OpenAI offers up to 90% discount to poach Anthropic's customers: Replit, Rippling, and DataCamp make the switch

OpenAI offers up to 90% off to poach Anthropic's customers: Replit, Rippling, and DataCamp switch over

🔎 OpenAI buys recurring revenue, Anthropic defends its margin

The figure is already making the rounds on news feeds: 90% off. Let's set the record straight right away — this is not an observed catalog discount, it's an estimate of total savings on the customer side. The documented reality is almost as brutal, and far more interesting.

According to a report from The Information published on October 25, 2025 and picked up by The Decoder, OpenAI has been running a targeted pricing offensive against Anthropic's enterprise portfolio for several months. Three companies illustrate it: Replit, Rippling, and DataCamp. Signed contracts, measured workload shares, quantified discounts — not hallway rumors.

Why now? Because the model layer is becoming commoditized, the quality gap between the best models is narrowing, and OpenAI suddenly has every reason to buy recurring revenue rather than build it. The enterprise price war is only just beginning.

The essentials

• According to The Information (October 25, 2025), OpenAI is granting massive enterprise discounts to take share from Anthropic, with code as the main battleground.
• Replit: discount of more than 30% on a contract worth over $100M per year; OpenAI becomes the default partner for Replit Agent.
• Rippling: OpenAI's share of code workload jumps from 16% to 52%, while Anthropic's drops from 57% to 18%.
• DataCamp: -40% observed on its internal models; up to 80-90% in total estimated savings (an estimate, not a list discount).
• The logic: sacrifice immediate margin to lock in recurring revenue before OpenAI's IPO.
• Anthropic is responding with value (the Stainless acquisition, an enterprise JV) rather than price — at the risk of seeing its "code" revenue base erode.

Recommended tools

Agentic benchmark figures we track (June 2025 data): GPT-5.5 98.2 · Claude Opus 4.7 94.3 · Claude Sonnet 4.6 81.4 · GPT-5.3 Codex 80.

The three contracts that tell the whole war

Three deals documented by The Information (October 25, 2025) and analyzed by MLQ.ai are enough to understand OpenAI's strategy: buy a default partner, prove a workload transfer, and showcase spectacular savings.

Replit: over $100M per year, and a changed "default"

Replit signed a contract worth more than $100 million per year, with a discount of over 30% off list prices. But the real strategic clause isn't the percentage: OpenAI becomes the default partner for Replit Agent.

A "default" in a development tool is a distribution channel. Every developer who launches a Replit agent generates OpenAI inference without any explicit purchasing decision. The 30% discount isn't a cost, it's an investment: OpenAI is buying a recurring stream of tokens, backed by a user base that grows on its own.

This is the most important deal of the three. And paradoxically, the least publicized.

Rippling: proof by workload share

At Rippling, the shift is measured by the machine, not by the press release. OpenAI's share of the code workload went from 16% to 52%. Anthropic's dropped from 57% to 18%.

These figures are valuable because they come from usage telemetry, not a survey. Two things can be read in them: a complete crossover in just a few months, and coexistence — even after the switch, Anthropic retains 18% of the workload. The real enterprise is multi-vendor, with weights that shift.

The trigger? A price-parity mix. In our June 2025 agentic benchmarks, GPT-5.5 (98.2) overtook Claude Opus 4.7 Adaptive (94.3), and GPT-5.3 Codex (80) is operating in the same range as Claude Sonnet 4.6 (81.4). When the quality gap closes, price decides.

DataCamp: -40% observed, -80/-90% estimated

DataCamp obtained a 40% discount on the models it uses internally. The 80-90% savings figure driving the headlines is a total estimate from the client side: the discount, plus workload optimization — routing to cheaper models, caching, reducing token waste.

This is the figure to handle with the most caution. An estimate of total savings is not an observed list-price discount. Confusing the two means telling a price war story that is deeper than it really is.

One table, and the whole strategy fits inside it: a discount to lock in, a measurement to prove, an estimate to make noise.

Why OpenAI is willing to earn less per contract

Because at this stage of the race, a dollar of recurring revenue is worth more than a dollar of margin — especially just months away from an IPO.

OpenAI has officially filed for its IPO, and the biggest tech listing of the decade is imminent. Yet public markets don't value the margin of a hypergrowth platform: they value ARR and its trajectory. A multi-year enterprise contract, even discounted by 30 to 40%, is predictable ARR. That's exactly what an IPO filing needs to show.

Second reason, a structural one: land-and-expand. The default partner position (Replit Agent) turns every user into inference volume. At renewal time, the incumbent vendor negotiates with the advantage of switching costs — and that cost, the customer has already paid once.

Third reason: hitting Anthropic where it hurts. Code is the core of Anthropic's enterprise revenue. Every share point wrested away on a coding workload strikes at its foundation. It's no coincidence that all three documented deals involve companies that code at scale.

Finally, the broader context: both groups have each launched their own $10 billion enterprise JV to deploy AI across SMBs and large corporations. The enterprise market has become the main battleground, and price is the weapon. The comparison with the cloud discount wars of the 2010s is unavoidable: at AWS, Azure, and Google Cloud, list prices were never the actual prices paid by major accounts.

Anthropic, the Target: A "Code" Foundation Under Direct Attack

Anthropic built its enterprise franchise on code, and OpenAI is attacking exactly that foundation.

The Claude models — Claude Sonnet 4.6 in volume, Claude Opus 4.6 and 4.7 for critical use cases — have established themselves as the benchmark for agentic coding. That dominance has fueled Anthropic's enterprise growth. It also explains its vulnerability: when 57% of a customer like Rippling's coding workload runs on your platform, every switch translates directly into market-share points.

The problem isn't just the opponent's counterpunch. It's technical parity. On our June 2025 data, GPT-5.5 dominates the agentic rankings (98.2 vs 94.3 for Claude Opus 4.7 Adaptive). As long as Claude was the only one able to sustain long, reliable coding tasks, the price premium held up. With parity, it becomes a cost that CIOs challenge in executive committee meetings.

Anthropic's response is consistent with its positioning: defending value, not price. The acquisition of Stainless for over $300M locks down the SDK and integrations layer — where technical retention is decided. And the valuation tells the same story: a round targeting $900 billion can't be justified with commodity margins.

The risk is crystal clear: revenue concentrated in coding, under attack from a competitor slashing prices. If technical parity persists and the switches keep piling up, growth slows at the exact moment valuation expectations peak. That's precisely the scenario The Information is documenting.

Discounts, lock-in, and switching costs: the real mechanics of a switch

A discount doesn't flip a customer: it's the combination of discount + parity + reduced switching cost that does.

What a CIO actually evaluates when changing model providers: rewriting prompts and evals, differences in tool-calling behavior, latency profiles, the risk of regression on agents in production. This cost can represent months of engineering — and it's this, more than price, that froze customers in place at Anthropic.

OpenAI's playbook is to reduce these frictions all at once. Benchmark parity reduces the re-evaluation work. Default-partner status removes the adoption decision. Discounts fund the migration. Rippling didn't switch for a percentage: it switched because the total cost of change became lower than the savings realized.

A detail that matters: even after switching, Anthropic keeps 18% of Rippling's workload. The enterprise no longer chooses a model — it allocates shares across providers according to cost, latency, and reliability, use case by use case. For providers, the strategic question is no longer "how do we win 100% of a customer" but "how do we own the default workload."

How to negotiate your LLM contract before the next renewal

If your organization consumes AI-generated code at scale, you're in a position of strength — provided you show up with numbers.

First prerequisite: telemetry. Tokens per team, agent success rates, cost per merged pull request, breakdown by model. Without this data, you negotiate blind; with it, you turn every renewal into a competitive tender.

Second lever: competitive tension, now documented. The Information's report (October 25, 2025) is itself a negotiating argument — your vendor knows that you know. Put the two head-to-head on your own evals, not on global scores: an average benchmark says nothing about your workload.

Third lever: the self-host alternative. Open-weight models lend credibility to your exit option — Kimi K2.6 in self-host scores 88.1 on our agentic benchmarks, GLM-5 Reasoning 82. For a self-hosted agent PoC, a Hostinger VPS from ~€7/month (October 2025, check hostinger.com) is enough to test a stack before committing to a three-figure contract.

Finally, negotiate the clauses, not just the percentage: price protection over the term, portability of data and fine-tunes, exit clauses, performance commitments. The most expensive discount is the one that forces you to migrate twice.

The systemic risk: growth that's bought, not earned

Discounts of 30 to 40% on three-figure contracts is growth financed by margin. The question nobody asks: what is the real price of the token?

If list prices aren't the real prices, the reported ARR overstates the sustainable revenue base. It's the kind of gap that eventually attracts regulators' attention — the FTC is already investigating industry practices, a matter we were tracking in our analysis of Anthropic's warnings to its investors.

My read: the 2025 discounts are opening bids, not a new permanent state. After the IPO, OpenAI will have to show margins. The 2026 renewal cycles will be the moment of reckoning — and customers who signed long exclusivities without price protection will pay the bill.

Anthropic, for its part, is betting it can hold the premium by reopening the quality gap: controlled integrations via Stainless, agent reliability, enterprise JV. It's a coherent strategy if the gap widens again. It's a painful strategy if parity persists. The three documented contracts suggest that, for now, parity persists.

❌ Common Mistakes

Mistake 1: Taking the "90%" as an observed discount

The figure circulating is an estimate of total savings at DataCamp: 40% observed discount, with the rest coming from workload optimization. Quoting "90% discount" as a list price amplifies the price war beyond what is documented. The solution: cite 40% as the discount, 80–90% as an estimate of total savings, sourcing The Information (October 25, 2025).

Mistake 2: Choosing a model based on a global score

An average agentic score (GPT-5.5 at 98.2, Claude Opus 4.7 at 94.3) masks variance across use cases: large-scale refactoring, long-running agents, intensive tool-calling. The solution: evaluate on your own evals, on your codebase, with your regression criteria. A switch is decided by your workload, not by a leaderboard.

Mistake 3: Negotiating without usage data

Without telemetry — tokens, success rates, cost per task — your negotiating power is theoretical, while the vendor across the table knows theirs. The solution: instrument first, negotiate second. That's the difference between a courtesy discount and a documented 40% discount.

Mistake 4: Buying a discount with your exclusivity

The Replit deal shows the real trade-off behind discounts: default partner status. Signing a long-term exclusivity for a percentage means mortgaging your upcoming renewals. The solution: multi-vendor by default, exit clauses, price protection — and keep the ability to reroute your workloads.

❓ Frequently Asked Questions

Are OpenAI's discounts official?

No. They were documented by The Information (October 25, 2025) via sources close to the negotiations; OpenAI does not comment on its contractual terms. The figures cited — 30 to 40% discount, total savings estimated at 80-90% — come from customers or insiders, not from a public price list.

Why is code the main battleground?

Because it's the largest, most recurring, and most loyalty-building workload: an agent that codes every day generates millions of tokens. It's also where Anthropic dominates — every point of share gained on code strikes at the heart of its enterprise revenue.

Do these discounts also apply to SMEs?

Mostly no. The documented discounts target contracts in the hundreds of millions of dollars. SMEs go through standard offerings or partner programs — like the enterprise JVs launched by both companies. But pricing pressure always ends up trickling down to the rest of the market.

Will Anthropic match OpenAI's prices?

Nothing suggests so in the short term. Its premium positioning and valuation — a round targeted at $900 billion — make a frontal price war unlikely. The stated strategy: defend value, agent reliability, integrations and SDKs, rather than the price per token.

Should you renegotiate your contract now?

Yes, if you have a significant code workload and multiple possible vendors. The window of competitive tension is real and documented. Instrument your usage, compare using your own evals, and negotiate before renewal — not after.

✅ Conclusion

A 30 to 40% discount on a three-figure contract isn't generosity: it's the purchase price of recurrence that OpenAI intends to monetize long after its IPO. If you run an enterprise code workload, the negotiation window is open — it will close at the first post-listing renewal cycle. And if you should retain only one action item this week: instrument your usage before your next negotiation, it's the only discount nobody can take away from you.