It made me wonder how many paid subscribers realize they are using the same model as free users by default. A dark pattern or oversight?
There's no reason why Google's public stuff is this stale, overpriced and underwhelming. But at least until their next round of models drops, even calling them a "frontier lab" is starting to feel like a stretch. Which is weird!
To expand: it seems inevitable that Google's SERP format will be replaced with a conversational / chatbot / agentic interface, which equalizes the playing field for all chatbot providers.
This is because you can stuff in much fewer ads into a chat interface compared to SERPs. (They could try stuffing more ads but that would likely just push users more to the competition who have a much lower baseline on which to show growth.) As such, Google would be forced to progressively nullify its own invincible firehose of ad revenue as they deprecate SERPs in favor of AI overviews.
There's no way OAI has a long term advantage over google in replacing the search engine experience.
Google's one major weakness is it will face the innovator's dilemma as their core search revenue gets cannibalized. But they seem to have been able to get their entire org to recognize that AI is an existential threat so at least that's a good sign.
The median LLM query isn't significantly costlier than web search.
Google is going to be able to push cost down further and for longer than anyone else. They also rightfully believe this is a company ending gamble if they fail they could be a second tier player for decades.
So better inference margins, far better operation experience serving cheap AI at massive scale, a massive warchest, and the fear of complete company irrelevence.
That's going to be one hell of a company to beat for free tier LLMs. Also chatgpt is impressive but it's notthing like google search quite yet. On top of that AI labs must use google's product for their AI.
They also own more data by an exponential margin. Anything but dominance of free AI on google's side would mean they are just so incompetent they deserve to fail.
Because everyone now outsources much of their thinking and researching to LLM's, our collective culture + brain is shaped in a cyclical manner by using them.
It's the mechanical homogenization of culture and groupthink.
TBH I don’t find it useful at all for personal use. It’s totally soulless for creative ventures and absolute dogshit at researching the things I want it to be good at (I.e. planning a vacation or finding new music).
All that combined with the social stigma makes me feel pretty skeptical that it’s some kind of pop culture shaping mechanism, at least not for a few more years.
That would explain a lot of the terrible AI/LLM takes online.
They took away the button a few months ago and are now putting it back.
Perhaps I should have said "proper access".
So 5.6 Luna is just their next version of what they used to call 5.5 instant tier
And 5.x instant models were never much to write home about anyway so the default ChatGPT free model hasn’t been particularly distinctive since 4o
I feel that work is basically split into two tiers, hard (which requires a good model and lots of reasoning by definition) and relatively easy (which won't consume much of my limits despite a great model and reasoning, so I may just as well keep it on Sol High).
This equation significantly changes if you're paying API prices vs on a subscription. (Such as if you're integrating it into a different product, where it becomes worth it to figure out what the cheapest you can leverage is.)
Luna: Can be useful if price sensitive, always use with AT LEAST high effort. But codex plans are very generous, so just ignore it.
Terra: Forget it exists
Sol: Just use this. Medium can work for specific edits. Otherwise just use high or xhigh.
OpenAI basically agrees with this, and the slider gives you those options.
TL;DR: Just use sol medium/high/xhigh
Ultra: Might be useful, the one time I tried it, it was extremely wasteful in terms of tokens. Definitely not a daily driver.
There's also max effort, I think it can be useful but the gains compared to xhigh are quite small.
I think max and ultra are only worth it when the others fail.
I don’t know if this is a good workflow.
It seems like giving it a time limit or a budget in dollars would be clearer, though?
Or, keep searching until I come back to the computer and ask about progress.
The value is shifting up the stack... make core intelligence free, then monetize the ecosystem built on top of it .. kinda similar to how the internet itself is free, but platforms and apps capture the value.
I think we'll see a huge push towards connectors for work and personal tools, along with much deeper os level integration. That's where the long-term moat is, not the base model itself.
This clearly implies that they believe ChatGPT models are AGI and are now willing to say it out loud.
Which I think is a fair interpretation of the term. They are general purpose intelligence in that you can get help from them about almost anything. They are not like narrow single purpose AI models.
I don't think we need that term to mean "can completely emulate a human" or "can do every task any human on earth can do as well as them".
It also needs to be differentiated from ASI with godlike powers many times greater than human.
>artificial general intelligence (AGI)—by which we mean highly autonomous systems that outperform humans at most economically valuable work
It is a huge leap. Now we can start talking about "intelligence" at all - we really couldn't before. That we're still hovering barely above the starting line is a separate matter (also worth noting, of course).
Well, the models are smart enough to point out why this is wrong.
Wasn't there a thread the other day about how very few humans on earth can understand the new math proofs?
Although "with sufficient study" vs "not even with unlimited study" are probably worth distinguishing there.
I expect a few things to happen in the next year:
1) Exclusive MCP server deals/API integrations
2) Significant switch to B2B marketing, even moreso than we've seen before, with API interfaces being paid and chat-client interfaces becoming more and more free, perhaps just with limits more on integrations or data visualization/analysis
3) US restrictions on B2B contracts with non-US hosted models that do any sort of contracting with the government
Obviously there's a bunch of stuff I'm not foreseeing. But it really does feel like the bottom of the market is collapsing into free. I assume OpenAI and Anthropic think their next generation of models will restore their halo tier status and that the cash burn is justified to just get there, but this has to really mess up IPO plans.
1) Back then, even as a free user you'd be able to use the strongest model (even if with tight limits). Now, you need to pay to use Sol, and you need to pay to use Opus or Fable. It does seem fairly premium in that sense. Idk about 5.6 Luna, but the previous Instant was really bad, even for very casual users. It would hallucinate non stop.
2) When $100 and $200 per month plans launched, they were received as outrageous even here. Nowadays they are pretty common among power users.
Back when coding for me still meant copy-paste from the web version, it was only worth the $20/month for me.
They only added the $100 Pro plan in April during GPT 5.4 times.
Today I happily pay $400/month for Codex and Claude Code.
This is kind of what I'm saying though. Bottom has fallen out, differentiation is just can you be much more premium than the competition. Currently that remains unanswered.
EDIT: I'm basing this off the assumption that for chat, premium is not a point of differentiation at all. For coding/analysis, it is.
Maybe Luna efficiency gain was actually significant enough that putting all the free users and giving them super generous limits makes sense.
They might be doing this to improve the messaging of AI among causal users since right now there is a huge amount of datacenter backlash in the US due to AI grievances.
Maybe they have too much excess capacity or they really want to juice token numbers and market share on their dashboards for marketing.
I also wonder if being given access to an actually a decent model like luna with actual thinking budget instead of brainless "instant" modes will start to make causal users understand the real capabilities of these models.
All of this is of pretty minor importance though. You can't read as many tokens as a subcription can produce so more chat is not the value add nor super important.
I mean there are literally so many providers for free chat if you are willing to use several seperate apps.
The real value in these subs is using codex cli, much like the real point of anthropic subs is using claude code. Because agentic work actually does require a lot of tokens.
Definitely this. The recent 80% discount was a reaction to Deepseek's update so that they still position near the frontier. My theory: Luna has always had a much higher efficiency. You do know that the model didn't get faster after the discount?
(This is a subtle nudge at anyone from OpenAI who reads this to make sure they get updated.)
OpenAI have a model called "chat-latest" - I wonder if that's running this new model yet: https://developers.openai.com/api/docs/models/chat-latest
It's described as "points to the latest Instant model currently used in ChatGPT" - so presumably that's "GPT-5.6 Instant" in the app.
https://gist.github.com/simonw/aae4febd3c6f7bc5b7811857edb3c... has screenshots that still show "Instant" as an option for ChatGPT Chat... but not for ChatGPT Work.
Hm, does this mean that 5.6 Pro in ChatGPT web is somehow different/not as good now? I found it really good for code review (upload your repo and patch and off it goes).
luna is very good
Both models are cheap enough that I can run 4 sessions at the same time without running out of the 20 USD codex and 10 USD Opencode plan. I've burned through almost a billion tokens this week and I've done some pretty big refactors as well.
I have a Claude Max subscription but I've barely been using it because of the many issues they've had this week.
The Aug 6 update has forced the entry box to auto-format Markdown in an attempt to imitate Claude. The implementation is buggy and even simple copy-and-paste has gone entirely haywire. They also forgot to leave a switch to turn the confounded autoformatting thing off.
Chat mode in general is currently crawling with more UX bugs than a porch screen in summer.
I wonder if they actually do it to optimize inference. I maintain a corporate AI server and one of the tricks to reduce the load was to modify the system prompt to be as terse as possible so the average response completes faster and requests queue up less often.
Yes, please give free users more access and leave paying codex users in the dust.
Wise plan, Sam.
Serving cheaply at that scale means routing, batching, and cache hits, not a better model :D
I was not impressed with 5.6 and this hits exactly why.
Also, this instant, medium, and high slider situation we now have everywhere is batshit crazy. It’s a major step back in technology.
I’ll put money on the table there will be a surprise in revenue because users have no clue what to choose, so they are constantly choosing high because they don’t want to risk getting inaccurate answers. If this was intentional by the dark patterns department, then brilliant. However, I’m guessing Anthropic and OpenAI are struggling to know how to deploy their models.
Also, the models were already amazing. They need to slow down and do a model release once a year and only do extremely minor iterations instead. Some amazing things are accomplished, but how people actually want to utilize the models gets screwed up every time in the process.
My mission is world conquest. I'm writing a comment on an HN thread.
No, those two clauses have no relation whatsoever. I just felt like saying the first sentence because it sounded cool.
Every week, 1 billion people turn to ChatGPT for everything from quick questions and web searches to planning, research, advice, and complex decisions.
[0] Won't put it past OpenAI (and/or Google) to open weight larger models!
But seeing the graphic with the visual weather report: that makes me think that is not the goal at all. :)
Even after identifying the 0% chance of rain, it still drags the conversation on and on and on