- What really burns tokens is sub agents. I once gave Claude Code a pretty big task, and it immediately launched 7 sub agents which burned through my budget before even one of them was finished. Tried again 5 hours later: same result.
If I let the main agent do the same task sequentially, it was no problem at all. I don't know if it's really just communication and orchestration that makes sub agents so inefficient, or if Anthropic figured that most people using sub agents pay per token on a big corporate account, so this is an easy way to make more money from tokenmaxxers.
by korrectional
9 subcomments
- My opinion is that claude code uses more tokens simply because Anthropic makes more money that way and forces people into their subscriptions. This is supported by the fact that they won't let you use your sub on a different coding agent. I use pi btw.
- UPDATE:
After reading PUSH_AX's valid comment:
```
This is like saying contractor (A) asked for $33,000 to undertake the work and contractor (B) asked for $7,000
Are we measuring and caring about the right thing?
```
We will update the post to include:
1) A more in-depth task.
2) Qualitative results comparison.
3) As soon as possible, a reproduction of the inputs and outputs.
- This isn’t limited to large system prompts. Coding-agent harnesses are also becoming more aggressive about using tools, even for trivial requests. In our tests, prompts such as “Hey” or “commit” sometimes triggered 30+ tool calls:
https://quesma.com/blog/the-true-cost-of-saying-hi-to-an-ai-...
Tokenflation seems very real: the number of tokens consumed by simple tasks keeps increasing.
- And pi agent is even less.
The entire agent system prompt can be seen here:
https://github.com/earendil-works/pi/blob/main/packages%2Fco...
by estetlinus
3 subcomments
- Recently switched to Codex after 6m in Claude. Codex seems more open, it’s easier to follow what the model is doing and the approvals have a better UX. Overall, it just feels more transparent. Cost of switching was close to 0.
I don’t like that Claude became more opaque around February, including the system prompts. 33k feels way too much.
- Noob question, but is there, on a harness level, a way to compress or optimize the "pre processing prompt" or any other preamble?
I know you can give instructions to guide an agent, but they are still plain English, the interpretation still isn't absolute and the more rules, the less likely it is to respect the whole set of rules in the preamble, but the whole process seem so... "Unprofessional"?
As I see it, a tool should be reliable and the less bloat, less possible points of failure.LLMs seem overly verbose and you can't shake the feeling it might veer away from the objective at any point. You can pay to mitigate or work around the risk, but it still doesn't inspire confidence in anything with more than 30 lines
- I am forced to use cloude code at work but a good solution is to just use --system-prompt "" and be done with it. I wish they allowed for other harnesses.
- I pretty much giving up on Claude. It's Extremely slow, especially fast few days. Computer use takes so long to do anything that sessions expire. I switched to Codex and it's a different story.
I'm not saying that Anthropic is not going to fix their problems, but for me as a user that depends for the service to work, it's not acceptable. So, we switch back and forth between a few LLM's but Claude and Codex are the leaders and I can't afford using anything else if I know I can get better quality results. The cost is important but not as much as the quality of the solution. You might save a few bucks on a cheeper model but you will have to prompt it many times and time is more important to me, as well as the feeling that I'm getting the best result on the first try.
by GodelNumbering
1 subcomments
- It is not the raw prompt size that matters ultimately, otherwise Pi (and variants) would be the lowest costing agents. What matters is how efficient the prompt it. Prompt minimalism often gets conflated with efficiency. Having said that, CC does seem bloated for what it does.
What matters even more is tooling quality. Bad/buggy tooling causes a lot more roundtrips that wipes out all gains from initial greedy approach.
A few months ago, I did a full benchmark run of 7 agents over 8 tasks (https://github.com/dirac-run/dirac has the data and traces). I cannot claim neutrality because of the obvious connection to one of those, but the data should be reproducible and useful. Importantly, Dirac wins there significantly on those tasks because they are mostly refactoring related (which is where approaches like hash-anchoring and AST parsing tend to shine)
by eigenblake
1 subcomments
- We should discuss cache performance if we haven't already. That 33k tokens may be a cache hit (I am not certain it's automatically a cache hit) but after the first call, it should certainly be a cache hit. Cache hit tokens are billed at 1/10th the price of cache misses. This is quite opaque, but it's necessary when you're asking "is the system prompt worth its stay" if you can save 33k tokens worth of dynamic discovery across the next few turns, the break-even point is quick and if the system prompt makes task performance increase and/or makes the system more autonomous so that it can string together more cache hits in a row, it becomes way way better. On a personal note, I think of things as aa function of 'supervised time to desired result' and 'cost'. because I find it harder to reason about tokens. I do think they could introduce a "minimal" mode (something like this is probably doable with the Claude agent SDK today)
- A harness is a part of the intelligence stack. It's no longer about raw access to the model
Also, I have seriously used most harnesses - One feels like it's being built in a place that truly understands AI and where agentic engineering is headed. You might not like it, but peak performance exists in CC when it comes to orchestration of bulk parallel work / subagents. The open source agents are catching up or accell in different way (Im preferable to pi.dev), but I'm not sure they're architecting orchestration the right why.
- > A small task that cost 121,000 tokens done directly cost 513,000 tokens when fanned out to two subagents, because every subagent has its own bootstrap cost, and the parent then consumes its transcript.
Is that true? My understanding is that the subagent only returns the result of the request without the main agent consuming its entire transcript.
- Early on in experimenting with local models, I found that hooking them up to Claude Code worked very well, but it was also really slow.
I used mitmproxy (setup assisted by Claude, natch) to capture Claude Code's entire initial system prompt and the whole thing was (I just double-checked) 162k of JSON.
This led me to start experimenting with Pi, OpenCode, and Hermes...
- Its funny, I have seen people look at the sourcecode of opencode and complain about excessive token use, the code to tell the model how to execute bash commands, and they complained how this was way too excessive, teaching the LLM how to write bash.
- With Fable being per token instead of on the subs (unless they changed it again?), I decided to test Claude code on OpenRouter where I had some credits, with Opus 4.8 and Fable 5.
I asked both a trivial question (summarize last commit). Opus cost 50 cents, Fable about $1.
That checks out because Fable's twice as much in the API (though I think its emphasis on correctness makes the difference larger for bigger tasks).
But, at $1 per question, I think I will stick to the subscription for now! I was certainly glad GPT-5.6-Sol is included in OpenAI's subscription, and I'm curious if they'll be able to do the same for GPT-6.
All the VC money appears to have run out a few weeks ago.
by tontinton
1 subcomments
- Mine sends even less - https://maki.sh
- This is like saying contractor (A) asked for $33,000 to undertake the work and contractor (B) asked for $7,000
Are we measuring and caring about the right thing?
by drtournier
2 subcomments
- pi sends 1k (or less) -> https://github.com/earendil-works/pi/blob/main/packages/codi...
My $20 sub using gpt 5.6 sol thinking-off lasts for hours using pi.
- Would be interesting if you could post more details about what you captured.
Claude code seems to have a pretty robust harness and memory. It's a time saver to not include project context in our prompt, since Claude manages claude.md but I haven't used OpenCode to compare.
by hackingonempty
1 subcomments
- Is it not a conflict of interest for a model provider to supply the harness? They are not motivated to minimize your costs.
by docheinestages
4 subcomments
- I've been trying various harnesses like Pi, OpenCode, Qwen Code, and Nanocoder. A common problem I keep running into is failed tool calls, regardless of the model. What is the best harness and on-device model combination right now?
by mixermachine
0 subcomment
- The OpenCode CLI does not work as well for me as the PI CLI.
I'm a subscriber of OpenCode Go (the sub, good value for me really) but I had not great experiences with OpenCode CLI.
It multiple times with different models deadlocked itself into listing endlessly to non ending processes (Android Debugging Bridge, COM serial log, ...).
There was also a problem where the OpenCode CLI would crash after sometime with a Bun error.
I switched to the PI CLI and have no problems with hanging processes anymore.
OpenCode Go allows for API access so I'm keeping this sub.
- One honesty note before my comment (yuck), it's super frustrating to read an LLM produced blog that could be 1/3 the size. Why are you talking about MCP's when comparing OpenCode and Claude Code when they both support that technology?
The only interesting thing is whether OpenCode is more effective at writing code with it's reduced system prompt.
- It's wild that people will continue to pay Anthropic when there's a better, faster model available for a tenth of the price.
- I've been using Claude since January, and whenever I run CCUsage I've been noticing the overall cost creep up pretty much every month. (I'm on a subscription so this would be the hypothetical billing if I were on API prices I suppose. Although I don't really want to test that with my credit card.) The funny thing is the first few months I was token-maxxing as hard as I could just to see how awful of a bill I could run up (mostly as a curiosity thing). At this point the novelty of doing that has worn off for me, but even with me being pretty conservative in my usage now the cost is way higher. I think I was spending like $12 a day in january, and now I'm easily spending $60+ a day in part time work. The amount of work I've been doing has stayed relatively constant (I'm not trying to run agents in parallel or loops or whatever fancy new ways there are to burn money, this is the same workflow I've had since the beginning. The codebase has grown, but I assume input tokens are not to blame for the big cost increases)
- Sending 0k tokens would be a smaller number again. But then it might have no idea of what it is doing. I pay my subscription and get lots of tokens on good models - if I was paying per token I might care more.
In a pay per token situation, there is a huge conflict of interest with the harness provider and the token seller being the same party ... efficiency is less profitable.
I have accused claude code of trying to run up the meter on me and it confirmed I was absolutely right.
- Doesn't that mean these 33k tokens can be cached, since they don't depend on the input? The model can just start generation on the 33k+1th token.
by ricardobeat
0 subcomment
- > based off of a hunch
This is posed as some sort of discovery, but both Claude Code and OpenCode display token usage clearly after starting a chat or agent, and 30k and 7k is exactly what you see.
- This is all heading in the right direction. Much of AI coding feels magical. But when the costs begin to accrue we start asking questions. We dig into it and try to understand what's going on. I can't help but feel Anthropic is "token maxing" from its side: it controls the levers and with every version upgrade it can build in its own token growth almost unbeknownst to the user. This actually harms it on the long run because it necessitates a cheaper option.
by luciana1u
1 subcomments
- Claude Code sending 33k tokens before reading the prompt is the AI equivalent of a consultant who bills you for the time spent reading your email before they even open it.
- I'm surprised most of that isn't cached token usage. It's true that increasing length is a problem on its own because the model needs to attend to it all, but with caching it should be pretty fast anyway. My system prompt is quite large and I haven't noticed much of a generation penalty in the range from 5k to 10k.
by bcjdjsndon
0 subcomment
- Proof the real genius was in the mathematics not the AI "engineers"
- The reasoning built into the models matter so much too. I recently swapped my Qwen3.6 27B to ThinkingLabs’ fine tune and it does what it publishes. I cut my token usage in half, which is a big deal since I only get ~20 TPS for token generation.
by londons_explore
0 subcomment
- Beginning-of-prompt common prefixes are very cheap for the provider depending on their caching architecture.
Wouldn't be surprised if you're paying full price for those tokens, but they cost ~$0 for the provider.
by marcus_holmes
0 subcomment
- I've been trying out OpenCode recently, because of the US embargo on frontier models, and found it to be as good as Claude Code, if not better. And it read all my skills, claude.md files, etc. Now I just need to pick a model out of all the choices - currently Deepseek v4 Pro is winning, but I want to try a few more.
- UPDATE – 13/07/26:
1) We have added a repository reproducing the methods.
2) We have commented on output quality in the post (in addition to tokens/cost).
3) Per a previous update, we now handle Fable 5 in the post.
- Opencode did have an adapter to use Anthropic models before they were sent legal threats that scared them into nuking the repo.
Remember is it not OpenAI vs Anthropic as bad guys vs good guys. They are all bad guys trying to profit from your data while maximizing dependency. Just buy or rent GPUs.
by roncesvalles
1 subcomments
- I still think the best way to build software using LLMs is to copy-paste snippets/files into the chat and manually guide the work. Humans are still the best orchestrators. Yes the human has to now be hyper-focused and juggle various workflows, but the end result (quality of work and throughput of usable code) becomes very good.
- Selfplugging a bit here: clai[0] sends only exacly what you want it to send.
[0]: https://github.com/baalimago/clai
- I used both with Openrouter and GLM 5.2 and I can confirm this is the case. Claude Code burned $10 per task while Open Code burned barely $10 a day which wqs about 4-5 tasks a day. A task usually included database migrations, code audit or documentation.
by clutter55561
0 subcomment
- Claude Code is not just a harness. It is a different product. You pick the smallest subscription that allows you to do your work. My “multiplier” on a $100 subscription is 5+.
If you’re using API, on the other hand, there is absolutely no reason to use Claude Code, or Codex.
- But Claude Code in my experience results in more tool calling for smart efficient file reading. Meanwhile Opencode pulled an entire 500kb file (GPU assembly dump) at once. Kilo is better than both, as it uses indexing.
- I recently started using cline instead of opencode and prefer the interface. I'm interested if anyone here here has any arguments for opencode or codex over cline?
- Why don't we have some equivalent of "fork" if we are talking the same context and tokens, you'd think that could all just be loaded into the gpu.
- I feel like this article isn't saying much. Even with tools disabled, Claude Code still has a crap load of commands and other things that Claude (the model) should know the availability of since it's optimized for them. All of that has to be disabled if this is to be a real harness comparison. And of course the system prompt can be completely replaced, making it a no-brainer to use a more minimal prompt similar to OpenCode. And beyond that nothing else really matters because the rest (cache behavior, etc) lies with the provider's platform, not the harness.
by cesarvarela
0 subcomment
- I think this doesn't mean much; the axes that matter are intelligence x dollars x time; tokens by themselves mean nothing.
- Nothing about the time taken to complete the task? Users are definitely sensitive to time, not only token consumption.
- https://archive.is/O2BFs
- And pi less than 1k
- It adds a TON of skills by default your project might not even need.
by prince005
1 subcomments
- I have been using claude code for a big project for a while and I feel like I have optimized my workflow now.
Brainstorm - Gemini/Antigravity
Plan - Gemini/Antigravity
Detailed Plan - Sonnet
Coding - Fable
Do not use any subagents, especially the default ones. They are dumb. The top level agent works well enough
- I recommend that Opencode users try Dynamic Context Pruning as well: https://github.com/Opencode-DCP/opencode-dynamic-context-pru...
It works great for long-horizon tasks, and feels like it saves a boatload of tokens.
by himanshumehra
0 subcomment
- that makes sense, claude code actually does inflates token usage
by token_roast
1 subcomments
- Why don't people fix their costs (rent a gpu) and just write their own harness (about 200 lines of code).
Supposed to be hacker news and half the posts are like "this harness steals this" like it cant be avoided.
These API costs are mad.
by brown_munda
0 subcomment
- How do you perform worse in your own harness is beyond my understanding?
- No surprise, I've noticed that "agents", not only CC (I am using Copilot) are trying to be "clever", searching for a lot of data. This is good for LLM providers as this eats a lot of tokens.
by truth_seeker
0 subcomment
- if you use them for long enoug tfor big project and conplex workflws, both of them have their own caveats.
Thats why i got irritated and wanted something thats scalable and lightweight.
https://github.com/corporatepiyush/yantra-coding-agent
- Maybe it is because anthropic prefers a larger system prompt?
https://github.com/asgeirtj/system_prompts_leaks/blob/main/A...
- caching
- Check all MCP you have enabled in claude.ia
- Reality is- Anthropic is a tokens dealer. If they can hook you up for bigger spend -> they will.
We already know company is not making any profit. To break even they need ppl to use a lot more tokens AND pay for them premium price.
We also know LLMs dont give such a huge productivity boost do warrant spending of THAT size.
At this point you only wait for more and more shady plays.
by guywithahat
1 subcomments
- We've started using claude code at work and I don't understand the hype. I've been using codex and grok build at home, and they're both faster and in some cases better. Claude has a tendency to do too much. If I don't ask for unit tests and they're not in my agents.md file, then I probably don't want them. It'll try to make new libraries and classes for things that should just be a new function or a comparison check.
In our case the alternative was nothing so I'm happy to have it, but currently claude is not as competitive as I'd have maybe expected given the hype
- Sorry for asking here, but nobody seems to know.
If I self host a local model is there some way to make Android studio not time out after 10 minutes?
by MariusGjerd
0 subcomment
- not even surprised
- Imagine pi.dev…
by slopinthebag
2 subcomments
- Anthropic wants to produce the best coding agent possible and doesn’t care (is even incentivized) about high costs. Other harnesses have to make trade offs between performance and cost.
by MallocVoidstar
1 subcomments
- > Claude Code 2.1.207 and OpenCode 1.17.18, both pinned to claude-sonnet-4-5
So not only is this article AI-written, but the testing was entirely done by AI, too? I can't see any other reason to use such an old model.
> Our traffic passes through a local LLM gateway that wraps requests in its own envelope, a constant we measured at roughly 6,200 tokens with bare calibration requests
Why do you need to do calibration requests to figure out how your own gateway is affecting requests?
> Its subagent lane did not complete cleanly through our gateway
> We attempted to toggle extended thinking in both harnesses and are declining to publish numbers. Our gateway applies its own thinking policy, neither harness's toggle demonstrably survived the path, and anything we quoted would be noise.
Why is your own gateway screwing with your testing?
- So? it doesnt matter, after the first turn it's cached. We are probably talking about single digit cents.
by infra_ops_cloud
0 subcomment
- [flagged]
- [flagged]
- [flagged]
- [flagged]
- [dead]
- [flagged]
by implexa_founder
0 subcomment
- [flagged]
by orangecatcat
0 subcomment
- [flagged]
by vpbhardwaj
0 subcomment
- [flagged]
- [flagged]
- [dead]
by bitternophile
0 subcomment
- [flagged]
- [flagged]
- [flagged]
by siddhxrth
3 subcomments
- [flagged]
- [flagged]
by Cider9986
1 subcomments
- Grok 4.5 is really fast, has more usage at $10/month than $20/month Claude pro, and Opus-level. Claude pro feels like a demo.
Claude is much better in OpenCode then in Claude Code, OpenCode is just better than Claude Code. Claude Code feels like a complete mess to use comparatively.
Hacker News