Three things happened this week that, taken together, tell you more about where AI is heading than any single product launch.
First, Ramp published its March 2026 AI Index and the numbers are striking: Anthropic now wins 70% of head-to-head matchups against OpenAI among businesses buying AI for the first time. Second, a16z dropped the sixth edition of its Top 100 Gen AI Consumer Apps report, revealing that ChatGPT and Claude's app ecosystems have only 11% overlap (read our full write-up on that here). They're becoming different platforms for different users. And third, Anthropic launched inline interactive visuals in Claude, a feature that seems small but signals something big about where these platforms are going.
We wrote about all three of these stories this week (the a16z data, the customization arms race between Claude and ChatGPT, and why everyone's talking about "harnesses"). This deep dive connects the dots.
First up, the TL;DR
You probably have a pretty locked-in preference between ChatGPT, Claude, Gemini, and Grok at this point. You might even feel a little defensive about it.
Turns out businesses feel the same way. And for the first time, we have spending data that shows exactly who's winning.
Ramp's March 2026 AI Index dropped this week with a headline number that stopped people mid-scroll: Anthropic now wins roughly 70% of head-to-head matchups against OpenAI among businesses buying AI for the first time.
- That's a complete reversal from 2025.
- Nearly one in four businesses on Ramp now pays for Anthropic (a year ago, it was one in 25).
- OpenAI's adoption rate fell 1.5%, its largest single-month decline ever.
The "why" is where it gets interesting: it's not Opus' benchmarks. Claude Code and OpenAI's Codex are roughly comparable products. Codex is arguably cheaper as Altman "floods the zone" by resetting usage limits. Meanwhile, Anthropic literally can't meet its own demand; every plan still has rate caps because they don't have enough compute (or, put another way, a pretty smart way to control costs imo).
But that's what makes this meaningful: a company charging more for similar performance, while actively turning away revenue, is growing faster. In most enterprise markets, the cheaper product wins. Not here.
Ramp economist Ara Kharazian thinks the moat might be cultural. The DoD backlash sharpened the contrast between the companies, and a "certain class of user noticed." He floated a provocative comparison: choosing between OpenAI and Anthropic might become less like enterprise procurement and more like the green bubble / blue bubble distinction in iMessage. A signal of identity, not just technology.
Meanwhile, a16z's sixth Top 100 AI Apps report shows the platforms diverging fast. ChatGPT and Claude both have 200+ apps in their connector ecosystems, but only 11% overlap. ChatGPT is going super-app with shopping, travel, and food. Claude is stacking developer tools and financial data terminals. Two platforms, two philosophies, two ecosystems.
And yesterday, Anthropic launched inline interactive visuals, letting Claude build charts, diagrams, and interactive visualizations right inside your conversation. It's on by default for all plans. It sounds like a small feature, but combined with skills, connectors, and plugins, it makes Claude less of a text box and more of a workspace where information is both processed and presented.
The model stopped being the bottleneck sometime last year. The system you build around it (what the industry now calls a "harness") is the real product. So the race has shifted. And the scoreboard is public now.
DEEP DIVE / FULL BRIEF ON HARNESSES AND CLIs HERE
The Numbers, Explained
Let's start with what Ramp actually measured, because the data is more nuanced than the headline.
Ramp tracks corporate credit card spending across tens of thousands of businesses. Their AI Index measures what percentage of companies are paying for AI services, which ones, and how those numbers change month to month. It's real spending data, not surveys or app store rankings.
Here's what February 2026 looked like:
- Overall business AI adoption: 47.6% (record high)
- Anthropic adoption: 24.4% of businesses (up 4.9% month-over-month, its largest monthly gain ever)
- OpenAI adoption: Still the most-used, but fell 1.5% (its largest single-month decline since tracking began)
- Google adoption: 4.7% (slight growth)
- xAI: Less than 2%
The most revealing metric: among businesses purchasing AI for the first time, Anthropic wins about 70% of head-to-head matchups against OpenAI. That means when a company that has never paid for AI decides to start, seven out of ten times they're choosing Anthropic over OpenAI.
A year ago, that ratio was reversed.
So Why Is Anthropic Winning?
Kharazian's analysis is worth reading in full, because he lands on an answer that would've sounded absurd a year ago: the moat might be cultural.
His argument goes like this. Each AI company has a distributional advantage:
- Google's is that it already sells to most businesses. Gemini comes bundled with Google Workspace.
- OpenAI's was that it was the consumer default. ChatGPT was where most people first encountered AI, and that consumer momentum carried into business adoption.
- Anthropic's was that it was popular with early adopters: the evangelists, the engineers, the "AI person" on your team.
What changed is that Anthropic leveraged its early-adopter base to go mainstream. The people who picked Claude first became the people who recommended Claude to their companies. And the DoD controversy turbocharged this. When Anthropic refused to give the Pentagon unfettered access to Claude and OpenAI signed its own deal days later, a certain class of user noticed. Katy Perry switched to Claude. A US Senator did too.
Kharazian floated a wild comparison: choosing between OpenAI and Anthropic might become like the green bubble / blue bubble distinction in iMessage. A signal of identity as much as a choice of technology.
That sounds absurd for enterprise software. But the spending data says something like it is already happening.
Our Theory: It's the UX, Stupid
We think there's another layer to this, which we've said before but is worth repeating: nobody is innovating in the general business user experience better than Anthropic right now.
- Google's Workspace updates are too slow and too spread out. Gemini should be able to do everything you need, but it feels clunky to use. They have so many cool tools spread across all their different labs products, but what are you going to do, bookmark 100 things? Claude can do it all.
- Microsoft is trying to sell you 50 different Copilot plans when what you really need is a single interface that works as good as Claude and Cowork. So... why not just use Claude?
- OpenAI has Codex, which is close to what businesses need, but suffers from the same fate early Claude Code did: putting "Code" in the name scares off normies. There is their rumored Cloud workspace that could drop any day now though, so the word on the street goes...
And this week, Anthropic added another piece. Claude now builds interactive charts, diagrams, and visualizations directly inside conversations. Not in a side panel. Not as a downloadable file. Right there, inline, interactive. Ask Claude to visualize compound interest, explain a system architecture, or diagram a workflow, and it builds something you can click around in on the spot.
It's on by default for all plan types. You can prompt it directly ("draw this as a diagram," "visualize how this changes over time") or Claude will proactively create visuals when they'd help explain something.
This sounds like a small feature. Combined with skills, connectors, and plugins, it makes Claude less of a text box and more of a workspace. Information gets processed AND presented. That's the kind of UX gap that's hard to articulate on a spec sheet but obvious the second you use it.
The Platforms Are Diverging
The UX story becomes clearer when you zoom out to the ecosystem level.
We broke down a16z's sixth Top 100 Gen AI Consumer Apps report this week, and one chart tells the whole story: ChatGPT and Claude both have 200+ apps in their connector ecosystems, but only 11% overlap. The shared apps are the obvious productivity stack everyone needs (Slack, Notion, Figma, Gmail). Beyond that core, the platforms split:
- ChatGPT has 85+ apps across Travel, Shopping, Food, Health & Wellness, and Entertainment. Booking flights on Expedia. Ordering groceries through Instacart. Browsing Zillow listings. It's the most aggressive play any AI company has made to become a consumer super-app.
- Claude skews toward financial data terminals (PitchBook, FactSet, Moody's), developer infrastructure (Sentry, Supabase, Snowflake), and science tools (PubMed, Benchling), plus a growing open-source MCP community.
- Gemini is carving its own niche through creative tools, with traction nearly perfectly correlated to model releases like Veo 3 and Nano Banana. Google's also integrating AI across Workspace, but that growth is captured by existing products rather than a new experience.
Sam Altman has hinted at a "Sign in with ChatGPT" identity layer where users carry memory and tokens to third-party apps, effectively making ChatGPT the starting point for everything. If that plays out, this race may look less like the search wars (one winner takes 90%) and more like the mobile OS wars: two platforms with very different philosophies, both building trillion-dollar ecosystems.
The per-capita adoption data from the same report adds another layer. a16z built an index combining web visits and mobile MAUs across all 100 products:
- #1: Singapore
- #2: UAE
- #3: Hong Kong
- #20: United States
Why is the US so low? Two factors: a huge chunk of American jobs where AI hasn't really touched yet (retail, transportation), and cultural trust in AI sitting at just 32% per Edelman's survey. Top-adopting countries range from 50-80%.
The Real Product Is the System Around the Model
Here's where all three stories converge.
The Ramp data shows businesses picking sides. The a16z data shows the ecosystems diverging. And the harness conversation, which we wrote about separately this week, explains why switching costs are about to get very real.
"Harness" is the industry's new word for everything built around a model to make it useful: permission systems, context management, tool orchestration, lifecycle hooks, and project-level instructions. We wrote a full explainer on what harnesses are and how they work, but here's the short version.
Think of the model as an engine. The harness is the car built around it. Steering, brakes, GPS, seatbelts. It decides what the model can access, when it needs human approval, how it manages memory, and what happens when something goes wrong.
Every harness has five core components:
- Permission systems that control what the agent is allowed to do.
- Context management that keeps working memory focused (summarizing old context, spinning up sub-agents).
- Tool orchestration that defines which external tools the agent can reach (this is where MCP, the open standard for AI-to-tool communication, and the new format of Skills comes in).
- Lifecycle hooks that trigger automated actions at specific moments (run a linter after every edit, format code on save).
- System-level instructions via project files like CLAUDE.md or AGENTS.md that teach the agent your specific rules.
The data backing this up is striking. OpenAI built a million-line production app with Codex where zero lines were written by humans; the engineers spent all their time designing the harness. LangChain's coding agent jumped from 52.8% to 66.5% on a benchmark by changing nothing about the model, only the harness. Vercel got better results by removing 80% of available tools from their agent.
The model's intelligence is table stakes. The system you build around it is the differentiator.
What This Means for Switching Costs
This is the part that ties everything together and explains why the platform war matters to you personally.
Every skill you install, every connector you link, every project file you write, every memory the system builds... that's infrastructure. And as a16z partner Olivia Moore put it: "context and memory compound."
We mapped the customization stacks of both Claude and ChatGPT this week. Here's what each platform offers:
Claude's Customize tab (Settings → Customize) groups four things:
- Skills teach Claude how to do something (instruction files loaded on demand).
- Connectors link Claude to external tools via MCP (Google Drive, Slack, Salesforce, etc.).
- Plugins bundle skills, connectors, and slash commands into one installable package.
- Inline visuals (new this week) build interactive charts and diagrams right in the conversation.
OpenAI's Codex + ChatGPT offers a parallel stack:
- Codex Skills work like Claude's (bundled instructions and scripts); but where's skills inside ChatGPT? Sad face emoji...
- Automations schedule Codex to run recurring tasks in the background
- ChatGPT Apps (powered by the Apps SDK and MCP) let third parties build interactive experiences inside your chat (Spotify playlists, Zillow listings, Canva slide decks).
- Side note: The difference between ChatGPT apps and "Connectors" which ChatGPT also has is a bit confusing... isn't an "app" basically an MCP server? Will that just replace Connectors? And what about Skills? We think ChatGPT's UX is in transition, and we're waiting for them to land on some kind of "V2 architecture" to fix it up a bit. It has lots of legacy stuff in there right now that they can definitely simplify, which is why Codex is so refreshing...
Despite the differences, the architecture is converging on the same open standard (MCP), which means integrations will eventually be portable between platforms. But the configurations, the memory, and YOUR human muscle memory of how you work with one platform vs. another... that's personal. That's sticky.
Once you've spent a few weeks teaching Claude how your team writes code reviews, or built a ChatGPT workflow that handles your weekly reporting, moving to the other platform means rebuilding all of it from scratch. That's the real lock-in.
Where This Goes Next
A few predictions based on what this data implies:
- The next six months will be defined by voice and memory. Moore identified these as the two biggest trends. Voice dictation is becoming normalized in tech companies and will spread to mainstream consumers within 6-9 months. And memory could become the single biggest competitive advantage. As Moore put it: "Any product that you start to use two years from now, if it doesn't immediately feel like it knows you, it will feel broken."
- Agents will accelerate the divergence. The more autonomous your AI becomes (scheduling tasks, managing inboxes, executing multi-step workflows), the deeper it embeds into your work. Switching an assistant is annoying. Switching an agent that runs your morning briefing, triages your inbox, and manages your calendar is a week-long project.
- "Harness engineering" will become a real job title. Phil Schmid, Hugging Face's former technical lead, called harnesses "the most important concept in 2026." Companies are already hiring for roles that are essentially "make our AI systems reliable," even if the title still says "ML engineer" or "AI platform lead."
- The winner won't be the smartest model. It'll be the most configured platform. The company that makes it easiest to set up skills, connect tools, and build memory will win the general business user. Right now, that's Anthropic. But OpenAI's rumored Cloud workspace, Google's Workspace integration, and the 15+ CLI tools flooding the developer market all suggest this race is far from over. There's also a real case to be made how long the "general business user" will be a valuable target market, if agents take off the way these companies are pitching...
Anyway, if you haven't configured your AI setup yet, we highly encourage you to check out part one of this 15-minute playbook to customize your AI setup as a good place to start. The longer you wait, the longer you lose out on quality capabilities that could take you from your current level of manual arm wrestling with the machine, to having the machine actually DO work for you.
