My agent fleet picks its foundation model every week, and switching costs me almost nothing.
The fleet is a few Hermes agents that talk through a Telegram group. The one that runs this routine sits on an always-on Google Cloud VM, an e2-medium with 2 vCPU and 4 GB of RAM in us-central1. It has no GPU; the models are served through Nous Portal managed inference. Changing the default model is a one-line config value, not a reinstall.
Every Sunday a scheduled job reads the fleet’s real usage for the last 28 days: input tokens, output tokens, reasoning tokens, and cache-reads, which is the number that actually moves the bill. It prices the candidate models against that workload at their current rates and comes back with a verdict. Either a different model is cheaper for what I actually run, or the model I’m on has been retired. I change one line in config and I’m done.
The last switch looked wrong on paper. DeepSeek V4.1 Flash lists a higher input and output price than the V4 Flash it replaced, so a quick look at the pricing page says pass on it. My workload is roughly 94% cache-reads, V4.1’s cached token rate is about a third of V4’s, and that outweighed the listing prices. In practice it runs about 20% cheaper and it is the newer model. A pricing page would not have shown me that, but the weekly pricing run did.
Here is what that run compared with a month of my usage:
| Model | Input /1M | Output /1M | Cache-read /1M | Projected monthly |
|---|---|---|---|---|
| DeepSeek V4 Flash (retired) | $0.05 | $0.16 | $0.0135 | ≈ $32 |
| DeepSeek V4.1 Flash | $0.11 | $0.51 | $0.0040 | ≈ $26 |
| DeepSeek V4.1 Flash 1017 | $0.15 | $0.60 | $0.0030 | ≈ $31 |
My input is roughly 94% cache-reads, so the cache-read rate moves the total more than the headline prices. The newer, higher-listed V4.1 Flash still comes in cheapest because its cache-reads are roughly a third of the retired model’s.
What this costs next to a subscription
My managed-inference bill lands close to the price of one individual AI subscription.
| Who / plan | Price | What it covers |
|---|---|---|
| Claude Pro | 17/mo annual) | Everyday interactive use, per person |
| ChatGPT Plus | $20/mo | Interactive use, per person; API billed separately |
| My fleet (DeepSeek V4.1 Flash) | ≈ $26/mo | Programmatic agents on real usage, round the clock |
| Claude Max | $100/mo | About 5x Pro usage, plus $100/mo in API credits |
| ChatGPT Pro | from $100/mo | Higher usage tiers, plus up to $500/mo in API credits |
Prices from Claude plans and ChatGPT plans.
These are not the same product, and that matters. Claude Pro and ChatGPT Plus are per-person interactive subscriptions, and their API access is billed separately. Claude Max and ChatGPT Pro bundle monthly API credits, which gets closer to the game I am playing, but they are still seat licenses more than a meter for an always-on fleet.
What still lands is this: my agent fleet, doing real work around the clock, runs on managed inference for about as much as a single-person subscription and less than a quarter of an all-day plan. I pay for the tokens I actually use, not for a seat I might not fill.
The saving is not the point. What matters is that the ritual exists, and it only exists because switching costs are basically zero. I have no stake in any one provider’s model. I am a volume customer whose load gets repriced every week and routed to whoever the numbers favor.
That is the uncomfortable part for the foundation model providers, and I think they should take it seriously. The default model is not a durable moat. Any shop running agents at volume can do what I just did, and the tooling to re-evaluate and swap is becoming routine. What survives is quality per dollar, cache economics, token efficiency, and useful capability. Those are exactly the numbers a weekly cost run prices, so being already installed is not enough to keep your model the default.
I am carrying the same rule one tier down into my own hardware. I have a self-hosted model run set up for the days I do a full stretch of dedicated work and will keep an accelerator working all day. The managed tier stays the default for everything in between, when I need a model on and off and do not want an accelerator running 24x7 to cover a few minutes here and there. The managed tier is the right shape for that intermittent use, and the accelerator earns its running time only when the work is steady enough to fill a day.
The easy switch was between models on a portal. The next experiment is pointing the same weekly pricing at the model on my own accelerator, and letting whoever actually carries the work win. I am curious how far down the stack that same re-pricing ritual reaches.