Pi Agent vs OpenCode token usage A lot of people recommended Pi Agent so I decided to check Pi Agent took 1.1k tokens in first turn OpenCode took 11.5k Setup: 1) Trimmed OpenCode (from usual 30k first turn to 11.5k) - 0 MCPs - 2 lightweight plugins (opencode-env-protect and... openslimedit) - 8k char AGENTS.md - 11585 deepseek-v4-flash input tokens - for $0.0016 2) Vanilla Pi - 0 MCPs - 0 system prompt - 8k char AGENTS.md - 1114 kimi-k2.6 input tokens - for $0.0008 I think using better models with capped tokens per turn can keep usage nearly the same as uncontrolled DeepSeek V4 Flash? The challenge now is finding the sweet spot. Can't cap tokens if quality drops. We'll see. Video below - OpenCode vs Pi side by side "say hi back" first turns test, with OpenCode Go usage for eachshow more

I'm also disabling all my MCPs and rebuilding my workflow with CLI-based skills Token economics are changing fast, with subscriptions becoming more token constrained 50-100k tokens per turn from MCPs alone is a recipe to hitting usage limits fast As an MCP maximalist I'm going to miss how predictable MCPs are. Models are going to get one CLI argument wrong most of the time

@pseudokid now try SoulForge

wow thanks for sharing!

pi has less number of tools and extensions available, so its going to be more token efficient

I agree

i think need to compare and understand where this tokens going and if its help or not the daily coding... i dont think its useful to just compare raw vs modded... maybe pi with 5k prompts can be even better than pi 1k or opencode 11k (but yeah, 11k is a lot )

Totally. I agree. I'm about to figure out if less tokens with Pi and a more capable but slightly expensive model is enough to do the things that I can do with the cheapest in OpenCode Go which is deepseek v2 flash

Now try the same model from different providers. Like @InferXai and someone else.

@InferXai $20/mo for @InferXai right?

@InferXai Yes. While we are in beta and limited spots very limited available.

have you tried @CommandCodeAI $1 Go plan? also i think a great system prompt helps you!

@CommandCodeAI saw their ad a couple of days ago and i might actually try is system.md better than agents.md?

I have the same experience for local models. Was trying Gemma 4, qwen 3.6 locally on my M5 Max 128GB with Opencode+Ollama, found it is pretty slow, until I accidentally tried pi with the same repo…damn, it is night and day…is it because it has a gigantic system message? No MCP, no skills, just factory default for both. Also, Codex CLI is awesome to use👍

I'm seeing a lot of people using Qwen 3.6 locally. I also find Pi being too light and fast. Users start from nothing so you they control what they put in it. Saving tokens is just a side effect of that imo. Which one you find better, Gemma 4 or Qwen 3.6?

Just go ahead and run some of those longer tasks with context in the 140k range. You’ll really see who is cheaper then

About to do that, that's the real test

Have you compared it with @CommandCodeAI yet? I haven't tried it yet (hopefully before the end of the month) but I have been seeing a lot of wonders about them and their process. Like

I agree with you! Don’t know much about stats. But I used an agent workflow in opencode and then just gave a try in pi, there’s slight usage difference that I noticed, then optimized in pi itself now for each workflow run it hardly cost 10 cents to me with MiniMax M2.7

I feel like MiniMax M2.7 and even M2.5 will run better on Pi just because of Pi's minimalist approach I have problems with these models for long-running tasks in OpenCode. Easy to lose key and recent details that even older gens Mimo V2 and Kimi K2.5 will never lose My theory is it easily gets distracted so it has to be given one single coherent point of focus Can't have system prompts competing with system wide AGENTS.md and task brief

Because opencode it's a coding harness with longer instructions.

@grok is Pi Agent better than Opencode in terms of perf and cost ?? What's opinion of devs on X ??

How does no system prompt work? It jut gets tool schemas and runs with it?

I'm still new to Pi, but I simply didn't use a SYSTEM.md for it I also noticed that Pi relies on the terminal, just like Codex Desktop. OpenCode has its own Read and Write tool and bash.

Try with @justsisyphus too

So you put two different models into two different harness and concluded?

Did you save the first turn and look at it? Bc I am skeptical the pi is including your Agents file

I thought so too. So I tried probing it and it understood the AGENTS.md I have to test more and learn about the internals of Pi agent I only have the usage record from my OpenCode Go dashboard. Tokens used and price and that's it.

in output which seems better?

still testing it!

This is kinda misleading in longer tasks.

I'm about to figure it out, the real answer lies there

@thdxr

@Teknium

there's support for rules.md, commands and agents on pi besides just the skills and agents.md ?

Same pattern I've seen, most of the burn isn't the model, it's re-reading everything every turn. Building ThinkForge, search-first indexing (pull only relevant chunks, cache what's read) cut first-turn cost more than any model swap did.