There are huge token efficiency/bloat differences between agents while working on the same tasks, using the same model, in the same environment
Yesterday I ran 10 agentic tasks using GPT 5.6 Sol in an ubuntu 26.04 vm a couple of times with different harnesses and got vastly different token usage.
+-------------+-----------+-----------+-----------+-----------+--------+
| Harness | API total | Input | Cached | Uncached | Output |
+-------------+-----------+-----------+-----------+-----------+--------+
| smol | 172,807 | 142,334 | 8,704 | 133,630 | 30,473 |
| Pi | 427,211 | 392,767 | 137,216 | 255,551 | 34,444 |
| OpenCode | 1,564,429 | 1,523,957 | 1,204,736 | 319,221 | 40,472 |
| Codex | 3,005,744 | 2,953,154 | 2,649,344 | 303,810 | 52,590 |
| Hermes | 3,856,611 | 3,808,231 | 3,167,232 | 640,999 | 48,380 |
| Claude Code | 5,073,137 | 5,029,969 | 4,587,008 | 442,961 | 43,168 |
+-------------+-----------+-----------+-----------+-----------+--------+
I'm not surprised that Claude Code is not optimized for an OpenAI model but I was still quite shocked re how much of a difference the harness makes.
Disclaimer: I'm working on 'smol' which is a minimalist harness but it's really nothing special, just a minimal system prompt, no skills files, only tool is shell
Do not underestimate how much popular harnesses are spamming the context window. The context window is very important.
Do we have any insight into whether it is actually spam and not useful info such as project or programming language specific context?
but even injected context that when I read it sounds useful can oversteer the model and make it second guess or take a more complicated route than it normally would
(you can see this when looking at traces with and without that injected context)
often harnesses also mention in their system prompt locations of markdown files that the model can consult if the model thinks they might help
that hint alone as part of the system prompt can be strong enough to make the model read in more tokens than would have been necessary
'spam' is maybe a harsh way to say it
unfortunately I don't see an easy way other than to invest time and tokens into finding out which parts of the added context (in system prompt, injected in turns etc etc) are actually helpful or harmful and when
I'm just doing the easiest thing I could think of: start from nothing or close to nothing
that seems to work better than what most harnesses are doing
turns out GPT 5.6 Sol is all you need
The difference in tokens between the two also makes super curious. The system prompt can't be that different (I'd even bet Pi's shorter) and the 4 tools shouldn't make as much of a difference. I'm gonna have to try it.
the system prompt of smol is shorter than the system prompt of Pi
smol has no system prompt
system prompt of Pi 0.83.0
""" You are an expert coding assistant operating inside pi, a coding agent harness. You help users by reading files, executing commands, editing code, and writing new files.
Available tools: - read: Read file contents - bash: Execute bash commands (ls, grep, find, etc.) - edit: Make precise file edits with exact text replacement, including multiple disjoint edits in one call - write: Create or overwrite files
In addition to the tools above, you may have access to other custom tools depending on the project.
Guidelines: - Use bash for file operations like ls, rg, find - Use read to examine files instead of cat or sed. - Inspect PI_* environment variables for current model and session details. - Use edit for precise changes (edits[].oldText must match exactly) - When changing multiple separate locations in one file, use one edit call with multiple entries in edits[] instead of multiple edit calls - Each edits[].oldText is matched against the original file, not after earlier edits are applied. Do not emit overlapping or nested edits. Merge nearby changes into one edit. - Keep edits[].oldText as small as possible while still being unique in the file. Do not pad with large unchanged regions. - Use write only for new files or complete rewrites. - Be concise in your responses - Show file paths clearly when working with files
Pi documentation (read only when the user asks about pi itself, its SDK, extensions, themes, skills, or TUI): - Main documentation: /usr/local/lib/node_modules/@earendil-works/pi-coding-agent/README.md - Additional docs: /usr/local/lib/node_modules/@earendil-works/pi-coding-agent/docs - Examples: /usr/local/lib/node_modules/@earendil-works/pi-coding-agent/examples (extensions, custom tools, SDK) - When reading pi docs or examples, resolve docs/... under Additional docs and examples/... under Examples, not the current working directory - When asked about: extensions (docs/extensions.md, examples/extensions/), themes (docs/themes.md), skills (docs/skills.md), prompt templates (docs/prompt-templates.md), TUI components (docs/tui.md), keybindings (docs/keybindings.md), SDK integrations (docs/sdk.md), custom providers (docs/custom-provider.md), adding models (docs/models.md), pi packages (docs/packages.md), environment variables (docs/environment-variables.md) - When working on pi topics, read the docs and examples, and follow .md cross-references before implementing - Always read pi .md files completely and follow links to related docs (e.g., tui.md for TUI API details) Current working directory: /workspace """
You should use --disallowed-tools to prune any tools not needed for the task. Note that this is also a perpetual game of whack-a-mole since they’re always adding new tools.
It's so unfortunate they don't let you use the subscription with other harnesses anymore - since even if I used OpenCode they'd still get a bunch of useful data from the API calls, meanwhile I could stretch their tier limits way further.
200$/month is a lot of money on Luna/DS4 flash, like really a lot and the results are much better than clowning on bloated CC.
It's absurd how you have more and more organizations encoding their processes on LLMs and "engineers" (charlatan coders) don't even bother optimizing the tool they use most.
I won't argue with the cost effectiveness, but the results are very much not better. Opus and Fable are in a different league than DS4 Flash. Even GPT Terra, which I really like overall, sometimes gets stuck in weird loops and starts to do stupid stuff once its context window fills up. Whereas I can more or less trust the big models to just Do The Thing™ on the first try.
With that said, you get way more value out of a GPT subscription than you do from Claude, partly because of the ability to use more efficient harnesses.
I’m not sure if they let you skip the cache write cost on the first turn. That would imply cross-user caching infrastructure or special casing the default system prompt to give you a discount. Maybe? Away from the computer but you could try a “hello” in a fresh session and see what was billed.
It is everything. My experience with Claude Code is that you have to decide when to compact to make it efficient. It defaults everything to 1M context and it will never keep it in check. It is strange how little cache reads you hit in smol, that may be a configuration issue.
I’m asking because I’ve been looking for agent harness comparison tools too. I’m interested in more than just the inputs and outputs—I also want the system prompts, traces, and tool calls. It’s useful to understand why Codex, for example, uses more tokens while Pi doesn’t.
Fewer tokens aren’t necessarily better if the agent skipped important checks. On the other hand, using more tokens could just mean it’s overthinking the process. Either way, seeing the full execution trace for the same task is really valuable.
I agree fewer tokens is not necessarily better but a bit counter-intuitively often the harness using fewer tokens is not only done faster but has better results
(that said: of course check the results, look at the full traces, agree!)
the uncached tokens are also from runs where smol finished a task below 1024 tokens (the minimum amount of tokens needed to activate caching) which is less tokens than other harnesses are using for their system prompt (!)
> GPT-5.6 and later models: Caching is available for prefixes containing at least 1,024 tokens. This is a strict minimum.
https://developers.openai.com/api/docs/guides/prompt-caching
so in this specific case the count of uncached tokens for smol makes it look worse than it actually is
that said: it does makes sense to add more tasks that are difficult enough to fill the context window to compare the harnesses for how well they deal with compaction
staying below compaction (or with compaction at fewer compactions) is not only cheaper and faster, it also helps the agent stay on track
also, are you using a tool to collect those metrics? what is it?
https://news.ycombinator.com/item?id=49006862
I'll have more about it in the next hours/days, you can follow me on twitter in the meantime (https://x.com/__tosh)
like creating a checksum of a file, merging csvs and so on, fixing a makefile pipeline
with known 'good' outcomes
all harnesses could reach the outcomes, only cost, time, number of tool uses and so on were different
(Claude Code failed once in 1 task but I think that was just an unfortunate outlier, the tasks aren't that difficult)
Be careful here. Remember these are non-deterministic models at the end of the day, and even with everything being "the same" you can have two runs where the same model, same harness, same tools can arrive at the same conclusion through a wildly different sequence of events.
I will add more tasks (esp longer ones) and think more about grading, the current tasks were easy to grade because the desired outcomes are well specced but I will also look into more open ended tasks and how to grade those
thank you!
the "spamming" is mostly compaction appending files, tool artifacts, images in its summary there is a known github issue for codex. only solution is periodic clean up but its also how a lot of the agentic orchestration is performed and able to work for days.
I will look into how token usage looks like for longer sessions and more complex tasks
re caching: the cache ratio for this bench looks 'bad' for smol because it often finishes a task before caching kicks in (caching starts at 1024 tokens)
thank you for flagging this
These days I “write” code with claude code and codex, and read/review it on GitHub. If I need to read it locally, I use a plain text editor.
Can someone help me understand what value cursor offers in 2026?
1) It's still an IDE. Because of pricing I mostly use Codex, but I always have a VSCode/Cursor IDE open, thus have to juggle between the two. Working directly in the IDE is more comfortable. For full on vibecoding that might be worse, but when you want to do a deep review of the changes, an IDE is way better than reading a diff on github.
2) It supports every model. It's often very helpful to try different models when you don't like the result of the first.
These days I've switched to Zed, which is good enough (and wicked fast), but I still miss Cursor as an IDE.
The $20 price point is just too competitive now and I'd rather use the Claude/Codex plugin over Cursor's agentic coding sidebar.
To answer your question though, to me your workflow seems cumbersome. Cursor is more integrated and more frictionless.
In agent mode it’s horrible for editing files etc but better if you are juggling multiple chats
Should the AI sidebar be on the left, or on the right? At some point they swapped them automatically, which was jarring (maybe a bug). But now I realize that if I'm not the one primarily writing code and navigating it, then I prefer agents in the "primary" position on the left (esp. if I have the browser on the left half of the screen, the agents are in the middle).
What annoys me the most about Cursor is the random stuff that breaks (due to its nature of VS Code fork).
The tab experience is still unparalleled. The cost might be worse than Codex/Claude sub, but they are all much, much, much cheaper than API pricing, so depends on your POV.
Especially if you are doing "remote" development through SSH. If you are doing stuff where you still have to write some parts of the code manually or you have to fix few things here and there that the AI outputs, you still need a real editor.
Unless by "text editor" you mean Neovim or Emacs ?
The potential benefit of Cursor CLI (vs CC and Codex) is that you can easily between all major models (by Anthropic, OpenAI, xAI, as well as Kimi K3 and GLM 5.2). I found it useful when reviewing work - e.g. I implement using Opus then review using Sol, etc. Models by different providers tend to have different perspective on things and they can find different issues with the code.
You can still see what you’re billed on the Spending page. We did accidentally break dollar costs in the Usage CSV export yesterday while cleaning up an old feature flag. That was not intentional and the CSV export is fixed now.
That feature flag also showed a dollar usage graph to some self-serve users. The confusing part was included plan usage shown as $, which is not what you’re billed (on-demand usage is). Some people read it as actual spend, so we decided to remove that graph.
No, you can not: https://www.pasteboard.co/dNXUdT-h8Giy.png
If you want to say that "admin can" - it doesn't matter, I'm not going to ping admin every day to check how it goes. I'm not going to ask admin about every session to check how cost efficient a model was.
Double-edged sword. It is also simple to move back to VS code and agent extensions.
I imagine she’s just weighed all the details and options in her mind and plans to french fry tokens from my plate instead of getting her own.
And I hate that. I mean, I love her. She’s definitely value-add. But those are my tokens, right? My brain doesn’t do books. I can’t sit on the beach and read. I must always be swimming somewhere meaningful. So I’ll be vibing the next great Canadian web app and she’ll just casually ask, “hey what do you want to do about dinner?” So now I’m asking Copilot to tell me what I want for dinner. And it’s just… c’mon lady get your own tokens.
Gotta justify a $60B purchase of an IDE and (at the time) a single, decent model.
These days I'm using Codex and Claude Desktop with Zen when I need to look at code. Codex's real time audio chat feature (not dictation) is also second to none when paired with their agentic flow.
Kind of (barely) like how Facebook has "friends". Or how a Snickers bar costs $1.99 and not $2.
Abstracting meaning of "cost", reducing value of information. Maximizing profits. Enshittification.
You came back from the dead pretty much and now you're pissing it away for what exactly?
Do not spite your individual developer customers or you will perish yet again.
It might be called grok out it was trained by the composer team using a large chunk of their training data.