New spending data on OpenAI vs Anthropic enterprise market share says the two labs are swapping the lead every few months, and that churn is the part worth your attention. TechCrunch reported on Ramp's corporate card data showing Anthropic still ahead but OpenAI closing through Q3.
Most people will read that as a scoreboard. I read it as evidence that lock-in barely exists right now. If you build software from Sri Lanka on a small budget, that is leverage you should be using.
📊 What the data says, and what it doesn't cover
Ramp is a corporate card and expense-management company. It tracks spending across more than 70,000 American businesses, so it sees which AI vendors actually get charged, not which ones get announced.
| Data point | May 2026 | July 2026 |
|---|---|---|
| Anthropic share of AI spend | 41% | ~44% |
| OpenAI share of AI spend | 39% | ~40% |
| Ramp customers paying for any AI | ~50% (March) | ~56% |
Anthropic took the lead in May and has held it. But Q3-to-date growth is running OpenAI's way, which is the whole story behind the headline.
Before you treat those percentages as gospel, note the limits TechCrunch itself lists:
- American businesses only.
- Excludes large enterprises that run spend management through providers like American Express.
- Covers corporate card and bill-pay spending, not every purchasing route.
- Skews toward tech, because Ramp is popular in Silicon Valley.
- Percentages only. No dollar amounts were shared.
TechCrunch is blunt about it: "This isn't a measure of the total market." It's a directional read, not a census. Treat it that way.
🔁 Eight weeks is the new switching cost
Look at the shape of the movement rather than the values. A lab ships a model, spend moves. Another lab ships, spend moves back. Ramp economist Ara Kharazian told TechCrunch that GPT-5.6 Sol is "really good, increasingly the choice for developers," while Fable 5 "disappointed both in adoption and real-world application."
That is a market where a two-month-old quality gap rearranges buying decisions. Which means the switch itself has become cheap: same chat-completions shape, same tool-calling pattern, same streaming semantics. Changing providers is a config change and a round of eval runs, not a rewrite.
Key takeaway: Nobody has moat-grade lock-in on the model layer today. Design for that, and every future price cut and capability jump is yours to take. Design against it, and you're volunteering for a rewrite you didn't need.
🛠️ How to build for a market that flips every quarter
Provider-agnostic does not mean a heavyweight abstraction framework. It means being disciplined about four things.
| Layer | Hardcode to one vendor? | Why |
|---|---|---|
| Model IDs | Never | Put them in config or env. They change more often than your code. |
| Prompts | No | Store as data (files or DB rows), not string literals scattered in handlers. |
| Tool / function schemas | No | Keep one JSON schema definition, map it per provider. |
| Vendor-only features | Isolate | Caching, batch, computer use — behind a flag, with a plain fallback path. |
A minimal seam is often enough:
type LLM = {
complete(prompt: string, opts?: { model?: string }): Promise<string>;
};
const provider: LLM = process.env.LLM_PROVIDER === "openai"
? openaiAdapter
: anthropicAdapter;
Then three habits that make the seam real:
- Keep an eval set. Twenty to fifty of your own prompts with expected outputs. Without it, "the new model is better" is a vibe, and vendor benchmarks are marketing.
- Log every call with the model ID. When quality shifts, you want to know whether it was your prompt or their silent update.
- Re-test on release, not on renewal. Run your evals when a new model drops, then decide with numbers.
💰 What lab churn does to your bill
This is the part that helps anyone earning in rupees and paying in dollars. Competition this close is what keeps prices falling and free tiers alive. When 56% of a card provider's customer base is already paying for AI and the leaders are four points apart, neither side can afford to price like a monopoly.
Practical moves for a small team here:
- Price the workload before you commit. Our AI token counter and AI model comparison let you cost a feature before you write it.
- Split by task. Cheap models for classification, extraction, and formatting. Expensive ones only where reasoning quality shows up in the output.
- Re-price quarterly. If the leaderboard moves every two months, your cost assumptions from January are already stale.
⚠️ The retention clause matters more than the leaderboard
One detail in the piece is worth more to Sri Lankan dev shops than the market share numbers: Anthropic warned users of its higher-end Fable tier that it must retain their data for 30 days under regulatory mandates. That caused a round of public anger, though the article notes the criticism was somewhat oversimplified, since Fable targets specific use cases rather than general chat.
If you're doing client work for a bank, a hospital, or a European customer, retention terms are a procurement fact. Before you ship, get answers to:
- How long is prompt and output data retained, and can that be turned off?
- Does the tier you're on differ from the tier the marketing page describes?
- What does your own client contract promise about sub-processors?
- Can you point to the vendor's written policy, not a blog post?
A model that scores two points higher and violates your client's data clause is worth zero. Check the terms before the benchmarks.
💡 What this means for you
- Don't pick a "winner." The lead changed in May and is being chased in August. Anyone telling you one lab has permanently won is selling something.
- Build a seam, not a framework. One adapter interface, model IDs in config, prompts as data. That's a day of work and it buys you every future price drop.
- Own your evals. Your twenty prompts beat any public benchmark for deciding what actually works on your product.
- Read the retention terms first, especially for client work under an NDA.
- Treat volatility as a discount. Two well-funded labs fighting over enterprise spend is the best pricing environment a small team in Colombo is going to get. Stay in a position to switch, and you keep collecting the benefit.
The investors being warned about weak stickiness are right to worry. For the rest of us building things, weak stickiness is the good news.
Top comments (0)