Rippling ran 57% 10 of its coding work on Anthropic models in May. By October, Rippling CEO Parker Conrad told reporters, that share was 18% 01. Over the same five months, he said, OpenAI's share climbed from 16% 11 to 52%. His reason was cost.
The Decoder, relaying that reporting on October 25, says OpenAI is reportedly offering discounts of up to 90% 03 to win Anthropic's biggest accounts. One report says Replit shifted most of its AI spending to OpenAI after OpenAI offered more than 30% 04 off a contract worth over $100 million a year. Thirty percent of $100 million is $30 million 25. OpenAI gave up at least that much revenue per year to win one logo.
Techpresso notes that neither company has confirmed the deal terms. The 90% 12 figure is one customer's total cost drop, and no published price list shows it. Every builder buying tokens should still read it as a warning about the bill that arrives at renewal.
List prices turned into opening bids
The published price per million tokens is now just a reference price. What a large account actually pays is a negotiated package, and the October reports show that package getting built in at least three ways.
Three reported discounts and the renewal jump they hide.
The first is the straight cut, and Replit's reported 30%-plus 04 discount on a nine-figure contract is the plain version. The second is free seats. The Decoder reports a US bank with more than $7 billion 06 in assets doubled its ChatGPT seats to 500 employees at no extra charge. According to the same reporting, that bank pays under $7 09 per user against the usual $20 or more.
The third is the conditional discount. Techpresso reports DataCamp, a training company of about 200 07 people, was offered roughly 40% off models for internal product development if it adopted Codex, OpenAI's coding agent.
The gap between a 40% 07 offer and an 80% saving likely includes moving work onto cheaper models. Sort any vendor offer into those three buckets and you can see what it will cost you later. A straight cut expires at renewal. Free seats put more of your staff on the platform. A conditional discount buys your workflow, and your workflow is the expensive thing to move.
OpenAI is paying acquisition cost in margin
A discount is customer acquisition cost under a different name. Customer acquisition cost, or CAC, is what a company spends to win one customer. In OpenAI's case, the spend shows up as revenue it never books.
Apply that to Replit. If the discount is $30 million 26 a year, OpenAI needs lifetime value, or LTV, to cover it. LTV is the gross profit a customer produces over every year it stays. A common SaaS rule of thumb says LTV should be at least three times CAC. So OpenAI needs Replit to stay long after the discount ends, or to pay closer to list later.
What OpenAI is really buying is habit. Once engineers write prompts and test suites around one model family, leaving takes real work. For now, though, the move is still cheap. Rippling's OpenAI share jumped 36 27 points in five months.
Anthropic made the opposite call on margin. An investor memo cited by The Decoder says Anthropic's growth slowed from Q2 to Q3.
For builders, the headline token price is a distraction. It tells you very little about your bill. The invoice is decided by output-heavy agent loops and by whatever condition sits in the contract.
If your product only makes money at a 90% 03 discount, what are you selling? A subsidy with a user interface on top. That is a pass-through business dressed up as software, and its margin belongs to someone else's sales team.
There is a fair counterpoint. OpenAI's website says OpenAI has achieved real price cuts through its Sol and Luna model lines. So some of today's low price will stick. Whether that permanent share is a quarter of the drop or most of it is unclear, and neither company has said.
We think the boring move wins here. Measure cost per successful task every week across two vendors. Nobody will write a launch post about that spreadsheet, but it is the document that gets you a better renewal.
What the renewal invoice will actually charge
A 90% 03 cut turns into a 10x 29 bill.
A team paying $1,000 a month after a 90% discount faces $10,000 at list for the same usage. The bank's 500 06 seats at under $7 09 cost under $3,500 30, and at $20 they cost $10,000 31 or more.
Conditional discounts buy your workflow.
DataCamp was reportedly offered roughly 40% off if it adopted Codex. Every bundled tool raises the cost of leaving, and the workflow is the expensive thing to move.
Rippling moved 36 points in five months.
Rippling's OpenAI share rose about 7 points a month, from 16% 11 to 52%. At that pace a company could shift half its workload in about seven months, which gives buyers real leverage today.
Subsidies meet public-market scrutiny by 2027
On October 11, Michael Parekh wrote that both labs were trimming subsidies ahead of their planned public listings. Two weeks later came the reports of 90% 12 cost drops. The Decoder ties Anthropic's slowing growth to a planned November IPO. Public companies report gross margin every quarter, and under that scrutiny a $30 million 26 annual concession on one account is hard to repeat.
The subsidies are also aimed at different targets. Parekh, citing SemiAnalysis, found a $200 13 Anthropic plan buys about five times OpenAI's value at API prices. Each lab is spending to win a different kind of customer. Every builder falls into one of those camps and should know which.
Say you pay $1,000 29 a month after a 90% 03 cut. List for that same usage is $10,000, a 10x jump if the discount ends. The bank's 500 06 seats cost under $3,500 30 at $7 09 a seat. At $20 a seat they cost $10,000 31 or more, nearly three times as much.
Rippling's share moved about 7 points a month, 36 32 points over five months. At that pace a company could shift half its workload in about seven months. Because switching is cheap today, offers now carry conditions like Codex adoption. Every bundled tool raises the cost of leaving.
The risk is lopsided. Keeping a second vendor tested costs a few engineering hours each month. Getting caught with only one vendor at renewal can cost 10x 34.
Price every call at list before renewal
The build is a weekly cost report across two vendors for one workload. Pick the workload with the biggest bill.
Step one: put a gateway in front of your model calls. A gateway is a single internal service every request passes through, so your app never calls a vendor directly. LiteLLM and OpenRouter both do this. Move one workload behind it and leave the rest alone for now.
Step two: build a private test set of 50 16 real tasks pulled from your logs. Include the ugly ones, like long inputs, tool calls and the requests that broke last month. Write a pass or fail check for each task.
Step three: run the set on your current vendor and one rival every Monday. Log tokens in, tokens out, retries and pass rate. Divide total cost by passed tasks to get cost per successful task.
Step four: compute that number twice, once at your contract price and once at list. The gap is your exposure at renewal. If your margin disappears at list, your real margin today is smaller than your dashboard says.
Step five: keep a one-page ledger of every discount you hold. Write down the percentage, the condition attached and the expiry date. When renewal is 90 days out, send your vendor the rival's cost per successful task.
Expect breakage. The rival will fail tasks your current model passes, and prompts tuned for one vendor will need rewriting. Those failures are the data. A model that is 30% 04 cheaper but needs 20% 35 more human fixes may cost you more, and your Monday report will show it in week one.
Build a weekly cost report across two vendors.
- Put a gateway in front of one workload. Route your biggest-bill workload through LiteLLM or OpenRouter so your app never calls a vendor directly, and leave everything else alone for now.
- Build a 50-task 16 private test set. Pull real tasks from your logs, including long inputs, tool calls and last month's failures, and write a pass or fail check for each one.
- Price every passed task at contract and at list. Run the set every Monday on your vendor and one rival, divide total cost by passed tasks, and log each discount's percentage, condition and expiry in a one-page ledger.
Price every call at list before your renewal.
OpenAI is paying for customers in margin, and public-market scrutiny makes concessions like Replit's $30 million 26 a year hard to repeat. Some of today's price cuts will stick, and neither lab has said how much. We think the safe plan is a second vendor kept tested every week, a cost-per-successful-task number at list price, and a ledger of every discount's expiry. Send that number to your vendor 90 days before renewal.