On Tuesday we published a count of the AI coding tools companies name in their job posts, and the result (Cursor and Claude Code, each ahead of Copilot) kept nagging at us, because there was an obvious follow-up sitting one layer down: fine, that is what engineers code with. What about the models companies actually build on?

Fair question. The AI editor in a job post tells you how a team works. The model API in a job post tells you what the product is made of. Different signal, arguably a bigger one, because it comes with a roadmap attached.

So we counted again. Same method as last time: take every company on Remoet that is currently hiring (867 of them this week), read the tech stack we extract from their own job posts, and group every model and model-API tag by vendor family. Coding tools excluded this time. Claude Code, Cursor, Copilot and friends do not count here; this is Claude-the-model, GPT-the-API, Gemini-the-platform.

If you spend any time on tech Twitter you already know the consumer ranking. ChatGPT is a household name with hundreds of millions of weekly users, Gemini ships inside every Google surface, and Claude is the one your non-tech friends have maybe heard of.

The job posts rank them in the opposite order.

One in four hiring companies now names a model vendor

220 companies name at least one of these families. That is one in four of everyone currently hiring, and it undercounts the trend, because another chunk of companies writes the generic version instead: the plain tag "LLM" appears at 251 companies, nearly one in three. Two years ago that line barely existed. Now it sits in stack lists the way "REST APIs" does.

But the ranking is the story. Claude, the model with a fraction of ChatGPT's consumer footprint, is the one companies write down most. 169 hiring companies name Claude or Anthropic in their job posts, against 135 for the OpenAI family and 80 for Google's.

We double-checked this one, because it cuts against the headline narrative. It held. The exact tag "Claude" alone shows up at 126 companies, which puts it above Cursor (113), TensorFlow (110), and MongoDB (105) in our whole tag universe. A model API that did not exist four years ago is now named in more job posts than a database that has been around since 2009.

We also went looking for someone else who had run this count, expecting a bigger dataset to cross-check against. As far as we can find, there is not one. Indeed's Hiring Lab reports that about 74% of AI-related postings just say "AI" and only 2% name ChatGPT (no other vendor gets broken out at all), and Lightcast tracks generative AI as a skill category, not by vendor. So the table above is, to our knowledge, the first published vendor ranking built from what companies write in their own job posts. If someone beat us to it, genuinely, send it over.

The Indeed number is worth sitting with for a second anyway: across the wider market, naming any vendor is rare. Most companies write the generic line and move on. Which is exactly why the ones that do put a model name in writing are worth reading closely. It is a choice.

And no, this is not the coding tool leaking into the count. We measured the coding tools separately in Tuesday's post and excluded them here. 119 of the 169 Claude-family companies do not tag Claude Code at all. They are not telling you what editor to use. They are telling you what the backend calls.

Who writes Claude into a job post

Not the AI labs, mostly. The list looks like the regular software economy: Databricks, Snowflake, Datadog, Reddit, Pinterest, Block, Ramp, Robinhood, Intercom, GitLab, Asana, JetBrains, Affirm, Vanta, Oura, Toast.

Fintech, data infrastructure, consumer apps, a smart ring. When a company like Toast (restaurant software) puts a frontier model in a job post, the thing has stopped being a research artifact and become a line item.

The OpenAI column has its own marquee names (Stripe, MongoDB, Zscaler, plus most of the same data companies), so nobody should read this as OpenAI being absent from the stack. It is very much there. It is just no longer the default answer, which, given where the two companies were in 2023, is the actual news.

My favorite row in the raw data: Cohere, itself a model lab, names both Claude and OpenAI in its own job posts. Even the people training models are hiring people who can drive the other guys' models. (OpenAI also tops its own family's list, which is the vendor-in-own-list effect we flagged in the coding-tool post. Mentally subtract one from each column. The gap does not care.)

Nobody is marrying one lab

105 companies name both Claude and OpenAI. That is 62% of the Claude crowd. Only 64 name Claude without any OpenAI tag next to it.

So the realistic picture inside these teams is not "we are an Anthropic shop." It is a router, or at least a bake-off: one model for the long-context work, another for the cheap classification calls, something behind an abstraction layer so they can swap when the next release lands. Databricks, to pick the biggest listing, tags all three families at once.

For a job seeker that detail matters more than the ranking. The skill being screened for is not loyalty to a vendor. It is knowing how to evaluate models against each other and when to switch (which is a genuinely different skill than prompt tinkering, and much rarer in the wild).

The open-weight no-show

Eight companies. Out of 867.

The entire visible footprint of the open-weight world (Llama, Mistral, Qwen, DeepSeek, Grok) in our job-post data, and one of the eight is Mistral AI itself, so call it seven. DeepSeek, which dominated the discourse for a solid month, appears at exactly zero companies. Grok, zero.

We want to be fair to this number, because it makes open models look deader than they are. Teams that self-host models tend to write the serving stack into the post instead of the weights: vLLM shows up at 26 companies and Hugging Face at 30. If you count those as open-model proxies, the real footprint is maybe forty-something companies, still a rounding error against 169.

The gap between how much oxygen open models get in the discourse and how rarely anyone hires for them is the single most surprising thing we found in this count. Whatever "open source is eating AI" means, it does not currently mean jobs.

Why the job posts disagree with the app store

Because consumer AI and enterprise AI quietly stopped being the same market, and most of the coverage still only watches the first one.

On the consumer side there is no contest. OpenAI announced 900 million weekly ChatGPT users back in February, and third-party trackers now put the app around a billion monthlies. Claude's consumer app is somewhere in the tens of millions, an order of magnitude and change behind. (One funny detail from the Sensor Tower data: Claude makes more money per user, $2.76 against $1.74. Fewer people, more of them paying.)

The enterprise side flipped a while ago. Menlo Ventures' end-of-2025 enterprise survey put Anthropic at 40% of enterprise LLM API usage, OpenAI at 27%, Google at 21%, with OpenAI down from a 50% share two years earlier. For coding workloads specifically the survey had Anthropic at 54%. Our job-post count arrives at the same ordering from a completely different direction (nobody got surveyed; companies just wrote their stacks down), which is exactly what you want two independent measurements to do.

The same survey has open-source models falling from 19% to 11% of enterprise workloads year over year. Our eight companies are the hiring-side shadow of that decline.

And the money agrees with the job posts, not the app store. Anthropic disclosed a $47 billion annualized run-rate in May, with reporting around it putting roughly 80% of that revenue on business customers. Companies do not write a vendor into a job post because the app is popular. They write it there because a budget line and a backlog come attached, and right now more of those backlogs say Claude.

What to do with this if you are job hunting

A model API in a job post is a skill line you can prepare for, same as a database. Concretely, from the tag data:

• "LLM" at 251 companies and a named vendor at 220 means model-integration work has gone mainstream well beyond the AI-native startups. If your last two projects have no LLM surface at all, a third of the market is quietly starting to read that as a gap, senior roles included.

• The orchestration layer is hiring too: LangChain, LangGraph or LlamaIndex appear at 117 companies, RAG as an explicit tag at 80, and MCP (the protocol this site runs on, hello) is already named at 48. These are the tags that separate "has called an API once" from "has shipped the thing."

• Multi-model experience beats single-vendor depth. The 105 both-vendors companies are telling you their stack is plural. Being able to say "we moved a workload from GPT to Claude and here is what changed" is exactly the sentence those interviews want.

The cut a keyword search cannot give you

"Companies that name Claude, also run Go, and are hiring at mid level" is the kind of question this data exists to answer, and no job board search box will take it.

Connect the agent you already use (Claude, Cursor, whatever speaks MCP) to Remoet and ask it directly:

• "Which hiring companies name Claude or Anthropic in their stack?"

• "Of those, which also run TypeScript?"

• "Star the ones with mid-level roles open."

It reads the same tag-level stacks this post is built on, not the marketing copy. Star the handful that genuinely fit your stack and you get one email a week when they post something new. You decide where to apply; the agent just stops you from bookmarking forty tabs you will never reopen.

Where this data is soft

Same honesty section as always, because these counts bend in known ways.

Named in a job post is not a production-dependency audit. It measures what companies chose to write down when describing the work, which is a real, deliberate signal (someone typed it), but a floor, not a census of what runs in their clusters.

The OpenAI number in particular is probably a floor. ChatGPT is close to being the generic word for this technology, and generic defaults get written down less (the same argument we made about Copilot on Tuesday). Lightcast even measured ChatGPT mentions in postings falling through 2025 as demand shifted toward agentic skills, so some of Claude's lead is the newer, more deliberate name being the one worth stating. We cannot fully separate those two effects from tag data alone.

Tag families involve judgment calls. We grouped Vertex AI under Google and excluded Gemini CLI, Codex and Claude Code as coding tools; move those lines and the counts wiggle by a few, not by enough to reorder anything.

And this is the 1,073 companies on Remoet in early July 2026, tilted toward tech companies that post detailed engineering roles, not the entire economy. Counts are distinct companies, not seat counts or spend.

What survives all of that: a quarter of hiring tech companies now tell you, in writing, which frontier model they build on, and the one they name most is not the one with the users.

If your team names a model vendor in job posts, I would love to know what actually drove the choice (bake-off? one killer workload? a sales call?). Come tell us on Discord. And if you think our tag grouping put a model in the wrong family, say so and we will rerun the numbers. Hope this one was useful, thanks for reading!