"Which AI do you use?" is really two questions. What we build with, and what runs inside the system once it's yours. The second one hides a bigger question every team should ask: should we be using open-source models at all?
What we build with: Claude Code writes, Codex reviews
Claude Code and Codex are both coding agents. They write and change software with one of us at the wheel. We use them for different jobs.
- Claude Code builds. It writes the feature, runs it, and fixes what breaks.
- Codex reviews. Before anything ships, Codex reads the change cold, the way a second engineer reviews a colleague's work. It flags bugs, missed edge cases and anything risky.
- We decide. A person reads both, fixes what matters, and only then does it go live.
Why two tools? Because they're built by different companies and trained differently, so they make different mistakes. One checking the other catches problems a single tool would wave through. It's a big part of why a workflow can be running in weeks, not months, without cutting corners.
It also plugs into what you already use. The code lives in your repository, and every review sits next to the change it checked, so your team can see exactly what was built and why.
What runs inside your system: closed or open-source?
Closed models are the ones you rent from a company like Anthropic (Claude) or OpenAI (GPT). Open-source models, like Llama, Mistral or Qwen, are ones you can download and run on your own servers. Both are good. They're good at different things.
Closed models: the upside
- The strongest at judgment and writing. For anything a client or investor will read, they're still ahead.
- Nothing to host. You pay as you go, and there's no hardware to buy or look after.
- They keep getting better without you lifting a finger.
- Business and API plans don't train on your data by default.
Closed models: the downside
- Your data leaves your walls. It's covered by a contract, but it does travel to the vendor.
- You pay per use, at a price you don't control.
- The vendor can change or retire a model, and its behaviour can shift a little when they do.
- If they have an outage, you wait.
Open-source models: the upside
- They can run entirely on your own infrastructure, so sensitive data never leaves.
- No fee per use. At very high volume, that adds up to real savings.
- The model never changes unless you change it. What worked in testing keeps working.
- You can fine-tune them on your own examples, and nobody can take them away.
Open-source models: the downside
- A step behind the best closed models on hard reasoning and careful writing.
- Someone has to host, update and watch them. That's work, and it's yours.
- Powerful hardware costs money up front. At low volume, renting a closed model is usually cheaper.
- Safety checks and guardrails are your job, not a vendor's.
So, should you use open-source?
For most teams, not for everything. We reach for open-source when one of three things is true:
- The data can't leave. A contract, a regulator or a client says so.
- The volume is huge and the task is simple. Tagging, sorting or cleaning hundreds of thousands of records.
- It must never change. A process where the same input has to get the same answer, year after year.
Otherwise we start with a closed model, because it gets you the best results fastest, and build so you can switch later.
The four questions we ask
- How sensitive is the data? If it can't leave your walls, open-source on your own infrastructure keeps it there.
- How much judgment does the task need? Drafting a careful email to an investor gets the strongest model we can use. Tagging ten thousand rows doesn't.
- How often will it run? A few times a day, cost barely matters. A million times a month, a smaller model can save real money.
- How fast does it need to be? Anything a person waits on should feel instant.
Most systems end up mixing them: a strong closed model for the judgment work, smaller or open-source models for the bulk.
Keep it swappable
Models get better every few months, on both sides. We build so that switching one out is a setting, not a rebuild, and our maintenance keeps an eye on when it's worth it. Picked for the job, not out of habit, and ready for whatever ships next.