At the beginning of June I decided to run an experiment: see if I can replace my $100 Claude subscription with something cheaper – but still keep the Claude Code harness, and at least a similar quality level.
Almost 2 months later, I’m happy to report that I successfully cancelled Claude, and I’m running a combo of Grok-4.5, GPT-5.6, MiniMax M3 and Open Code free models. Same level of productivity, maybe slightly better.
Total monthly bill: $80.
Here’s the breakdown: $35 Super Grok, $25 MiniMax M3 Token Plus plan, $20 vanilla GPT subscription (includes the Codex starter package).
Now, the limits for these, taken separately, are not great. Except MiniMax M3, the other 2 seem to eat tokens for breakfast. That’s just one of the things I discovered during the experiment. There are a couple more that I want to share, so, here are my thoughts on how you can actually run Fable level inference, with decent limits, for the monthly cost of one gas tank refill for your SUV.
If you just pick 2-3 entry-level subscription that offer coding, like ChatGPT, Grok, even Gemini or Cursor, you get a cost per token way smaller than for a premium subscription, but still with a decent token volume. In other words, you pay way less for more or less the same.
So, my choice is 2-3 entry level subscriptions, but combine them for different capabilities. Here are 2 options, first the one that I’m using, second a fictional one.
Option 1
- Super Grok – difficult reasoning, large code bases, very specific logic
- GPT-5.6 – cleaning up, understanding and planning
- MiniMax M3 – execution
This is $80, as I wrote above.
Option 2
- Cursor – planning, reasoning, complex codebases, logic
- MiniMax M3 – execution
The Cursor plan I’m talking about here is the $60 subscription, which gives you Grok-4.5 AND Cursor-2.5, whereas MiniMax M3 is the same $25 Token Plus. So here you’re at $85, but you get 2 powerful models in the same interface, with decent limits and a solid executor.
But there’s an important caveat: using this kind of setup is different. You don’t just prompt something and expect the model to telepathically understand your mind and one shot everything. Fable might give you this sensation, but it can also force you to take a second mortgage to finance the token consumption just for that single prompt.
Instead, you change a little bit your approach, using the guidelines below.
Aggressive Execution Specs
Before starting to work you actually run through your own notes (or ask some of the agents to do this for you) and mentally allocate tasks per model. I call this aggressive execution specs. This step matches models with tasks and keeps the token consumption constantly low, while maintaining the output decently high.
But it requires you knowing what you are actually doing and spending some time outside the chat window.
In-flight Model Switch
Be used to switch model mid-session. This is something you probably never did with Claude or Codex. You just assumed everything is done for you – and 99% of the time it was.
In the approach I’m using, this is not true anymore. You need to build the mental model of having a small group of employees which are each good at one set of tasks. And juggle them accordingly.
Free Models Are Now Quite Good
When I did my experiment, I briefly considered Kimi 2.7, but discarded the idea in favor of MiniMax M3. Just a couple of days ago, Kimi 3, a 2.8 trillion parameter model was launched and the feedback so far is overwhelmingly positive. The cost is a bit too high, the pricing is still per API token, but my point is free models now are getting very, very good.
You may even consider a 3rd option, which may look like:
- OpenCode: includes Grok 4.5, Kimi (K2.6, K3), GLM (5.1, 5.2), MiniMax (M2.7, M3), Qwen (3.6 Plus, 3.7 Plus, 3.7 Max), MiMo-V2.5 (Pro), and DeepSeek (V4 Pro, V4 Flash).
Total cost: $10/month. Allegedly, you would hit the limits pretty fast, but even if you add 3-4 extra layers to your balance, worth in total, let’s say, $40, you will still be at $50/month, and in very good company.
*Caveat: by “poor guy” I meant “smart, financially aware AI-amplified builder”.
dragos@dragosroua.com (Dragos Roua)
Source link