Hacker News

points by nikolay 1 day ago | 0 comments

Opus 5.5 burns tokens like there's no tomorrow! Meanwhile, GPT 6.1 Sol on Extra High feels unlimited! Opus (on Max 20x) munches my weekly quota in a day, or two at best. Meanwhile, working with Codex and reviewing with a $50/mo Muse Code Wonder Full seems like it can go own forever (on the same codebase, the same skills, etc.) I am highly disappointed by the overly glorified Opus and Claude Code in general - there's no commit in which Codex or Muse Code or GitHub Pilot don't find critical flaws! I have long sessions of Claude Code having over 100 review cycles of Codex and still not addressing all the valid issues that Codex finds in Claude Code's work! Claude Code keeps injecting British English into prose and identifiers with nothing being able to stop it! Of course, there are almost never resets for Claude Code either - they would rather me focus on Codex being my primary and Claude Code being the reviewer (it works much better than in reverse). Anyway, enver trust a single model! Always review with another frontier one!

[–] jpardoe 20 hours ago

It's crazy how different everyone's experience is. I'm running 3-6 concurrent sessions in work hours and Opus 5.5 feels practically unlimited on $200/month. The value is incredible for how much I'm outputting. I finally managed to exhaust a weekly budget doing huge refactors and letting it run through the night with many sub agents so I thought I would try OpenAI for the first time. Even using Terra for most of my sub agents I can't believe how quickly I'm burning through my weekly quota. It seems significantly less generous than my Claude sub and at least in my experience Opus 5.5 is also clearly superior.

[–] nikolay 11 hours ago

Running in parallel and the new "thread coordination" is even worse! I "love" the cryptic messages threads share, and the endless loops of "main has changed, so, all tests done got invalidated". Maybe I have the wrong workflow, but I doubt. Mostly it's harness regular updates.

[–] hnart 10 hours ago

opus does feel insanely great with consumption at this point I'll let it rip for a long time on "high effort" and it'll move the needle on the 7d by 1%

my main gripe at the moment is how they structure ultracode workflows which I tried again last week after a long hiatus. It created insanely complex milestones/lanes that moved super slow

btw, I always mix and match models/harnesses but since Opus 5.5 I'm finding fewer reasons to go with Codex/Grok/Kimi/others

I expect this will change again soon

[–] SmirkingRevenge 17 hours ago

Same here. I think the programming language and project size/complexity probably have a lot to do with it

The code I own professionally has LOC in the 6 figures, and it's near impossible to hit weekly quota unless I'm running a bunch of experiments on the side or using Fable for everything

Probably a different story when the codebase is much larger

[–] thunfischtoast 23 hours ago

If you tell a model to find flaws, it will find flaws. What qualifies as critical to you?

[–] nikolay 20 hours ago

I am talking about functional flaws, not cosmetic ones. For example, in web apps, where a11y is mandated, Claude Code kept violating accessibility (regardless of the installed skills), and even the lackluster GitHub Copilot with some outdated models always finds serious violations. Codex constantly finds race conditions and other timing flaws for distributed services written by Claude Code. But I have to admit that Codex sucks at web design compared to Claude Code!

[–] deadnomore 1 day ago

100 review cycles sounds like a lot of arbitrary churn, I can get most models to review and endlessly find things to “improve”

[–] nikolay 20 hours ago

I constantly get into endless review cycles of 50+ rounds. Claude Code simply cannot get many things right regardless of the effort.

[–] HumanOstrich 20 hours ago

More paragraphs! Fewer bangs!

[–] xyzsparetimexyz 23 hours ago

What on earth are you working on that you're burning this many tokens? i can survive the regular 1x pro subscription just by using Opus on low and not medium/high

[–] nikolay 20 hours ago

A domain names web storefront with negotiations system, complex payment shapes, quite a few integrations with registrars and services, a Telegram bot for back office with mini apps, etc. A CLI, a TUI, an MCP, a WebMCP, text messaging. Built mostly on Cloudflare and with Ansible and OpenTofu. Runs also a local DuckDB database as I don't feel comfortable putting registry keys somewhere in the cloud, and so on. It is a scope creep, no doubt, but the idea is a to be fully autonomous, to ship features after I just pitch an idea to it. It writes blog posts, produces podcasts (not using Gemini Notebook!), advertises on social media, etc. Lots of work I could never do without coding agents, and I try to perfect it from all angles - following standards, and being as modern as possible. It's overkill, no doubt, but hey, why not. I'm trying to build a foundation for future projects, and get back into real development, and the last 10-15 I spent mostly with Infrastructure as Code. The maintenance of the entire system is probably less than $50/mo, which includes token costs for LLMs, ElevenLabs, etc.