I wanted to see how a bit of API is used, so I had Codex.app create a sample Xcode project and it went off and did exactly what I wanted (which wow, so cool).

Figured it was a good chance to try out Claude to see it was any better. Gave it the same prompt and it too eventually created the test app, but much slower. I tweaked some settings tried again and it went quite a bit faster, but then immediately hit its 5 hour API limit.

Same $20 plan, Codex was just a ton more efficient both time & $.

@paul codex gives more usage these days.

They are running promo for codex app launch with 2x the limits.

It used it be so slow but now better.

@paul I like Codex, but everything's fine and dandy until the enshittification hits the fan.

@paul It’s also a lot more efficient at funneling your money into certain hands 😞 https://www.theverge.com/ai-artificial-intelligence/867947/openai-president-greg-brockman-trump-super-pac

@paul Whenever you give money to OpenAI, you're giving money to these people:

https://archive.ph/WGTXo

At least Anthropic aren't selling sex chat bots.

@paul Plan with Opus and execute with Sonnet. That's really the only way to use the $20 plan.

@paul I think you’re right, but also Codex limits work differently. Claude is really aggressive with its 5 hour limits, but you can hit a weekly limit on Codex much quicker.

@jamesfmilne @paul meanwhile Anthropic stole books, and were one of the early identified parties of ignoring robots.txt leading to crawler overload.

@paul same conclusion here. Will switch from Claude Code to Codex very soon.

@wezm @jamesfmilne @paul The entire "AI" industry is a moral and ethical cess pit, and it drives me fucking BONKERS that people just don't seem to give a shit about any of it.

@paul huh, I’ve had the opposite experience. Time to try codex again.

@paul Didn’t use it for any Xcode projects, just web apps based on rails. My workflow consists of speckit to prepare the specs and the plan in Claude code then also implementing it. I could get quite a lot done with the 20$ plan. But switched to the 90$ one now. And I got so much done in one day. It nailed it with every time and it was working on existing projects, some with old codebase.

I'm still testing Claude (even switched to Sonnet instead of Opus) vs Codex. Gave them both similar prompts while cleaning up some code, mostly a mix of move this stuff around and review this change I just made. After about 30 minutes, my 5 hour usage on Codex is 3% vs Claude's 37%.

IMO Claude gives better insight at the "Review this" level of things, but just don't see how anyone can actually use it without going on the $100 or $200 plan. Unless I'm doing something wrong…

@paul The $20 plan worked for me fine when I was using Sonnet, but Opus definitely turned it into a bottleneck for me.

The $100 plan is a lot, but for me it's brought more than enough value for me. I'm eyeing Codex as well, though, it just hasn't clicked with me yet.

@paul IMHO Codex has more tokens included in the plan to attack Claudes leadership as coding agent.

With my €20 plan in Claude, I get roundabout 90min agentic coding for HTML/CSS/Javascript per 5h-session

@paul I’ve been using Claude this week on a project and hit the 5 hour limit three times so far, all after about 1.5-2 hours of coding.

Clearing the context between tasks helps. You can also plan an implementation, write the plan to a file, then start a new instance and tell it to “implement the plan in <file>”.

But yes, it’s hard to really lean on Claude with just the $20 plan.

@mikepj are you on the $20 plan? I doubt I could make it past an hour with anything resembling heavy use.

@paul I'm on the $100 plan and I can use Claude Code it pretty much nonstop without thinking about it, FWIW. I think OpenAI is running it as a loss leader. I guess they all are to some degree.

@manton that's what I wanted to know, most stuff I've seen just compares the two as if there isn't this huge difference between them.

@paul For me, I've never hit the Claude limit when the scope of the request is easy. Where it seems to run out of limit quickly is when I tell it to add a major feature and don't give it any direction

@paul I saw a YouTube video about something called “Carl” that reduces token usage by a lot by minimizing what is sent from claude.md files each prompt.

For my purposes, I use Gemini or codex for generating most code and Claude as the “senior engineer” to review and fix the code. I only start with Claude if I know it’s going to be something complex.

I’ve mostly been able to stay within my limits that way, and I have access to 3 engines to split work across for ~$60 a month.

@matt_birchler @paul would you say the $20 plan is fine for using it as a spicy autocomplete + some “apply this pattern on some other similar files”? I pay $15 for Windsurf and never got even close to exhausting its credits, but would like the remote Claude Code web access sometimes.

@paul Depends on what you’re doing. I have Claude Pro for personal use and it works great, but we have the Anthropic API plan at work. I’ll give Xcode 26.3 a prompt sometimes and then run /cost on the work plan and it’s used like 50¢. 😮

@paul please do publicly mention if/when you use it for any part of Ivory. I have quite liked using it and am happily subscribed presently, but I would be forced to cancel in that scenario.

@alahmnat I don’t know that people don’t care (most people aside from the boosters) but given the amount of hyperbole and flat-out misinformation around it I’m not sure how people are supposed to make good judgements about it. With almost any topic these days there seems to be a lot of noise on the sides and not much signal in the middle. Most things are never black and white, good or bad, and AI is no exception.

@paul AFAIK Claude Code still supports API-based billing, through the CLI at least. Newer features may require a pre-paid plan. You can see usage at platform.claude.com and set billing refill amounts and such. I set it up this way when Claude Code was new and it’s still working.

@paul I've been using it for reading through my recent BASIC Tutorial series. "Pretend to be 7 years old, and tell me if this makes sense."
It's been quite handy for that, if I'm honest. It's either that or I kidnap a local child to read through it, so..

I gave the recent Opus a try the other day and .. Good grief, that's powerful.

I wouldn't rely on it to write full things, but getting to do little snippets is probably no worse than googling for an answer.

It still lies, though!

@paul Yep, $20 plan using Opus.

@fantinel @matt_birchler @paul Yes! Refactoring, cleaning up, adding small features / UI changes - easy going.

Sure, if you add more complex stuff like big features, whole services from scratch … it gives to time out soon 😄

@paul im using it for cowork and also BLASTING through my allowance in no time. Results are better than ChatGPT but it’s costly af.

@nSonic @matt_birchler @paul ok good to know!

@matt_birchler @paul Codex certainly seems to have larger limits at the moment

@paul even $200 is nothing if you get that value in 1-2h in an afternoon.. the rest of the month is pure profit. it all depends on the user

@paul Codex has 2x limits for a limited time now to emphasize this :)

I see bigger token consumption (codexbar.app) with the higher levels and with large contexts. I feel like splitting up into multiple smaller tasks or compacting context before it auto-compacts saves more.

Therefore, multi-step jobs using Plan Mode use much more than planning in a PLAN.md file.

@matt_birchler @paul they’re including Codex in the free plan now for a while if you want to try.

@paul I get the biggest Claude plan from work and got Codex $20. Tried Claude$20 for side projects too and it’s unusable

@cmaciasjimenez the $200 plan? Assuming that one is hard to hit limits on?

@paul sorry was typing on the iPad. The $200 plan on Claude is hard to hit the limits yes. I sometimes hit them but only have to wait for 30 mins or so because the 5h time was up. But this is only a few times on the last 3 months and always using Opus which takes more credits. I meant the $20 was unusable.