. The savings would’ve been around $70 a month. I traded a system people trusted for $70, and I’m still a bit embarrassed by how casually I made that call.
I’ve gone back and forth on what the real lesson is. Build a proper test set first, sure. But maybe optimizing a $70 decision at all was the mistake, and honestly I’m not sure which of those two I’d defend at a dinner party. Both, probably.
Then there’s lock-in, the quiet strategic point of discounts like these. Ninety percent off is a very good price for your habits. Once your prompts, test sets, integrations, and team intuition are shaped around one provider, leaving in year two costs far more than it did in year one, discount or no discount. OpenAI isn’t uniquely villainous here; Anthropic would run the same play with the positions reversed. Just walk in knowing what year two looks like.
Three questions before you spend an hour on this
If you’re wondering whether to chase the discount, run this checklist first. Question one disqualifies most of us, which is kind of the point.
1. Pull up your real monthly API spend
The actual number, not the estimate in your head. Under a couple hundred dollars a month? Even a genuine 90 percent discount saves less than a day of migration labor costs, and no SMB discount has been announced anyway, so today you’re comparing list prices. If you’re pushing $1,000 or more a month through an API, though, you’re no longer invisible to sales teams, and a polite “what can you do on rate” email to both providers is worth your time.
2. Measure output quality before you touch anything
Save 30 real inputs from your current workflow. Run them through the current model and the candidate model, and compare the results with someone who uses those outputs daily. Without that, you can’t verify the discount didn’t quietly come with a quality downgrade, and you’re shipping on vibes. (The support-email story above was me shipping on vibes. It went how it went.)
3. Ask what happens when the discount ends
Discounts are introductory by design. Get four things in writing: how long the rate lasts, what you pay after it expires, what leaving costs, and who owns your prompts and data if you walk. Vague answers are their own kind of answer. I bring the same skepticism to any AI vendor, and the longer version lives in my 5-question checklist for buying AI without getting sold a chatbot.
So what should a small business actually do?
My honest verdict: nothing dramatic this week. Use the news as a nudge to do two cheap things instead.
First, look at your current spend, because you can’t negotiate or ignore a number you don’t know. Second, build a little portability into whatever you automate: keep prompts in your automation layer (n8n, Make, whatever you use) rather than buried in vendor-specific settings, and keep a folder of test inputs handy. Do that once and a future model switch becomes an afternoon of verification instead of a six-hour surprise like mine was. The AI automation for SMBs playbook covers how to structure workflows so the model layer stays swappable.
And if your bill is genuinely large, congratulations, you’ve just gained a bargaining chip you didn’t have last month. Ask your current provider what they can do now that a rival is offering your neighbors 90 percent off. Worst case is a no, and you’ve lost one email.
That’s my read, anyway. Ask me again in six months and this discount might be standard practice, or gone entirely. I’ve stopped pretending I can predict this market. I just try to keep our automations cheap to leave.
Talk soon,
Damian
