What Happened to the AI Race? Who Actually Codes With These Models
My take as a developer who uses AI in production every day: why only two companies compete at the coding frontier today, why benchmarks aren't enough, and why you should never marry a model.
In this article
Short answer: in my opinion, the AI race for coding is being run by two companies right now: OpenAI and Anthropic. Meta and Google were pioneers, but in my day-to-day work I don't see anyone shipping production code with their models. Their models do work well inside their own ecosystems. Chinese models post a lot of numbers, but for production work, in my experience, it's mostly smoke. My rule: use whichever frontier model wins your own tests, and never marry any of them.
This is an opinion piece, not a ranking or a benchmark. It's what I see working with AI on real projects every day, and I'm writing it because people keep asking me what happened to the companies that were "supposed to win."
Why did Meta and Google fall behind if they were pioneers?
Meta and Google were among the first in this race. Nobody can take that away from them. But being first isn't the same as being useful today.
What I see in the field is simple: nobody around me codes with their models. Not me, not the developers I work with, not the people I follow who show their real workflow. A model can look great on a benchmark chart, and reality can still be something else. Shipping production code isn't answering an exam question. It's reading a whole repo, touching ten files, running commands, breaking something, noticing, and fixing it.
A good model isn't enough for that. You need a harness.
What's a "harness," and why does it matter more than the benchmark?
By harness I mean the tool wrapped around the model that lets it actually do the work: read your code, edit files, run commands and keep iterating. The two I use every day are Claude Code and Codex.
What they do isn't my impression; it's in their official docs:
- Anthropic describes Claude Code as an agentic coding tool that reads your codebase, edits files, runs commands and integrates with your development tools, available in the terminal, your IDE, a desktop app and the browser.
- OpenAI pitches Codex CLI as a way to inspect, edit and run code from your terminal.
Here's the opinion part: Meta and Google don't have a harness at that level today, or at least not one I see people using for real work. And without a harness, the best score in the world stays on the chart.
There's another piece people rarely mention. A harness doesn't just help the user; it generates real usage, and real usage is what lets a model get better at real tasks. Whoever has thousands of developers in their tool every day learns faster what breaks in production.
Is Gemini 4 Argon good for coding?
In my opinion, not today. I'm speaking from my own experience, not from a scoreboard. For my coding work, Gemini 4 Argon isn't something I use or recommend right now.
I think Google is behind because it doesn't have as much data from developers coding with its model to help it improve. Same loop as before: less real usage, less learning about what actually fails.
That said, I'm not writing anyone off. Google could ship a good one tomorrow. It has the resources. If it does, I'll test it the way I test everything: on my own tasks, not on launch-day charts.
So what are Meta's and Google's models good for?
Plenty, inside their own ecosystems. That's the part that gets lost when everything is framed as "who's winning."
- At Meta: search and images inside their products.
- At Google: meeting summaries, search inside Gmail and Docs, the kind of built-in help in tools millions of people already use.
Those are real, useful jobs. They're just not the frontier. Summarizing a meeting is one thing; leaving an agent to work on its own against a production system is another.
Why is coding the engine of AI?
Because it's where AI does work you can check: the code compiles or it doesn't, the tests pass or they don't, the system runs or it goes down. There isn't much room for smoke.
As I see it, AI has three engines right now: coding, science and cyberdefense. And at that frontier, from what I see, only OpenAI and Anthropic are really doing the work. Everyone else is playing in other leagues, which are legitimate too, just not this one.
My own tool split reflects that:
- Claude for coding.
- Codex for computer automation.
- Grok Bot for everything social media.
What about Chinese models like Kimi?
Kimi, from Moonshot, and other Chinese models post a lot of numbers: charts, comparisons, viral threads. I've tried it, and I'm not saying it's bad. I'm saying that, in my experience, for production it's smoke.
Why? Because almost nobody shows what actually matters: how much it costs and how long it takes to get that result. In a test of my own, same prompt across several models, Kimi took several times longer than the fastest one. A benchmark that doesn't tell me what it cost or how long it took doesn't help me decide anything.
My rule: always use the frontier model that comes out on top in your own tests. Not the influencer's tests. Yours, with your tasks and your code.
Why shouldn't you marry a model or a company?
Because this changes all the time. The model I use for everything today might fall behind next month. If I marry a company, I'm stuck with its model even when it's no longer the best, out of habit or loyalty.
I do the opposite: I test every major release on real work, compare cost, time and quality, and switch without drama when something is better. Right now my main tools come from OpenAI and Anthropic because they win my tests. If someone else wins tomorrow, I switch.
Never marry a model or a company.
Are subscriptions subsidized?
In my opinion, yes, heavily.
I use CodexBar, an app that tracks Codex and Claude Code usage and, according to its official site, can run local cost scans on that usage. Based on what it shows me, my usage works out to roughly 18 to 30 thousand dollars a month at API prices. That's my own estimate from the tool, not an invoice.
Here's what I actually pay:
- Claude: $200 a month.
- Codex: $200 a month.
- Grok Bot: $30 a month.
The gap between what I pay and what the API would cost is huge. That's why I believe subscriptions are subsidized today: frontier companies are paying for developers to use their tools every day. It makes sense, because that usage is exactly what makes them better.
One fact that is official: Claude's Max plan comes in two tiers, Max 5x at $100 a month and Max 20x at $200 a month, and Claude Code is included in paid plans, drawing from the same usage limits. Anthropic's pricing page also notes that for heavy coding sessions you can switch to pay-as-you-go API credits.
My practical advice: if you code every day, take advantage of the subsidy while it lasts. I don't know how long that will be, and neither does anyone else.
FAQ
Are you saying Gemini 4 Argon is bad? I'm saying that, in my opinion and for my coding work, it doesn't do the job for me today. I don't publish scores or percentages about it. And I think Google could ship a good model at any moment.
Are benchmarks useless? They're a reference, but they're not enough. If they don't say what each task cost and how long it took, I can't decide based on them. What decides is how the model performs in your own tests.
What's the exact number CodexBar shows? I'm not giving an exact number. It's my estimate: 18 to 30 thousand dollars a month at API prices, based on what I see in the tool. What is exact is what I pay: $200 for Claude, $200 for Codex and $30 for Grok Bot, per month.
Which model should I use? Whichever frontier model wins your own tests on your real tasks. For me today, that's Claude for coding, Codex for computer automation and Grok Bot for social media. Tomorrow it could be a different mix.
Sources
- Claude Code, official docs (Anthropic): what Claude Code is and does.
- Codex CLI, official docs (OpenAI): inspect, edit and run code from your terminal.
- Claude pricing (Anthropic): Claude Code included in paid plans, plus the API-credits option for heavy coding sessions.
- What is the Max plan? (Claude Help Center): Max 5x at $100/month and Max 20x at $200/month.
- CodexBar, official site and official repository: usage tracking for Codex and Claude Code, with local cost scans.
- The author's opinions, experience and estimates, current as of October 24, 2026. Opinions about Gemini 4 Argon, Meta and Kimi are personal and not based on benchmarks.
