I'm working in large US corporation.
And I see that I already have access to 5.6-Sol Ultra on my corporate account.
I haven't really used it yet.
2 months ago management was showing us scoreboards, praising leaders who used most tokens.
Last few weeks, we're getting weekly emails, telling us that whenever we can - we should use cheaper models, and that we should watch the page which shows our tokens usage.
I’m in Finance and learned pretty quickly that point out the implicit future cost raises based on the cost the LLM-providers need to recoup was unpopular at best (STFU better describes the situation). Running full force into a bear trap.
Sounds like a bad idea in general. Any data use agreements get lost, shadow-IT brews and nobody knows what tools to use, oh and it's against the service terms.
Generally not, since the corporate account has all the privacy knobs turned up. I use my personal account on my open source projects, where code leaks aren’t exactly an issue.
The market is priced at expecting AGI levels of breakthroughs. Just a very useful tool for programming is definitely not enough to keep the music playing.
The thing with language models is they are tailor made to fool managerial types into thinking it’s the holy grail.
Just like many managers, the appearance of productivity is all that counts. And LLMs shine at giving the appearance of having solved all of the managers problems, and all they have to do to use it is spend on tokens.
This isn’t to say that LLMs aren’t truly useful, they absolutely are. But they’re very nature is one of simulating intelligence through next word prediction.
The chat modes and models are by their nature supremely attractive to management layers, because they give answers that sound so damn plausible even when they are complete fictions, and uttered with such confidence how could they be anything but the singularity.
> Additionally, we’re introducing a new ultra mode that goes beyond the capabilities of a single agent by leveraging subagents to accelerate complex work.
For pro mode the agents worked independently and only when they all finished did a new agent take a look at everything to merge the work into a single response. The new thing involves subagents that have been trained to cooperatively pursue a task and are allowed to communicate with each other along the way.
I tried a pro model out the other day and thought there must have been a bug in Pi’s cost calculations. But no, it’s absolutely fucking insane. Wasn’t even any better at the task.
I really suspect that the models are basically the same below, it’s all in the prompt. The way I use them, surgically, they seem to perform about the same. Fable certainly hasn’t blow my socks off.
Some of it feels boiled down to "opus works better when told not to be dumb, fable's prompt tells it not to be dumb."
If they know much of what the tool is used for, they can customize prompts to "do that usage right" even if the user doesn't know exactly how to ask for it.
Same. I suspect they'll get better at taking in terrible prompts over time though... Maybe that's what Fable does better, reminds me of Sora 2, it would take my crappy prompt and expound upon it. I told it once to generate a video of someone working at some company that changed its name, but the old name had historic relevance, it referred to the new company name without me telling it to, by virtue of me wanting a video of TODAY with a 90s icon.
This is where I think you see the distinction between two classes of LLM users:
1. Managers: those who generally know what needs to be done, and want it done faster, so they provide a lot of instructions and context (where many developers fall)
2. Executives: those who vaguely know the end goal, but are clueless about the process, and are willing to burn resources and cycles on a black box to get the result
> We generally treat GPT-5.5’s safety results as strong proxies for GPT-5.5 Pro, which is the same underlying model using a setting that makes use of parallel test time compute. As noted below, we separately evaluate GPT-5.5 Pro in certain cases because we judge that the setting could materially impact the relevant risks or appropriate safeguards posture.
Any idea what that means exactly? I vaguely remember that ChatGPT Pro was originally called "deep thought", just like Geminis "deep thought" feature (or "deep think"?), so it seems likely they are using the same approach.
Its widely accepted[1] that it runs the same query through the model in parallel and then has a model that either selects the best answer or synthesizes an answer from the multiple ones generated.
I believe most people think it runs 6 sub-models, but I think that is based on the pricing.
It's a pity that OpenAI doesn't publish details like this.
Basically like passes@6 or passes@5 if you’re doing a benchmark, except for your real tasks.
Pro is quite limited on the web UI I reckon. This approach can be highly effective for reasonably verifiable task, for example, write comprehensive unit tests pointing out a tricky bug, get multiple agents to swarm at it.
I asked Claude in the browser if it could do anything like that. It wrote a little frontend app that calls the Anthropic API (with fetch()), without including a key. I expected that to fail, but it worked!
Apparently in the web chat (and also in Claude Code?[0] Though I haven't tried yet) they can call the Anthropic API and your subscription key gets auto-magicked into the requests somehow.
Those are two separate things of course (aside from the key-injection) but I guess there's no reason it couldn't run completely in the front-end... hmm...
How is this any different than what we have already? We've had this ability for ages (6+ months, decades in the AI world), you can literally today easily prompt CC or Codex to use subagents to accomplish tasks and they'll do it well. My entire workflow is one top level orchestrator chat creating tickets to dispatch to subagents to implement, and other subagents to verify. Why is this being sold as a new thing? Have HN users never tried tried asking CC or Codex to use subagents?
I assume this is ~equivalent to ultracode in Claude Code, which can deploy a tree of hundreds of nested subagents and was just released experimentally 5 weeks ago IIRC.
"However, these inference optimizations, which rival Anthropic refers to as “compute multipliers,” are a big focus for all the labs. Anthropic CEO Dario Amodei has been publicly talking about the concept since at least mid-2023, when he said on a podcast that the company limits “the number of people who are aware of a given compute multiplier” because it could give other AI labs a leg up if they were to be able to replicate them. (Compute multipliers can also refer to efficiency optimizations in the model-training phase.)"
Yes, on a world with finite resources where your industry is singlehandedly siphoning ALL THE RESOURCES - hoard general efficiency optimizations and treat them as trade secrets - winning is all that matters, normal people and other species and the planet be damned.
Everything I hear about Dario these days makes me like him less and less. He sure did seem to speed run the 'tech leader with scruples' to 'tech villain' path! I guess all the cycles are compressing as we approach the singularity..
I agree that Dario is pretty annoying, but I think the "tech villain" archetype is essentially survivorship bias. The tech leaders who don't act that way are not nearly as visible because they're not nearly as successful.
HN is just a massive Anthropic hate fest now, probably funded/manipulated by OAI's $8B PR budget.
OP phrases it as a bad thing that Dario is keeping compute multipliers to Anthropic. How naive can one be? Compute multipliers are the whole business. Those are the trade secrets every lab is built on. It is the alpha of the business. How does protecting this make Dario evil?
This website is getting out of hand with the uninformed hot takes. I wish when HN was still people that knew what they were talking about.
Not sure I know where I fall regarding your point: Yes to trade secrets, but also science and AI should be for the good of all.
OpenAI seems to be trading roles back with Anthropic becoming misanthropic. I hope they both start heading in the direction of how the AI field was prior to LLMs.
Collaboration and benefit for all should always be the primary motivator.
> Collaboration and benefit for all should always be the primary motivator.
Of all the things to never happen, this is never going to happen the most.
That train left the station for good once hundreds of billions to trillions of dollars were involved.
On the bright side, in the long run I suspect the vast majority of the value of AI will not be captured by the model making labs and the vast investments in them are going to implode, so...
They're in a price war with the People's Republic of China running flat out with the full backing of a government that literally does not care if they ever see a financial return on the investment, they just want to drive the value of LLM training and inference to zero because we banked the market on it being arbitrarily high margin forever. China was like hold my beer.
They have a staggering surplus of grid capacity and can bring more online without any difficulty. We couldn't get a serious nuclear project done if Jeffrey Epstein was offering private flights to the ribbon cutting.
In the United States at any given time more than half of the FLOPs are badly misallocated, Meta has like, a double digit percentage of the total capacity going down the drain every day and has for years. That's a conspicuous example but on OpenRouter rankings it's rare to see more than one or two American vendors in the top 10, sometimes the top 20. But 3rd, 4th, and 5th place are all merrily burning half the compute duplicating effort and missing key innovations because we stopped publishing real results. In China if DeepSeek makes a breakthrough it's at Zhupai and Moonshot and MiniMax and MiMo and Qwen that week.
Our only lever, export restrictions, seems to do nothing but breed multiply antibiotic resistant super hackers who just get more efficient and immediately propagate all of those efficiencies to the rest of the Chinese AI industry.
At the beginning of 2026 there was one Chinese lab with a model that had any real relevance fielding modern tool users. Today in July there are like, eight lagging the absolute frontier by maybe 3-6 months. Barring some massive bend in some curve 3-4 of the top 5 and 6-8 of the top 10 will be Chinese and open weight by January.
The great irony in all of this is that our current playbook is straight out of the 1960s USSR, and the PRC's current playbook is straight out of 1960s USA. We're the ones with the opaque decision making and gross resource misallocation driven by the personal agendas of a shadowy cabal of frenemies wired back channel into government in the form of the individuals rather than the offices. They're the ones with a thriving marketplace of ideas powered by robust public/private partnership and a paved path running bidirectionally to the university system.
It's going to implode because the Kruschev system does. Theirs is going to thrive because the Kennedy system puts a man on the moon before the decade is out.
>full backing of a government that literally does not care if they ever see a financial return on the investment
There's no evidence of this, the parsimonious explanation is PRC AI, by virtue of being sanctioned, simply is not able to run magnitude more expensive compute model, and even if they could, they don't have the $$$ or market cap to do so. So they optimize and involute margins like they do in everything, and US misallocated expensive flops because the entire industry has been financially engineered for phat margins along the entire producer supply chain is just cherry on cake. Like wipe out the 50%+ margins from toolmakers, fabs, gpu/memory/data center components to some reasonable level and US is overpaying for tokens by a stupid multiplier on top of actual compute misallocation due to incompetent infra. Maybe PRC AI has unsound economics, but it's structurally simply not able to misallocate as much as US who will find a way to financialize compute to point of absurdity.
> the company limits “the number of people who are aware of a given compute multiplier” because it could give other AI labs a leg up if they were to be able to replicate them.
I wonder if that makes sense if the orgs within the industry are starting to shift their mindset towards "Tokens are expensive, we should use AI less." which feels like an existential threat to the status quo, if those AI providers can't find ways to keep costs affordable for their clients. Otherwise those orgs would just be using GLM 5.2 or DeepSeek V4 Pro but it seems like what they're doing instead is trying to use AI just less, period.
i really hope it's just what Deepseek V4 does. Deepseek V4 is very cheap and highly performant
OpenAI tried to pull off the same trade secret thing with RL when they announced o1 and o3, aka "Compute time scaling". Then Deepseek revealed it with Deepseek R1.
Could also be something like Deepseek DSpark. Or using diffusion like DiffusionGemma as a draft model. The timing between the release of those, and this article, makes me think its maybe one or both of those things
Dario tells the truth. If you look at everything through their safe AGI mission it all makes sense. They are not bs'ing about that. Also I think most people just read headlines or 10 second clips and make false extrapolations from there.
(BTW Anthropic only exists because Sam Altman is a liar, Dario admitted this.)
A no-investment policy would take them off the scene entirely. Essentially handing over the reins to OpenAI, Google, and others. Their position is something close to "if I don't do it, someone worse will".
There's a more nuanced discussion that could be had about how to balance relevance with outside influence. But at a foundational level it should be acknowledged that the tradeoff exists, and that receiving outside investment can't alone be seen as evidence of corruption.
Besides that, there's more that can be said about other things like their corporate structure or the degree to which they accelerated the AI race.
Of course that's what Dario thinks because that's what every tech CEO thinks. Dario, Sam, Sundar, probably many Chinese CEOs as well. It's what everyone thinks. That's why they're competing so fiercely with one another. That's why they basically make all the same decisions. That's why we need properly open source AI.
Open source AI fails first contact with sufficiently-intelligent-as-to-be-dangerous AI.
The day Mythos class models are open sourced will not be a good day. I don't think you understand the impact that will have on the world and on cyber defenders everywhere. It will be pure chaos.
Even if you don't think Mythos-class is the bar, open source has to stop at some point, you don't hand everyone a superweapon.
I find this kind of cynicism fascinating tbh. On the one hand, it seems so relatable in some ways, because there is something uncomfortable about being seen as naive, in a way that being seen as cynical or negative doesn't seem to carry. I guess it's just self-protective, almost like some kind of perverse Pascal's wager: it's better to think everyone is horrible and be wrong than to think the opposite and be taken advantage of?
The thing I can't quite square is that it doesn't really fit my lived experience. I have known sincere, genuine people in the types of positions that I'm sure someone like you would declare to be sociopathic.
But beyond that, I just don't know why it would actually be true that everyone at the top is a villain. Why couldn't someone like Dario (or even Altman, gasp) be sincere? Because if he is, it does seem like a lot of the moves he's made would make sense given his worldview.
But if you assume he's just a villain, then you can twist any of those moves to just be further evidence of that which you already believe.
I don't know, I just find cynicism interesting, and a little sad.
You don't have to assume anything. A true "good guy" doesn't openly say that he's fine with autonomous, AI-powered weapons being used against me, and mass surveillance applied to me and my family just because I don't live in the US. A true "good guy" doesn't say "privacy is a human right", and then immediately (and completely) bend the knee to an authoritarian government on this issue.
It's a lot easier to sound smart on the internet if you're a bitter cynic.
Lots of nerds for some reason have made cynicism a personality trait. They think optimism/honesty is hopelessly naive, therefor cynicism is the correct default.
It stops being interesting, or even sad, after a while. People get stuck in all kinds of places, mentally. Some get unstuck eventually. It’s only sad if you have come to a counter factual belief that it could have gone better.
I went in the opposite direction - how far can I push myself to see multiple facets of a story? That is a wild ride, and it gets progressively more wild.
It's the Tragedy of the Commons. There will always be a vacuum of predatory bullshit that can be filled, and the victor is always the biggest sociopath. Rockefeller, Cecil Rhodes, Elon Musk, it's the same traceable pattern all the way back through history. It's not that everyone is like this, but that a few crafty marketeers are able to ruin it for everyone.
Why should I treat Sam and Dario with special white gloves? Are they different, this time? They have peers in China that do the same research and actually release it to the public. They let you run the production weights on your own machine. Am I a cynic, for comparing these CEOs to their populist superiors? Am I stupid for assuming their hostility when they refuse to give us the benefit of the doubt?
I'll believe their actual altruism when I see it. Both are seeped in "boy genius" puffery and lie out their ass. If this is the future of intelligent innovation, then America is truly declining.
Semi-related, has anyone noticed their GPT 5.5 usage in Codex being cut in half as of a couple days ago? I got a lot more mileage out of my session usage yesterday for the same workload.
I've noticed less quota and 5.5 intelligence degrading. I didn't run the analysis like the post the other day, but I had noticed decreasing ability to complete tasks, more laziness. Switched back to 5.4 and it's much better. Maybe they're getting ready to launch 5.6?
Like google search, this does not work because of how common long tail use is.
What you think could be a big chunk, is more likely to be a fraction of a percent of queries.
And what use is similar query caching - so you (very often! if actually cost effective, maybe half the time) get a response to a query that was different from yours. Including for when you have a lot of context input already. You’re going to get trash.
If it were constrained to only very common initial prompts, and somehow the long tail did not actually dominate as it does with Google search (can't find the reference at the moment but it was a famous article some years ago), it also wouldn't account for serious enough cost savings. Long context is what is expensive.
This might only work in constrained domains like customer service where there’s tolerance for generic answers and escalation paths. For technical work? For general purpose use, with secretly canned responses charged at full price?
But there must be a ton of generic questions that people ask. Stuff like "What's the capital of country X?" - it's probably at least 10% of queries. Memories, custom instructions etc would invalidate them, but if you can return the answers basically free it's probably worth it.
Questions like that cost a tiny fraction of a cent. "What's the capital of Sri Lanka?" cost a fifth of a cent at GPT 5.5 API price, and would cost a fraction of that if the question were routed to a more suitable, cheaper model. The output was 78 tokens.
By contrast, when coding, devs typically have hundreds of thousands of tokens in the context window, and may use many millions of input tokens per day.
Caching requires the full prefix to match exactly. If a single word differs near the beginning of the prompt, nothing after that can share the cache. So this type of caching would save a few queries that cost virtually nothing, but wouldn't help with the stuff where cost matters.
“ Automatic Prefix Caching (APC in short) caches the KV cache of existing queries, so that a new query can directly reuse the KV cache if it shares the same prefix with one of the existing queries, allowing the new query to skip the computation of the shared part.”
This thread is about cutting costs in half for GPT across the board.
The technique you linked only makes a substantial difference for particular use cases where you are going to have many LONG CONTEXT queries with the same prefix. For instance, when having a set of documents that commonly get loaded in as context. It's a way for application developers to keep prefixes they manage (or prefixes managed by some set of their users) cached. It has no relevance for long tail general purpose use.
Please pardon the pure speculation incoming. Yes, caching the answer doesn't seem useful. Caching the progression, the graph, may be. This is similar to making code changes with ed(1) instead of editing in vi.
The transform script(s) are cached and can be played back or adjusted. Surely for some breadth of question inputs, they map more often to similar answers--but not static answers; instead, evented edits.
It's nearly untenable for a human to keep private edit scripts to generate code changes. The extra steps for custom regex, essentially one-offs for a shared codebase, is inefficient. But maybe not to an LLM.
Not including their best model in a max subscription would otherwise be truly a good reason for once to consider going back to openai for me. I'll at least try it.
Recently, I've been so eager to get new model releases in Codex. I'm hooked. I hope this accelerates development. Shows how dependant I have become to Codex.
Will it have similar limited access like Fable? It is an interesting timeline, as general access for Fable (without using extra credits) is coming to an end :(
Bruh when did understanding chatbots become like following pokemon? Wtf does any of this this mean. Tf is sol? Tf is ultra? Tf is codex? Tf happened to descriptive nomenclature?
It's a proper noun: the name given to their latest and greatest model. Means "Sun" in Latin. Similar to how Anthropic has been naming its models Fable, Opus, Sonnet etc. Their other models are called Terra (Latin: Earth) and Luna (Latin: Moon) [0].
> Tf is ultra?
The name of the "harness" around the model. It'll use deeper thinking, subagents and all that jazz in response to a prompt. Other options include max, high, medium etc I suppose.
> Tf is codex?
"Coding agent" similar to Claude Code [1], something with a more descriptive name.
> Tf happened to descriptive nomenclature?
Something like GLM-4-32B-0414-128K (not made up [2]) doesn't quite roll off the tongue I suppose.
I would assume yes - their goal is to capture consumer subscribers. Claude are going to take Fable away, and they're going to swoop in and give it to us.
Still avialable through the API. According to people that have tried both Fable nad 5.6, Fable is clearly better at coding. So i expect a lot of people to pay extra for it.
I still don't know why OpenAI doesn't put gpt-5.5-pro in Codex. It's one hell of a model and easily parallels Fable/Mythos. Sure, it'll use up your quota much faster but that's the price some users are willing to pay for absolutely high quality responses.
I think gpt-5.5-pro runs 12x parallel gpt-5.5 agents behind the scene and uses OpenAI's secret sauce to synthesize their answers into one insanely good response.
API pricing ends up being something like 20x more expensive for GPT 5.5 Pro than GPT 5.5 for actual work, even though the token cost is "only" 6x. On benchmarks where I've run both, I saw $1.12 mean per task with 5.5 and nearly $23 per task with 5.5 Pro, I guess it chews longer and harder on the problem.
If that's at all reflective of what it costs them to run it, I imagine they're in the same boat as Anthropic with Fable; they probably can't afford to offer it at subscription prices given current cost to operate it.
If 5.6 Sol Ultra has efficiency improvements (at one or more layers), and it allows OpenAI to offer a model that's competitive with Fable on the subscription plans, I'll guess a lot of folks will switch.
Fable is notably better than what came before. I watched it figure out stuff on its own over and over, on extremely hard problems, that I previously needed to guide a model to an understanding about, or work with them back and forth for several turns to figure it out together. Like, I've been reverse engineering a hardware device lately, and I've tried to tackle it a few times in the past with both some version of GPT and a couple of versions of Opus (most recently 4.7). In all cases, I barely made progress...would have gotten there eventually, probably, as I'm stubborn, but there were roadblocks constantly, with me and the models getting stumped and going around in circles in the end on every prior attempts.
Fable figured out other ways to find out what's happening, it dug into config files, found and extracted Boost-serialized data, compared that data to the observed behavior, built tools to compare the observed data with our emulated behavior, without being prompted. Would I have gotten there? Eventually, maybe. All prior models didn't; they mostly just tried the things I suggested and stopped at "well, that didn't work" or declared success after seeing results that matched their misunderstanding of the problem. I guess it's possible my prior attempts with other models had "loosened the lid" on the problem; we did already have a long list of documented "this didn't work" and a pile of tools for finding out if something worked. But, even so, I was impressed.
There probably will still be a "OK, let's rewrite this so it's not using lookup tables to precisely simulate the hardware behavior in software, because we don't need the noise, too" stage of the process...but, in one day with Fable, it solved a problem that I'd banged on for at least a week or too in the past with very little real progress. I don't think the models write exceedingly good code, even the best ones, but it sure does figure shit out quick.
I recently have been testing ChatGPT business at work and the quota seems to disappear almost instantly even using weaker models. Unless they dramatically increase their quotas it’ll be unusable.
I don’t know how anyone can realistically use the “business” plans - you blow through your quota so quickly. I use a consumer Pro account ($100 a month) and don’t hit the usage limits nearly as quickly. 5.5 Pro is so slow that it’s not a big deal to paste big prompts into it and come back and check on it an hour later.
My solution for the ones stuck with that: use 5.5 for planning and 5.3-mini for the grunt work. 5.3 is remarkably useful still but you need to hold its hand.
Is it as good as Fable..? Fable is the first model that mostly writes without the AI slop format for me, and so I can comfortably actually copy and paste most of what it spits out.
OpenAI models have always been the worst in my experience for verbose, slop formatted responses, with each generation increasing in sloppiness.
I hate that I have had to remove it from my writing style because people assume it’s AI generated. But I think that ship has sailed. I’ll have to do without now.
How do you type the em dash. I thought the point about the em dash "—" is, that it is longer than the normal minus "-". Humans normally have no way to produce it, cause there is no key on the keyboard.
Some text editors replace the -- (two separate dashes) with a proper em-dash. Literate people - who understand why em dash exists - have been using it all the time. Thats, after all, how the models learned to use it.
Many word processors (Microsoft Word, LibreOffice Writer, etc.) and some online editors will auto convert double hyphens "--" as they are typed into an em dash.
Macos, ios, google docs, and microsoft word will autocorrect two hyphens to an em dash, which is how I normally type it. On a mac you can also type an em dash with option-shift-hyphen.
Why can't you reply to hn_user2? I just did. Did you say something mean to hn_user2 and now you have a restraining order and can't go near them within one reply?
I haven't really used it yet.
2 months ago management was showing us scoreboards, praising leaders who used most tokens. Last few weeks, we're getting weekly emails, telling us that whenever we can - we should use cheaper models, and that we should watch the page which shows our tokens usage.
Everyone is insane.
Just like many managers, the appearance of productivity is all that counts. And LLMs shine at giving the appearance of having solved all of the managers problems, and all they have to do to use it is spend on tokens.
This isn’t to say that LLMs aren’t truly useful, they absolutely are. But they’re very nature is one of simulating intelligence through next word prediction.
The chat modes and models are by their nature supremely attractive to management layers, because they give answers that sound so damn plausible even when they are complete fictions, and uttered with such confidence how could they be anything but the singularity.
> Additionally, we’re introducing a new ultra mode that goes beyond the capabilities of a single agent by leveraging subagents to accelerate complex work.
https://openai.com/index/previewing-gpt-5-6-sol/
Can someone explain how this compares with Pro? I thought Pro was already something similar.
It’s far more careful than opus and puts far more effort into testing and validating by default.
Switching back to opus at work was a downgrade. Similar requests felt more clunky and needed far more hand holding.
If they know much of what the tool is used for, they can customize prompts to "do that usage right" even if the user doesn't know exactly how to ask for it.
Same. I suspect they'll get better at taking in terrible prompts over time though... Maybe that's what Fable does better, reminds me of Sora 2, it would take my crappy prompt and expound upon it. I told it once to generate a video of someone working at some company that changed its name, but the old name had historic relevance, it referred to the new company name without me telling it to, by virtue of me wanting a video of TODAY with a 90s icon.
1. Managers: those who generally know what needs to be done, and want it done faster, so they provide a lot of instructions and context (where many developers fall)
2. Executives: those who vaguely know the end goal, but are clueless about the process, and are willing to burn resources and cycles on a black box to get the result
The responses I get from pro don't feel like ensembles. They are often very one directional.
> We generally treat GPT-5.5’s safety results as strong proxies for GPT-5.5 Pro, which is the same underlying model using a setting that makes use of parallel test time compute. As noted below, we separately evaluate GPT-5.5 Pro in certain cases because we judge that the setting could materially impact the relevant risks or appropriate safeguards posture.
https://deploymentsafety.openai.com/gpt-5-5/model-data-and-t...
There have been multiple podcasts with people from OpenAI which have confirmed this.
Any idea what that means exactly? I vaguely remember that ChatGPT Pro was originally called "deep thought", just like Geminis "deep thought" feature (or "deep think"?), so it seems likely they are using the same approach.
Its widely accepted[1] that it runs the same query through the model in parallel and then has a model that either selects the best answer or synthesizes an answer from the multiple ones generated.
I believe most people think it runs 6 sub-models, but I think that is based on the pricing.
It's a pity that OpenAI doesn't publish details like this.
[1]eg https://news.ycombinator.com/item?id=48799977
Pro is quite limited on the web UI I reckon. This approach can be highly effective for reasonably verifiable task, for example, write comprehensive unit tests pointing out a tricky bug, get multiple agents to swarm at it.
https://www.erdosproblems.com/
Hopefully, 5.6 will automatically spawn sub-agents without needing to ask.
Apparently in the web chat (and also in Claude Code?[0] Though I haven't tried yet) they can call the Anthropic API and your subscription key gets auto-magicked into the requests somehow.
Those are two separate things of course (aside from the key-injection) but I guess there's no reason it couldn't run completely in the front-end... hmm...
[0] https://code.claude.com/docs/en/workflows
To what effect I don’t know… I thought subagents were useful because they were explicitly single purpose and bound to a narrow context
https://www.theinformation.com/newsletters/ai-agenda/openai-...
"However, these inference optimizations, which rival Anthropic refers to as “compute multipliers,” are a big focus for all the labs. Anthropic CEO Dario Amodei has been publicly talking about the concept since at least mid-2023, when he said on a podcast that the company limits “the number of people who are aware of a given compute multiplier” because it could give other AI labs a leg up if they were to be able to replicate them. (Compute multipliers can also refer to efficiency optimizations in the model-training phase.)"
Yes, on a world with finite resources where your industry is singlehandedly siphoning ALL THE RESOURCES - hoard general efficiency optimizations and treat them as trade secrets - winning is all that matters, normal people and other species and the planet be damned.
Everything I hear about Dario these days makes me like him less and less. He sure did seem to speed run the 'tech leader with scruples' to 'tech villain' path! I guess all the cycles are compressing as we approach the singularity..
OP phrases it as a bad thing that Dario is keeping compute multipliers to Anthropic. How naive can one be? Compute multipliers are the whole business. Those are the trade secrets every lab is built on. It is the alpha of the business. How does protecting this make Dario evil?
This website is getting out of hand with the uninformed hot takes. I wish when HN was still people that knew what they were talking about.
OpenAI seems to be trading roles back with Anthropic becoming misanthropic. I hope they both start heading in the direction of how the AI field was prior to LLMs.
Collaboration and benefit for all should always be the primary motivator.
Of all the things to never happen, this is never going to happen the most.
That train left the station for good once hundreds of billions to trillions of dollars were involved.
On the bright side, in the long run I suspect the vast majority of the value of AI will not be captured by the model making labs and the vast investments in them are going to implode, so...
They have a staggering surplus of grid capacity and can bring more online without any difficulty. We couldn't get a serious nuclear project done if Jeffrey Epstein was offering private flights to the ribbon cutting.
In the United States at any given time more than half of the FLOPs are badly misallocated, Meta has like, a double digit percentage of the total capacity going down the drain every day and has for years. That's a conspicuous example but on OpenRouter rankings it's rare to see more than one or two American vendors in the top 10, sometimes the top 20. But 3rd, 4th, and 5th place are all merrily burning half the compute duplicating effort and missing key innovations because we stopped publishing real results. In China if DeepSeek makes a breakthrough it's at Zhupai and Moonshot and MiniMax and MiMo and Qwen that week.
Our only lever, export restrictions, seems to do nothing but breed multiply antibiotic resistant super hackers who just get more efficient and immediately propagate all of those efficiencies to the rest of the Chinese AI industry.
At the beginning of 2026 there was one Chinese lab with a model that had any real relevance fielding modern tool users. Today in July there are like, eight lagging the absolute frontier by maybe 3-6 months. Barring some massive bend in some curve 3-4 of the top 5 and 6-8 of the top 10 will be Chinese and open weight by January.
The great irony in all of this is that our current playbook is straight out of the 1960s USSR, and the PRC's current playbook is straight out of 1960s USA. We're the ones with the opaque decision making and gross resource misallocation driven by the personal agendas of a shadowy cabal of frenemies wired back channel into government in the form of the individuals rather than the offices. They're the ones with a thriving marketplace of ideas powered by robust public/private partnership and a paved path running bidirectionally to the university system.
It's going to implode because the Kruschev system does. Theirs is going to thrive because the Kennedy system puts a man on the moon before the decade is out.
There's no evidence of this, the parsimonious explanation is PRC AI, by virtue of being sanctioned, simply is not able to run magnitude more expensive compute model, and even if they could, they don't have the $$$ or market cap to do so. So they optimize and involute margins like they do in everything, and US misallocated expensive flops because the entire industry has been financially engineered for phat margins along the entire producer supply chain is just cherry on cake. Like wipe out the 50%+ margins from toolmakers, fabs, gpu/memory/data center components to some reasonable level and US is overpaying for tokens by a stupid multiplier on top of actual compute misallocation due to incompetent infra. Maybe PRC AI has unsound economics, but it's structurally simply not able to misallocate as much as US who will find a way to financialize compute to point of absurdity.
I wonder if that makes sense if the orgs within the industry are starting to shift their mindset towards "Tokens are expensive, we should use AI less." which feels like an existential threat to the status quo, if those AI providers can't find ways to keep costs affordable for their clients. Otherwise those orgs would just be using GLM 5.2 or DeepSeek V4 Pro but it seems like what they're doing instead is trying to use AI just less, period.
OpenAI tried to pull off the same trade secret thing with RL when they announced o1 and o3, aka "Compute time scaling". Then Deepseek revealed it with Deepseek R1.
Could also be something like Deepseek DSpark. Or using diffusion like DiffusionGemma as a draft model. The timing between the release of those, and this article, makes me think its maybe one or both of those things
(BTW Anthropic only exists because Sam Altman is a liar, Dario admitted this.)
Except for, you know, all the outside investors and the forthcoming IPO.
Related: https://80000hours.org/2012/03/the-replaceability-effect-wor...
There's a more nuanced discussion that could be had about how to balance relevance with outside influence. But at a foundational level it should be acknowledged that the tradeoff exists, and that receiving outside investment can't alone be seen as evidence of corruption.
Besides that, there's more that can be said about other things like their corporate structure or the degree to which they accelerated the AI race.
Of course that's what Dario thinks because that's what every tech CEO thinks. Dario, Sam, Sundar, probably many Chinese CEOs as well. It's what everyone thinks. That's why they're competing so fiercely with one another. That's why they basically make all the same decisions. That's why we need properly open source AI.
The day Mythos class models are open sourced will not be a good day. I don't think you understand the impact that will have on the world and on cyber defenders everywhere. It will be pure chaos.
Even if you don't think Mythos-class is the bar, open source has to stop at some point, you don't hand everyone a superweapon.
What kind of rosy-eyed chump believes in the "tech leader with scruples" bullshit? It always lies.
Did some people just ignore Mark Zuckerberg and Tim Cook's sociopathy, somehow? Did anyone buy into their "privacy is a human right" nonsense?
The thing I can't quite square is that it doesn't really fit my lived experience. I have known sincere, genuine people in the types of positions that I'm sure someone like you would declare to be sociopathic.
But beyond that, I just don't know why it would actually be true that everyone at the top is a villain. Why couldn't someone like Dario (or even Altman, gasp) be sincere? Because if he is, it does seem like a lot of the moves he's made would make sense given his worldview.
But if you assume he's just a villain, then you can twist any of those moves to just be further evidence of that which you already believe.
I don't know, I just find cynicism interesting, and a little sad.
You don't have to assume anything. A true "good guy" doesn't openly say that he's fine with autonomous, AI-powered weapons being used against me, and mass surveillance applied to me and my family just because I don't live in the US. A true "good guy" doesn't say "privacy is a human right", and then immediately (and completely) bend the knee to an authoritarian government on this issue.
And about the mass surveillance, I don't see why the military should not use AI to do surveillance abroad.
Lots of nerds for some reason have made cynicism a personality trait. They think optimism/honesty is hopelessly naive, therefor cynicism is the correct default.
All have collaborated with the current US regime. All have shown signs of being quite willing to compromise their principles in order to make money.
I went in the opposite direction - how far can I push myself to see multiple facets of a story? That is a wild ride, and it gets progressively more wild.
Please, I'm dying to hear the optimist's take on Mark Zuckerberg's career. It wouldn't happen to be embarassingly foolish, would it?
Why should I treat Sam and Dario with special white gloves? Are they different, this time? They have peers in China that do the same research and actually release it to the public. They let you run the production weights on your own machine. Am I a cynic, for comparing these CEOs to their populist superiors? Am I stupid for assuming their hostility when they refuse to give us the benefit of the doubt?
I'll believe their actual altruism when I see it. Both are seeped in "boy genius" puffery and lie out their ass. If this is the future of intelligent innovation, then America is truly declining.
https://github.com/openai/codex/issues/30364
What you think could be a big chunk, is more likely to be a fraction of a percent of queries.
And what use is similar query caching - so you (very often! if actually cost effective, maybe half the time) get a response to a query that was different from yours. Including for when you have a lot of context input already. You’re going to get trash.
If it were constrained to only very common initial prompts, and somehow the long tail did not actually dominate as it does with Google search (can't find the reference at the moment but it was a famous article some years ago), it also wouldn't account for serious enough cost savings. Long context is what is expensive.
This might only work in constrained domains like customer service where there’s tolerance for generic answers and escalation paths. For technical work? For general purpose use, with secretly canned responses charged at full price?
By contrast, when coding, devs typically have hundreds of thousands of tokens in the context window, and may use many millions of input tokens per day.
Caching requires the full prefix to match exactly. If a single word differs near the beginning of the prompt, nothing after that can share the cache. So this type of caching would save a few queries that cost virtually nothing, but wouldn't help with the stuff where cost matters.
“ Automatic Prefix Caching (APC in short) caches the KV cache of existing queries, so that a new query can directly reuse the KV cache if it shares the same prefix with one of the existing queries, allowing the new query to skip the computation of the shared part.”
https://docs.vllm.ai/en/latest/features/automatic_prefix_cac...
The technique you linked only makes a substantial difference for particular use cases where you are going to have many LONG CONTEXT queries with the same prefix. For instance, when having a set of documents that commonly get loaded in as context. It's a way for application developers to keep prefixes they manage (or prefixes managed by some set of their users) cached. It has no relevance for long tail general purpose use.
The transform script(s) are cached and can be played back or adjusted. Surely for some breadth of question inputs, they map more often to similar answers--but not static answers; instead, evented edits.
It's nearly untenable for a human to keep private edit scripts to generate code changes. The extra steps for custom regex, essentially one-offs for a shared codebase, is inefficient. But maybe not to an LLM.
Somebody has to finaly pay for these heaps of accelerator hardware.
> Tf is sol?
It's a proper noun: the name given to their latest and greatest model. Means "Sun" in Latin. Similar to how Anthropic has been naming its models Fable, Opus, Sonnet etc. Their other models are called Terra (Latin: Earth) and Luna (Latin: Moon) [0].
> Tf is ultra?
The name of the "harness" around the model. It'll use deeper thinking, subagents and all that jazz in response to a prompt. Other options include max, high, medium etc I suppose.
> Tf is codex?
"Coding agent" similar to Claude Code [1], something with a more descriptive name.
> Tf happened to descriptive nomenclature?
Something like GLM-4-32B-0414-128K (not made up [2]) doesn't quite roll off the tongue I suppose.
[0]: https://openai.com/index/previewing-gpt-5-6-sol/
[1]: https://openai.com/codex/
[2]: https://docs.z.ai/guides/llm/glm-4-32b-0414-128k
> Something like GLM-4-32B-0414-128K (not made up [2]) doesn't quite roll off the tongue I suppose.
Surely it would sell better though if they could communicate what they're selling?
> This process is automatic. Your browser will redirect to your requested content shortly. Please allow up to 0 second…
Previewing GPT‑5.6 Sol: a next-generation model
https://news.ycombinator.com/item?id=48689028
https://news.ycombinator.com/item?id=44344246
107 comments, 1 year ago.
I think gpt-5.5-pro runs 12x parallel gpt-5.5 agents behind the scene and uses OpenAI's secret sauce to synthesize their answers into one insanely good response.
If that's at all reflective of what it costs them to run it, I imagine they're in the same boat as Anthropic with Fable; they probably can't afford to offer it at subscription prices given current cost to operate it.
If 5.6 Sol Ultra has efficiency improvements (at one or more layers), and it allows OpenAI to offer a model that's competitive with Fable on the subscription plans, I'll guess a lot of folks will switch.
Fable is notably better than what came before. I watched it figure out stuff on its own over and over, on extremely hard problems, that I previously needed to guide a model to an understanding about, or work with them back and forth for several turns to figure it out together. Like, I've been reverse engineering a hardware device lately, and I've tried to tackle it a few times in the past with both some version of GPT and a couple of versions of Opus (most recently 4.7). In all cases, I barely made progress...would have gotten there eventually, probably, as I'm stubborn, but there were roadblocks constantly, with me and the models getting stumped and going around in circles in the end on every prior attempts.
Fable figured out other ways to find out what's happening, it dug into config files, found and extracted Boost-serialized data, compared that data to the observed behavior, built tools to compare the observed data with our emulated behavior, without being prompted. Would I have gotten there? Eventually, maybe. All prior models didn't; they mostly just tried the things I suggested and stopped at "well, that didn't work" or declared success after seeing results that matched their misunderstanding of the problem. I guess it's possible my prior attempts with other models had "loosened the lid" on the problem; we did already have a long list of documented "this didn't work" and a pile of tools for finding out if something worked. But, even so, I was impressed.
There probably will still be a "OK, let's rewrite this so it's not using lookup tables to precisely simulate the hardware behavior in software, because we don't need the noise, too" stage of the process...but, in one day with Fable, it solved a problem that I'd banged on for at least a week or too in the past with very little real progress. I don't think the models write exceedingly good code, even the best ones, but it sure does figure shit out quick.
https://github.com/agentify-sh/desktop
OpenAI models have always been the worst in my experience for verbose, slop formatted responses, with each generation increasing in sloppiness.
I'm not that impressed by Fable's writing to be honest, still has the AI giveaways like em dash.
I hate that I have had to remove it from my writing style because people assume it’s AI generated. But I think that ship has sailed. I’ll have to do without now.
more competition is always good for consumers.
Also, from the Guidelines:
> Be kind. Don't be snarky. Converse curiously; don't cross-examine. Edit out swipes.