OpenAI is cutting deep, private discounts to pull big customers away from Anthropic, and it is working. According to reporting from The Information, relayed by The Decoder, Rippling, Replit and DataCamp have all moved meaningful AI spend toward OpenAI after aggressive offers. At the same time, public API prices from both labs have fallen fast. For anyone paying per token, the result is simple: the price you pay now depends on who you are, what you commit to, and whether you ask.
Key points
- OpenAI is offering private discounts of 30 to 40 percent to large customers, often tied to using Codex or committing spend.
- Rippling says OpenAI went from 16 percent to 52 percent of its coding workload, while Anthropic fell from 57 percent to 18 percent.
- The widely shared “up to 90 percent” figure is an estimate from one CEO, not a confirmed discount.
- Public list prices are dropping too. Anthropic’s Opus 5.5 lists at $4 input and $20 output per million tokens, with cache reads at $0.20.
- For builders, cache pricing and switching costs now matter more than headline input and output rates.
What happened
The clearest data point comes from Rippling. CEO Parker Conrad shared that OpenAI’s share of the company’s AI coding workload rose from 16 percent to 52 percent over a few months, while Anthropic’s share dropped from 57 percent to 18 percent. That is not a pilot. That is a primary vendor swap inside a large software company.
Replit is the most visible case. It reportedly received more than 30 percent off a contract worth over $100 million a year. The trigger, per the report, was Anthropic declining to extend Replit’s existing discount unless Replit committed to higher spending. CEO Amjad Masad first said on stage that Replit was “no longer working with Anthropic,” then walked it back. Replit now calls itself model agnostic, with OpenAI as the default partner for Replit Agent while Claude models stay selectable in its docs.
DataCamp got 40 percent off for internal product development, on the condition that it uses Codex. CEO Jonathan Cornelissen estimated that if similar terms reached the models inside DataCamp’s product, costs could fall 80 to 90 percent. That estimate is where the “up to 90 percent” headlines come from. It is a projection, not a deal on paper.
The report also describes an unnamed US bank whose ChatGPT seat count was doubled to 500 at no extra cost, which works out to under $7 per user against a typical $20 or more.
Why OpenAI can do this now
Part of this is strategy. Anthropic’s growth reportedly slowed from the second quarter to the third, according to an investor memo cited in the report, and OpenAI sees Anthropic’s higher margins as room to undercut. Both companies are also under pressure to show revenue momentum ahead of possible public listings.
Part of it is real efficiency. Max Woolf’s analysis, “OpenAI and Anthropic cut their LLM prices by up to 80% within a few months. Why?”, which drew heavy discussion on Hacker News, walks through the cuts:
- GPT-5.6 Luna fell 80 percent, from $1 input and $6 output per million tokens to $0.20 and $1.20.
- GPT-6 Sol launched at $2 input and $10 output, far below the $5 and $30 of GPT-5.6 Sol.
- GPT-6.1 Sol kept $2 and $10 but halved cache reads to $0.10.
- Anthropic made Sonnet 5’s $2 and $10 pricing permanent and cut Fable 5.1 cache reads by 75 percent.
- Opus 5.5 came in at $4 and $20 with $0.20 cache reads, which Anthropic says reflects a model that “requires less compute to serve.”
You can check Anthropic’s current rates on its pricing page.
The number most teams ignore: cache reads
Woolf’s sharpest point is about where agent spend actually goes. In agentic coding, cache reads make up roughly 96 to 99 percent of tokens processed, because each turn re-reads the same growing context. That means a cut in cache read price can matter more than any change to input or output rates.
It also explains a paradox. Data from Vercel’s AI gateway shows the average cost per token fell to about 43 percent of its spring level, yet total spend kept climbing. Cheaper tokens mean agents run longer, call more tools and re-read more context. The bill does not shrink on its own.
Open-weight competition adds pressure from below. DeepSeek V4.1 Flash sits around $0.15 input and $0.60 output with near-free cache reads, served by dozens of providers on OpenRouter, and Moonshot’s Kimi K3.1 Pro matches Sonnet-level pricing at $2 and $10.
What this means for builders
- Ask for a discount. If your monthly AI spend is meaningful, list price is a starting point. These reports show both labs negotiate, and competitive offers move the number.
- Price on cache, not just input and output. For agent workloads, compare cache read rates first. A model with cheaper outputs but pricier cache reads can cost more in practice.
- Stay portable. Replit could switch defaults quickly because it kept multiple models wired in. A thin abstraction layer or an AI gateway gives you leverage at renewal time.
- Read the strings. Many of these deals are tied to specific products like Codex or to spend commitments. A big percentage off can lock you in harder than the list price would.
- Watch subscription limits too. Cheaper tokens have not meant more generous plans everywhere. Woolf notes Codex’s top tier usage allowance was cut in half, so seat-based pricing can move the other way.
What to watch
Woolf’s thesis is that when switching between frontier models is cheap, prices drift toward the marginal cost of serving them. If that holds, the real moats become enterprise contracts and agent products such as OpenAI’s Dots and Anthropic’s Claude Duet, not raw model access.
The next signals to track are whether Anthropic answers with its own private discounts or another public price cut, whether more large customers publish workload splits like Rippling did, and whether either lab starts pulling back on discounting once fundraising and listing pressure eases. For now, buyers have more leverage than they have had in a long time.
