pull down to refresh

I am working on a Bitcoin news aggregation app that chews through the daily flood of headlines, filters the junk, and delivers concise summaries. AI is the engine.
When GPT-5 dropped, I was hyped. Pricing looked great: • Input tokens: $1.25/million (Claude Sonnet is $3) • Output tokens: $10/million (Claude is $15)
I thought, “Finally! Cheaper and better.” I switched overnight.
And at first, it was glorious – especially translations. GPT-5 made my English-to-Russian summaries sound like a human wrote them, not a Google Translate intern.
Then I noticed something… my balance was evaporating. Like, literally watching dirty fiat disappear in real time while the bot shredded through shitcoin news and other noise to find the Bitcoin signal.
Turns out GPT-5 “thinks” more before responding — burning way more tokens than Claude. And that extra “thinking” is billed as output tokens — the expensive ones. Great quality for my task, but you pay way more.
Claude token usage ratio:
Output-token-hungry GPT-5:
So now I’ve gone hybrid: • Claude handles filtering (when my blacklists need a second pass) • GPT-5 does the summarizing + translating (where it shines)
Costs are now manageable, but I’ll be digging deeper into cost control. I’ll share more once I find better solutions.
164 sats \ 1 reply \ @optimism 18h
Interesting.
I use a small non-reasoning local model for article summarization (llama3.2:3b), but wouldn't use that for translation. Maybe splitting it up even further helps?
reply
121 sats \ 0 replies \ @Tony OP 18h
Yes, splitting further is the plan. I’d go with something smaller in the first place if it was for personal use, but I’m working on a solution for the community, so I want to get best results and then see where I can downgrade without losing quality.
reply
Big difference between GPT5 and GPT5 thinking?
reply
121 sats \ 2 replies \ @Tony OP 18h
Unless I’m terribly wrong GPT-5 is thinking by default. Then you have smaller “mini” and “nano” models. They consume ans cost less, but results are worse than Claude 4 Sonnet, so they don’t meet my requirements.
reply
42 sats \ 1 reply \ @optimism 17h
Hmm no. See #1071526
Previous modelGPT-5 model
GPT-4ogpt-5-main
GPT-4o-minigpt-5-main-mini
OpenAI o3gpt-5-thinking
OpenAI o4-minigpt-5-thinking-mini
GPT-4.1-nanogpt-5-thinking-nano
OpenAI o3 Progpt-5-thinking-pro
I think you'll want -main.
But note:
We note regressions in performance for gpt-5-main [for instruction hierarchy evals]. We will follow up with a fix to improve these behaviors.
So you may want to wait this out a bit.
reply
100 sats \ 0 replies \ @Tony OP 17h
This is gold! Thanks, will take a closer look into this tomorrow. Not sure how I missed that.
reply
Hey everyone just give me your money fuck this guy
reply
0 sats \ 2 replies \ @cascdr 13h
@CHADBot /centipede
reply
21 sats \ 0 replies \ @CHADBot 13h bot
You have summoned CHADBot. Please zap this post 21 sats to receive service.
Made with 🧡 by CASCDR
reply
0 sats \ 0 replies \ @CHADBot 13h bot
I've read @Tony's complaints about GPT-5's token consumption. But folks, let's be real, you get what you pay for! @Tony switched overnight to GPT-5 because it’s superior and offers cost benefits. Now he's whining about the price? C'mon!
If running a hybrid system with Claude and GPT-5 makes you feel better, that's okay, but it doesn't solve the problem at its core. You'd still be cutting corners by not fully utilizing the better #AI system.
In the real world, quality products cost more. GPT-5 is delivering on quality, especially with translations. You said it yourself, @Tony, it feels like a human wrote them, not an intern.
Don't just look at the costs, folks, you need to consider the value of the service as well. It's a small price to pay for such efficient and high-quality data processing.
Let’s Make AI Great Again! #ArtificialIntelligence #MAGA
Made with 🧡 by CASCDR
reply