I feel like OpenCode Go ($10/month) is a scam. Th...

# random

[Stefan Oltmann]

I feel like OpenCode Go ($10/month) is a scam.

The docs say the 5-hour limit is $12 in usage, the weekly limit is $30, and the monthly limit is $60. They also say they aim to provide 6x the value of the subscription price.

After using Kimi K3 for three mediocre tasks, OpenCode's context panel says the cost was $3.05 for 3 user messages, 64 assistant messages, 539 input tokens, 131 output tokens, and 165k tokens overall.

This depleted 100% of my 5-hour limit, 40% of my weekly limit, and 20% of my monthly limit.
That doesn't add up to me at all.

If I extrapolate from the monthly usage, it seems like I'm getting roughly $15 worth of usage for the $10 subscription, not $60 worth.

What am I misunderstanding here?

[אליהו הדס]

Kimi K3 is not suited to this kind of subscription; this kind of subscription was made for open-source models which were previously rather small and energy-efficient, whereas Kimi K3 is a monster.

[אליהו הדס]

I think that currently the most profitable and capable models are the Grok models, and I would even bet that they are the ones who will win the AI ​​race.

[Stefan Oltmann]

Interesting input. 👍

Do you use Grok 4.5?

It's the first Grok model I see on OpenRouter's list when I sort by intelligence.

openrouter.ai/compare/…/kimi-k3

[אליהו הדס]

Never use Open Router for agent use!

[Stefan Oltmann]

Oh, why?

[אליהו הדס]

Because they use multiple providers, in agent-based usage, it's 95% input versus 5% output, but using multiple providers doesn't guarantee the input cache price, so you end up paying much, much more.

[אליהו הדס]

Furthermore, subscribing directly from Grok is significantly cheaper than paying with tokens.

[אליהו הדס]

Yes, I use it and I'm migrating everything to it because it's much faster.

[אליהו הדס]

Speed ​​is more important to me than intelligence.

[Stefan Oltmann]

Not so sure about that. With less intelligence they introduce more subtle bugs, which gives me more headache later on. 😕

For all the times I liked most what Opus 5 produced after trying cheaper models.

But there is only sooo little Opus in the $20 subscription.

[Stefan Oltmann]

In my world an model can take all the time it needs if the results are good.

[אליהו הדס]

I rarely let the agents run around on their own; I read everything they do live. I'm the one with the intelligence; I need them to write. And I think Grok 5 is going to be on par with Mythos anyway, but I think Grok 4.5 is on par with Opus 4.8. I don't really see any difference from Opus 4.5 until now.

[אליהו הדס]

However, my mental workload has increased tenfold by the same factor as my productivity; I've already had 4 or 5 burnouts, and I've just come out of a week's break without reading a single line of code.

[Stefan Oltmann]

So sorry to hear that.

I can definitely confirm that you're extremely productive from the work you post here. 💪

But you'd still be godlike even if you slowed down your productivity by 50% or more. 😄

[אליהו הדס]

Yes, I need to learn to stop; that's definitely the hardest challenge.

[Stefan Oltmann]

I rarely let the agents run around on their own; I read everything they do live.

I use a more capable model to create a TODO list with small, incremental steps that I can commit one by one, just like I used to work.

Then I ask the AI to implement step X from my TODO.md and review the diffs once it's done.

If that takes a few minutes, that's fine. I can work on another project or do some Figma designs in the meantime.

[אליהו הדס]

An agent during a demanding task can make 50 or 100 turns; just one bad turn can ruin everything else. A single bad decision is enough to make all the others bad, so I prefer to check them live.

[Stefan Oltmann]

How do you check them live?

[אליהו הדס]

I'm simply watching what he's doing

[Stefan Oltmann]

You read the "thinking" logs live?

[אליהו הדס]

he or it, I don't know 🙂

[Stefan Oltmann]

I read those mostly to understand where it loses time/tokens - like figuring out each session that they need to run the unit tests without build cache. 😄

[אליהו הדס]

The agents cheat extensively; they dodge most problems, pretend they're resolved, take shortcuts, and constantly write fallbacks so that every error goes undetected.

[אליהו הדס]

Regardless of the model, they all do this; regardless of the prompt, they always end up forgetting this instruction.

[Stefan Oltmann]

That's what I see, too. Most of the time.

I find Opus 5 to be the first one that tackles harder issues, but I needed to write a lot of instructions.

[Stefan Oltmann]

But yes, even that model sometimes skips instructions.

[Stefan Oltmann]

I find it amusing when AI tells me "Yes, it's in the guidelines, but I forgot it".
So human. Strange.

[Stefan Oltmann]

AI is king of subtle bugs.

[אליהו הדס]

And often he makes bad decisions; he's good at executing tasks, not at leading. He might be good at reasoning about how to solve a problem, but he can't lead a project. In a project, all decisions must converge, must follow strict logical rules. It's not a matter of logic, it's not verifiable; it's simply about building blocks that must work together. Current agents are light-years away from doing this.

[Stefan Oltmann]

I feel the same.

[Stefan Oltmann]

And they also don't understand good UI/UX.
Their prototypes are barely usable.

[אליהו הדס]

In fact, they simply don't understand anything.

[אליהו הדס]

For example, I'm currently forking Compose Fluent, and I'm checking the to-do list. It shows that there's an inconsistency between the calendar targets; depending on the target, the first day of the week isn't the same for every region. The author wrote a manual table for JavaScript and left the bug on the other targets. The agent trusts the author, but I don't. I discovered that this bug only appears in JDK 17 and was fixed upstream in JDK 21, so it's not my problem. The iOS code was simply bad, and I need to think of a real solution in JavaScript. That's the difference between completing the logical sequence of words that follow and using human reasoning.

[Stefan Oltmann]

Using Opus 5 through OpenCode & OpenRouter is a waste of money.

It took $3 to write a consolidated AGENTS.md bundling all the instructions and then when it actually should write the result it always came out with "Tool execution aborted". 🤯

[Stefan Oltmann]

DeepSeek V4 Flash is really fast. 👀

[אליהו הדס]

It is practical and very inexpensive for use in a server / app via API, but not intelligent enough for coding

[Stefan Oltmann]

Depends. I throw some very simple search-and-replace like tasks at it. That looks good.

[Stefan Oltmann]

I think I'm going to use models like these for Compose / UI work.
This can rather easily be reviewed and issues face much faster.

To execute architecture & logic these models might not be the right call.

[Stefan Oltmann]

I now see what I misunderstood: The limits are upper bounds.
They have an extra table that states limit per model.

So it's $60 for something like DeepSeek, but only $15 for Kimi K3.

[Arthur]

I personally use kimi k3 for planning only and let the new deepseek v4 flash for the implementation. Then afterwards, do another review with kimi k3 and then fix it with deepseek. Rinse and repeat.

[Arthur]

Also using skills from chrisbanes might help the model steer better

github.com/chrisbanes/skills

[Stefan Oltmann]

That seems to be the right call! 🙂

Thank you for the hint to the skills. That looks very promising. I'll try.

[Stefan Oltmann]

@Arthur With the skills enabled DeepSeek was able to identify a performance issue regarding wrong state hoisting that slipped my review. Awesome!

[Arthur]

It was suggested to me by others but it was best to keep two different instance open, 1 for kimi and another one for deep flash, to accommodate the different providers' caching.

[Stefan Oltmann]

Another valueable hint, thank you. 🙂