This is honestly insane. With a subscription, sure, Opus makes sense. It's better and more heavily subsidized, so it's the better choice in both price and quality.
But man, the quality of GPT models at these prices is crazy. Imagine companies using pay-as-you-go pricing. Would you spend $700 per user on Opus 5.5, or accept slightly lower quality with GPT-6.1 and spend just $70 per user?
Google’s Gemini naming is getting interesting.
Argon (18) is already out, Carbon (6) is rumored, and Barium (56) has also popped up as an internal codename.
Could the atomic numbers be a hint at model size?
Carbon < Argon < Barium
I know PR acceptance rate isn’t a real measure of seniority, but what would be considered a good rate for a senior developer contributing to open-source projects?
And what rate do you think models like Sol 6.1 or Opus 5.5 could achieve working almost autonomously?
One pretty wild thing about Argon is how it avoids taking lots of turns, while each turn tends to run much longer than with other models. Not sure if that’s better or worse, but it’s interesting to watch that behavior
Gemini 4 Argon’s cache hit ratio has improved a lot since launch, and it really shows in the cost. It still does dumb stuff, but at least the dumb stuff is cheaper now
The probability of Gemini 4 dropping tomorrow just jumped from ~50% to ~80% in the last hour.
Looks like we might finally see Google's new model tomorrow.
Simulated what happened in Santiago, Chile a few days ago.
Wildfires earlier this year left the Andes slopes around the city with no vegetation. Then the rain came, nothing was holding the soil, and a debris flow hit Las Condes by surprise.
~88M tokens, ~$47.6, Opus on xhigh. Not a one-shot: 3 prompts and a couple hours of work. Incredible explanation and simulation.
Haiku 5.5 is a good launch. Good model, cheap model. But I don't really see it as a model meant mainly for coding. Two reasons:
1. Under 100k context the price drops 80%. So where the price/quality ratio really shines is low-context work: support chatbots, in-app assistants, classification, extraction, routing. Not long agentic coding sessions.
2. Anthropic is giving API credits to Max users, the people they figure are probably building something that could run on the API.
Personally I'm jumping on it right away. Gonna have Opus add Claude support to the AI assistant in my project and test Haiku 5.5 there.
Honestly I think Luna 6 will still be better, cheaper and faster for my use case. But $100 in credits every month is something I can't ignore.
Polymarket has Gemini 4 dropping this week as more likely than not.
The model SKUs are already showing up in Google Cloud.
Worth adding some context. This is true: Claude gives you more usage and Opus is a better model than GPT 6.1 Sol or Astra. But the gap isn't as big as the image makes it look.
First, GPT models are way more token efficient and cheaper, so a lower $ number still translates into more actual work done.
Second, Claude has a real bottleneck: the 5-hour limit. I've used both the $20 and $100 plans, and the weekly limit has never been what stopped me. The 5-hour window is, and it seriously caps how much of your weekly limit you can actually use.
Even so, Claude is still the better option over Codex on models and total usage. Even more if you learn to optimize your 5-hour sessions.
It’s insane seeing people complain about GPT 6.1 Sol’s speed. Bro, you’re getting intelligence around Astra/Fable levels, cheap as fuck, totally usable on a $20 Plus account, and you’re crying because it’s too slow and can’t oneshot stupid games like Opus? Bro…
Are Claude limits variable? When I first got Claude, I could burn through around $40 per 5 hour session. Two days later, I’m down to roughly $20 per session.
Does this fluctuate, or am I missing something?
