If you buy AI models for a company, the story to watch this week is not a new model. It is a price. The Decoder reports, citing The Information, that OpenAI is offering discounts of up to 90 percent to pull Anthropic's biggest customers over. We could not open the paywalled original, so this post separates what is reported and attributed from what is confirmed by the labs, which is nothing yet.
It follows earlier explainx.ai coverage of Replit saying it no longer works with Anthropic and of Meta and Microsoft cutting internal Claude use.
What is claimed, and by whom?
| Claim | Attributed to | Confirmed by a lab? |
|---|---|---|
| Rippling raised OpenAI's share of its coding workload from 16 to 52 percent; Anthropic's fell from 57 to 18 percent | Rippling CEO Parker Conrad, via The Information and The Decoder | No |
| Replit switched after OpenAI offered more than 30 percent off a contract worth over $100 million a year | The Information, via The Decoder | No |
| DataCamp's costs fell 80 to 90 percent after pricing changes | DataCamp CEO Jonathan Cornelissen, via The Decoder | No |
| A U.S. bank with over $7 billion in assets had ChatGPT seats doubled to 500 at no extra charge, under $7 per user instead of $20 or more | The Information, via The Decoder | No |
| Anthropic growth slowed from Q2 to Q3 | An investor memo cited by the report | No |
| Chinese labs such as DeepSeek and Kimi add price pressure | The Decoder's framing | Prices are public |
The Decoder's own summary is that OpenAI is partly buying market share at the expense of its margins but has also achieved real cuts through its Sol and Luna model lines, thanks to more efficient models and infrastructure. That second part matters: not every cheaper bill is a subsidy.
What came before this report?
This is the latest in a run of reporting about the two labs competing on terms rather than only on benchmarks.
- August 2026. The Information reported OpenAI was pushing steep discounts through intermediaries. A secondary write-up says OpenAI offered its GPT-5.6 lineup at 50 percent off on OpenRouter and that Luna traffic there reportedly exceeded the combined traffic of Claude Opus 5 and Sonnet 5. That summary is from a crypto-news site relaying the original, so treat the traffic comparison cautiously.
- September 28, 2026. The Information reported that Anthropic ends enterprise discounts once customers use up the tokens in their commit, pushing them to renegotiate or pay list. The report puts large-commit discounts at around 15 percent off list. OpenAI was described as more lenient: customers keep the rest of the month plus one more month to renegotiate before reverting to list. Amazon, Microsoft and Google reportedly keep discounted overage pricing during the contract term. Anthropic declined to comment, and CodeRabbit's CEO said OpenAI had displaced Anthropic as its main supplier.
- October 6. We covered Meta and Microsoft cutting internal Claude use.
- October 16-17. At TechCrunch Disrupt, Replit CEO Amjad Masad said the company is no longer working with Anthropic, citing cost, partnership and performance. He later said Replit remains model agnostic and that OpenAI is its preferred partner for Replit Agent. TechCrunch noted insiders disputed his account and that Replit had promoted Anthropic products as recently as late September.
Read together, the October report is consistent with the September one: discount design, not just token price, is where the two labs are fighting.
How can a customer see "up to 90 percent"?
A price tag with a downward arrow beside scissors, showing how a headline OpenAI discount of up to 90 percent can come from stacked cuts
A headline discount of that size almost certainly stacks several things, and the report does not break them apart. The plausible components, based on what is public:
- Model mix. Moving from a premium model to a cheaper tier changes the bill by a large factor on its own. Our GPT-6.1 Sol and Astra cost-efficiency post and the small-model comparison covering Haiku 5.5, GPT-6 Luna, DeepSeek V4.1 Flash and others show how wide the price spread already is.
- Contract discounts. A 30 percent offer on a nine-figure contract is a negotiation outcome, not a price list.
- Free or bundled seats. The bank example, doubling seats at no extra charge, is a promotion on seats rather than API tokens.
- Workload fit. A customer that moves a simple classification job to a cheap model can see an 80 to 90 percent drop without the vendor discounting anything. Our post on Anthropic's Haiku 5.5 launch and cache cut shows Anthropic is also moving prices.
So "up to 90 percent" is best read as one customer's best-case total bill change, not a rate card.
What is not verified?
- No lab statement. We found no OpenAI or Anthropic statement confirming the specific offers.
- No primary document. The Information is paywalled and we relied on The Decoder's summary of it, plus the earlier September reporting.
- Quality-adjusted cost is unknown. Rippling's shift from 16 to 52 percent tells you where volume went, not whether output quality held or what was paid per task.
- Replit's account is disputed. TechCrunch reported that insiders said the partnership remains in place.
- The investor memo. We have not seen it, and slowing growth is not the same as shrinking revenue. For broader spend data, see our write-up of the Ramp AI Index for October 2026.
Why it matters for what you pay
A pie chart with one wedge pulled out, picturing enterprise share of wallet moving between OpenAI and Anthropic
There are three practical consequences even if every number above is partly inflated.
Leverage is real. If two top labs are competing on commit terms, a buyer with meaningful spend has room to negotiate. The September report suggests the biggest differences are in overage and renewal terms, so those are the clauses to read first.
Lock-in cuts both ways. Companies that built a harness around one vendor's models pay a switching cost. Tools that route across providers (our Wally write-up on running seven open models in coding agents is one example) lower it.
Price pressure is global. The Decoder notes Chinese labs add pressure. Our coverage of DeepSeek V4.1 Flash pricing and MiniMax M3.1 Flash shows how low open and low-cost prices have fallen, although government policy (see the DHS DeepSeek procurement ban) limits who can use them.
What to do if you buy AI at scale
- Run your own bake-off. Take 100 to 200 real tasks, run them through both labs' current models, and score quality and cost per completed task, not per million tokens.
- Ask both vendors for a written quote with the same assumptions: volume, commit length, what happens after the commit is exhausted, and whether prices reset at renewal.
- Read the overage and grace terms. The September report shows these differ more than headline discounts.
- Be careful with share-of-wallet clauses. Reporting says both labs are adding that language to large deals, while consultants say neither can verify compliance. If you accept a clause tied to your usage share, make sure you can measure it.
- Keep a second provider warm. Even if you stay put, a tested fallback is what makes a quote credible.
- Do not anchor on one customer's percentage. Your workload, model mix and volume decide your real number.
How should you read the Rippling and Replit numbers?
The Rippling figures are the most specific in the report, and also the easiest to over-read. A move from 16 to 52 percent for OpenAI and from 57 to 18 percent for Anthropic describes a share of one company's coding workload. It does not say how large that workload is, whether total spend rose or fell, or whether the work moved because of price, quality, latency or a new model release. A company can shift most of its volume and still spend similar money if usage grows.
The Replit figure is a contract offer, more than 30 percent off a deal worth over $100 million a year, which implies a discount of well over $30 million annually at that scale. That is a large number, but it describes an offer rather than a signed price, and Replit's own public statements have been more cautious than the headline suggests. The DataCamp figure, 80 to 90 percent, is the executive's description of his cost change after switching, and a drop that large usually reflects a different model tier as much as a different vendor.
None of this makes the reports wrong. It means each number answers a narrower question than the headline implies.
What would change this assessment
We will update this post if OpenAI or Anthropic comments, if The Information's article becomes accessible with more detail, or if a named customer publishes its own numbers. A confirmed company statement or a customer's published invoice comparison would move specific claims from reported to verified. Until then, the sensible reading is that competition for enterprise contracts has intensified, the exact discounts are customer-reported, and the label "up to 90 percent" describes a best case.
Related reading
- Replit says it is no longer working with Anthropic: claimed vs verified
- Meta and Microsoft are cutting internal Claude use
- Ramp AI Index, October 2026
- GPT-6.1 Sol and Astra cost efficiency
- Claude Haiku 5.5 launch, pricing and benchmarks
- Small models compared: Haiku 5.5, Luna, DeepSeek V4.1 Flash and more
- Sources: The Decoder, TechCrunch on Replit, Dealroom summary of The Information
Figures are as reported on October 25, 2026 and attributed to the outlets named; pricing and contract terms change frequently.