The Theme

Codex is no more, and I do not miss it. That is the theme for this week.

This is Week 97 of building Shokken, my restaurant waitlist management tool. It is available on Google Play and the iOS App Store, and the app itself is in good shape. What I want to talk about is not the app. It is the development journey, and specifically what happened after I let my Codex subscription lapse on the last day of October.

What I Expected to Miss

I expected to miss the Codex app. It is a genuinely polished product, and I do not think anybody would argue with that. The Hermes desktop app is anything but polished by comparison. It is rough. If you came from Codex you might reasonably conclude that you are downgrading.

So I went in expecting to resubscribe within a week. That is not what happened. This is the sixth day, and I have not once felt the urge to go back.

What I Actually Used Agents For

To be fair to Codex, I have not done much Shokken development in these six days. Most of what I needed was outside product work: server administration, research, and ordinary lifestyle questions like finding a product in an unfamiliar city.

That is where the surprise came from. Hermes has been consistently better at those tasks than the Codex setup I was paying for, and it is not because the model is smarter. Codex Sol 6.1 is one of the closer frontier models. It is large, it is expensive, and I could not get consistently good results from it.

The problem was never intelligence. It was instruction following. I would tell it, in the system instructions and in the agent instructions, to use specific tools and a specific memory setup. It would not check. It would assume. And when you have a custom environment, assuming is exactly how an agent breaks things.

Hermes has been the opposite. It does not arrive with strong opinions about how I should work, because it has almost no built-in instructions beyond the ones I give it. So when I ask it to use my OpenViking memory backend, or Exa for search, it does. There is no provider-side override quietly discouraging tool calls to save cost.

The Bill That Is Coming

Exa is good enough that I am using it far more than I expected. There is no local cache for search queries, so every search hits the API and is billable. I am still inside the free quota, but I can see the usage line, and I expect to cross it soon.

My rough estimate is a few dollars a month, maybe ten or twenty if I keep going. That is not a reason to stop. It is a reason to pay attention, because the current arrangement is unusually cheap and I know it.

DeepSeek 4.1 Flash is the other half of that arrangement. I run it through OpenCode and CommandCode, and both are handing out multipliers that are frankly ridiculous: something like eight times on one and ten times on the other. A modest top-up turns into close to two hundred dollars of usage. I am actually struggling to spend it.

None of that is permanent. The AI market is moving fast, and subsidies like this do not survive contact with a real business model. I am riding it while it lasts and assuming it ends.

The Failover

When the quota does run out, my failover is DeepInfra. I moved there from OpenRouter. DeepInfra does not charge a fee when you recharge the account, and its 4.1 pricing is among the best on the market for non-subscription API calls.

The annoying part is that OpenRouter offers DeepInfra as one of its providers at the same cost, and then adds a five percent surcharge every time you put credit in. I would rather not pay a middleman five percent for the same backend, so I went direct.

The Local Backup

There is a second layer under all of this that does not depend on any provider.

My local setup runs Halogen Qwen 3.8 Flash Next, the 4-bit quantized version, on an AMD 395 Plus AI Max chip with 128 GB of unified memory. Alongside it, a 5090 runs Qwen 3.8 27B quantized to 6-bit at the full 262K context.

That matters for two reasons. The obvious one is that even if everything goes dark, I still have local intelligence I can summon that is actually capable. The less obvious one is that DeepSeek itself can be hosted by other people. If my current provider starts behaving badly, I am not trapped; there are a million places to flee to.

That is the real goal here: not to replace one vendor with another, but to get to the point where no single vendor can hold my work hostage.

The Reviewer Trick

One thing worth passing on before I close. If you take an agent’s plan and send it to a reviewer, it almost always produces a better outcome than a single agent working alone. That holds even when the reviewer is the exact same model running on the exact same computer.

I use this most with models that are not state of the art. Qwen 3.8 Flash Next is a good model, but nowhere near DeepSeek 4.1. Asking it to review its own plan adversarially catches substantive problems that the implementer missed. You trade latency for quality, and for planning work that is usually a good trade.

What I Am Watching

I am not declaring victory. The real test comes next week, when I go back to actual development in front of an actual computer in my actual environment. That is the acid test for whether I can be fully off Codex.

I still want OpenAI to be a viable competitor. I do not want to go back, but I want the option to exist. And I hope Hermes sticks around long enough to make this stick.

If you are interested in the restaurant waitlist tool itself, visit Shokken. It is available on Google Play and the iOS App Store.