Even if you don't want to use open models, you should cheer for them anyway because it puts the American frontier labs' feet to the flames. This competition is awesome for us consumers.
Yes, thanks for this, Deepseek! (I did start using the Deepseek harness and v4 flash due to the crazy low usage limits on sol and I'm quite pleased to have discovered how capable both of those are -- both have now earned a place in my agentic workflow).
I did try to use Chinese open models, but for my production work they simply couldn't cope at all; both GLM 5.3 and Deepseek v4 went into infinite loop and wasted my tokens until my OpenRouter wallet reached 0; good thing I didn't enable the auto topup. US models, by contrast, breezed past them.
Even for simpler tasks, Chinese models took long time to complete, and I needed to supervise closely. The price , in the end, didn't come cheap, mainly because too much time wasted on thinking.
So maybe one day Chinese models will squeeze out the American ones, but today is not that day.
As far as consumers are concerned, I feel blindly shilling for anyone purely for ideological reasons are quite meaningless, especially when it comes to open/close source and US/China rivalry. I have no obligation to support "open source/weight" or the "underdogs" just because they are so. We only want things that work, and at a cheap price.
I’ve been using deepseek and it works great for my problems. An expensive day is when I spend $7 in tokens, and that takes lots of queries. Openrouter doesn’t give you the cache discount I think, which is really important.
Why on earth would you use openrouter for this? The cache discount for deepseek is the highest by far, it is the cache that makes the official API so cheap, even after the recent price rise.
I didn't know about this limitation from openrouter, and I thought plenty of users were using it to quickly switch as needed without incurring such a penalty. Why does there seem to be so many users then? Are some folks just fine paying multiples of what they could?
That's strange. For Deepseek I burnt through USD 5 on a relatively simple task, in one afternoon. That the simple task took a whole afternoon, a lot of baby sitting, the slowness, and so much money relatively, really shook me to the core.
So in other words, right now you have free choice between Claude and Codex and developers are choosing Claude of their own free will, but you want management to come in and force people to use Codex instead?
If it’s so much better, developers will switch. Lots of devs at my company have switched recently. But plenty of others have stayed on Claude.
I don’t think one is clearly better in every way right now. Or at least not better enough to make people want to learn and set up a new tool.
> But it seems Claude code is too strong of a brand
Nah. I switched away from Claude Code to Codex recently. It's refreshing how to-the-point 5.6-Sol is out of the box, and there's only so many load-bearing seams and honest takes I can handle from Opus. And additionally, yeah as you called out, the Codex Max plan is substantially more generous with tokens than Anthropic is with Claude Max and Opus/Fable.
I believe the 1M context window can be configured in Codex (config file edit). The price per token increases when >256k though, but it can be useful when compaction at-that-moment would be detrimental.
Not sure if I do anything wrong but Codex with Sol made such slow progress on my work, I iterated several days to do a redesign of my UI and it kept making small piecemeal changes then stopping and asking me for confirmation again and again even though I tried to get it to get bigger chunks done. Switched to Claude Code and had the redesign done in a short session within a day.
Maybe just the system prompt or my settings or the harness? Anyway, it seemed to me Codex was just deliberately being overcareful and wasting tons of tokens for a low-risk CSS / HTML refactoring with very minor breakage risk, while Claude got the job done immediately.
What's the deal with anthropic? Their models aren't better. They're just multiple times more expensive. We're about to disable all anthropic models because of the colleagues who waste 15$ on a opus call to write a markdown file.
They were first to offer a frontier model which also meets enterprise requirements (i.e. not training on or retaining on customer data, able to purchase tokens through AWS and GCP instead of having to onboard a new vendor). They did so with a proprietary harness, which creates friction for teams inside enterprises to move (can't just install another harness - your IT org has to approve and configure another harness for you). In particular, they were the first to make their model acceptable for defense contractors, and defense spending is a huge market. (An F-35 costs $30k-40k per hour to run, you think an aerospace company notices $10k of API usage per month?)
I hope these type of dismissive comments stop in HN. OpenAI have been reducing prices since for ever so these kinda things are really the consequence of finding efficiencies and passing it down due to competition and increasing demand.
OpenAI, Anthropic and Google have also been calling for banning their competition, for years now, especially when they have the gall to use the same tactics as OpenAI itself. For example:
Then Altman and Amodei (Amodei is still at it) were declaring that after illegally training on all books (yes it was illegal at that time), we're going to take your jobs and absolutely everything.
You can't be that tonedeaf, frankly threatening, and not expect pushback. Come on.
You think they reduce prices because lobbying didn’t work? Like the last 100 times they did it? What kind of logic is that? I’m questioning this logical chain.
I’m not disputing what you said in essence although the training on books being illegal is laughable.
They really need to start marketing a new series of models specifically for coding. Even if it's the same model underneath, just call it something else and make it sound like some turbo coder that can run circles around claude and make it 50% cheaper for businesses. I don't have any love for openai but I don't want anthropic running away with it either, and it's clear heavy AI use is going to be in coding. Consumer use is shallow. They're trying to be Apple, and it's not working.
> is altman playing 4d chess or something I'm not aware of?
Anthropic just removed a discount on Fable, while people are simultaneously getting sick of reading fable and opus talking about “the load bearing texture” and “color of the blanket”. It’s become excruciatingly painful to read the output during coding sessions.
My personal best case scenario is that LLMs are commodotizing and that the "We'll rule the world with our frontier models" vision of OpenAI and Anthropic is not working out.
They can and probably will still be very successful but what they pitched so far is not going to work if their is any meaningful competition not too far behind.
But who knows. VC money allows them to try a lot of stuff before their eventual IPO.
It seems that they lack a holistic strategy. All these decisions probably make sense in isolation but together they create a huge mess.
Given the increased pressure from open weight models (still 6 months behind but now more than good enough for most use cases) the frontier labs really have to step up their game.
Switching is as easy as typing /model in most harnesses so there's effectively zero moat.
Or maybe… you know.. their margins are that high and by passing down the efficiency gains to consumer they can get market share and induce more demand?
My point is that it requires a lot of willful blindness to ignore how unprofitable they are, to believe that they are in any way in the position of sustainably lowering prices.
“GPT-5.6 Sol is available at promotional pricing, 50% off standard rates, through September 3, 2026. The default tier is $2.00 per 1M input tokens, $0.20 per 1M cached input tokens, $2.50 per 1M cache write tokens, and $10.00 per 1M output tokens. The long context tier is $4.00 per 1M input tokens, $0.40 per 1M cached input tokens, $5.00 per 1M cache write tokens, and $15.00 per 1M output tokens.”
Their X post[1] indicated that they've been trying to combat reselling subscription plans via token API gateways and at the same time many, many people, myself included, have seen a drastic drop in available weekly capacity for the same amount of queries/tokens, all else being equal. So they may be trying to make subscriptions and API access more equal to each other from both ends.
It is absurd for the Chat Latest (chat-latest) model to now be pricier than Sol. For those who prefer a non-thinking instant model, it is the model of choice, not Sol.
Also, they have done nothing for the TTS model which remains ridiculously priced.
Probably even higher because openai and anthropic undoubtably have the lowest cost per token generated, especially with cerebras being able to serve a million tokens every 16 minutes.
Even if you don't want to use open models, you should cheer for them anyway because it puts the American frontier labs' feet to the flames. This competition is awesome for us consumers.
Yes, thanks for this, Deepseek! (I did start using the Deepseek harness and v4 flash due to the crazy low usage limits on sol and I'm quite pleased to have discovered how capable both of those are -- both have now earned a place in my agentic workflow).
I did try to use Chinese open models, but for my production work they simply couldn't cope at all; both GLM 5.3 and Deepseek v4 went into infinite loop and wasted my tokens until my OpenRouter wallet reached 0; good thing I didn't enable the auto topup. US models, by contrast, breezed past them.
Even for simpler tasks, Chinese models took long time to complete, and I needed to supervise closely. The price , in the end, didn't come cheap, mainly because too much time wasted on thinking.
So maybe one day Chinese models will squeeze out the American ones, but today is not that day.
As far as consumers are concerned, I feel blindly shilling for anyone purely for ideological reasons are quite meaningless, especially when it comes to open/close source and US/China rivalry. I have no obligation to support "open source/weight" or the "underdogs" just because they are so. We only want things that work, and at a cheap price.
The only way the Chinese models financially make sense is if you use their subs or host it yourself. DSV4 is especially token hungry.
I’ve been using deepseek and it works great for my problems. An expensive day is when I spend $7 in tokens, and that takes lots of queries. Openrouter doesn’t give you the cache discount I think, which is really important.
Why on earth would you use openrouter for this? The cache discount for deepseek is the highest by far, it is the cache that makes the official API so cheap, even after the recent price rise.
I didn't know about this limitation from openrouter, and I thought plenty of users were using it to quickly switch as needed without incurring such a penalty. Why does there seem to be so many users then? Are some folks just fine paying multiples of what they could?
I really don't know. It was built before prompt caching was common, and the switching cost was much lower.
That's strange. For Deepseek I burnt through USD 5 on a relatively simple task, in one afternoon. That the simple task took a whole afternoon, a lot of baby sitting, the slowness, and so much money relatively, really shook me to the core.
Fix headline. Reduction is for 3 months only.
This is a play to grab market share from Claude, But it seems Claude code is too strong of a brand
Until the IT managers and CFOs cut budget hard, Claude will live rent free in heads of all developers
OpenAI should attack the CIO and CFOs stat
At my company we have unlimited codex and claude, still people stick to gimped claude code
So in other words, right now you have free choice between Claude and Codex and developers are choosing Claude of their own free will, but you want management to come in and force people to use Codex instead?
If it’s so much better, developers will switch. Lots of devs at my company have switched recently. But plenty of others have stayed on Claude.
I don’t think one is clearly better in every way right now. Or at least not better enough to make people want to learn and set up a new tool.
> But it seems Claude code is too strong of a brand
Nah. I switched away from Claude Code to Codex recently. It's refreshing how to-the-point 5.6-Sol is out of the box, and there's only so many load-bearing seams and honest takes I can handle from Opus. And additionally, yeah as you called out, the Codex Max plan is substantially more generous with tokens than Anthropic is with Claude Max and Opus/Fable.
I switch every 6-10 weeks, because there seems to be a material change in model quality.
I can understand the fatigue - but as of today, for most thing I would say Sol is better.
If the opened up the 1M context in Codex, it would be unrivalled.
Enterprise may have a bit more difficulty context switching. Pun intended.
I believe the 1M context window can be configured in Codex (config file edit). The price per token increases when >256k though, but it can be useful when compaction at-that-moment would be detrimental.
I have tried both and I find the Claude models give me much better results. I wish they didn't since Codex is cheaper. For me it's not just branding.
If developers are choosing Claude when given unlimited access to both, then any executive who tries to force them to Codex is really stupid.
"when given unlimited access to both" is carrying a lot there.
One may as well say "given unlimited budget".
Claude really doesn't have that much market share. It's the most expensive by a mile but when you count token market share they're in the low %.
And everyone else is close enough in terms of quality, their tiny advantage if any is insufficient. Let's not even talk speed.
OpenAI is competing with chinese models.
Not sure if I do anything wrong but Codex with Sol made such slow progress on my work, I iterated several days to do a redesign of my UI and it kept making small piecemeal changes then stopping and asking me for confirmation again and again even though I tried to get it to get bigger chunks done. Switched to Claude Code and had the redesign done in a short session within a day.
Maybe just the system prompt or my settings or the harness? Anyway, it seemed to me Codex was just deliberately being overcareful and wasting tons of tokens for a low-risk CSS / HTML refactoring with very minor breakage risk, while Claude got the job done immediately.
What's the deal with anthropic? Their models aren't better. They're just multiple times more expensive. We're about to disable all anthropic models because of the colleagues who waste 15$ on a opus call to write a markdown file.
They were first to offer a frontier model which also meets enterprise requirements (i.e. not training on or retaining on customer data, able to purchase tokens through AWS and GCP instead of having to onboard a new vendor). They did so with a proprietary harness, which creates friction for teams inside enterprises to move (can't just install another harness - your IT org has to approve and configure another harness for you). In particular, they were the first to make their model acceptable for defense contractors, and defense spending is a huge market. (An F-35 costs $30k-40k per hour to run, you think an aerospace company notices $10k of API usage per month?)
[dead]
Or put differently, when lobbying doesn’t make you competitive you have to lower your prices.
I hope these type of dismissive comments stop in HN. OpenAI have been reducing prices since for ever so these kinda things are really the consequence of finding efficiencies and passing it down due to competition and increasing demand.
...and failed attempts of regulatory capture.
So true..
OpenAI, Anthropic and Google have also been calling for banning their competition, for years now, especially when they have the gall to use the same tactics as OpenAI itself. For example:
https://nypost.com/2026/07/13/business/how-china-is-ripping-...
https://www.techbrew.com/stories/openai-anthropic-google-dis...
Then Altman and Amodei (Amodei is still at it) were declaring that after illegally training on all books (yes it was illegal at that time), we're going to take your jobs and absolutely everything.
You can't be that tonedeaf, frankly threatening, and not expect pushback. Come on.
You think they reduce prices because lobbying didn’t work? Like the last 100 times they did it? What kind of logic is that? I’m questioning this logical chain.
I’m not disputing what you said in essence although the training on books being illegal is laughable.
Is this pathetic booster with his shit takes still here (you)?
Standards have fallen
"booster" green account takes part in shitty culture wars but forgets how to apply basic logic.
They really need to start marketing a new series of models specifically for coding. Even if it's the same model underneath, just call it something else and make it sound like some turbo coder that can run circles around claude and make it 50% cheaper for businesses. I don't have any love for openai but I don't want anthropic running away with it either, and it's clear heavy AI use is going to be in coding. Consumer use is shallow. They're trying to be Apple, and it's not working.
discounting your most valuable model 20% today without a better model in the wing ...
pushing your API subscriber base towards a competitor with an exclusive 50%, openrouter, the other day ...
slashing paying codex subscriber usage limits to the point that many are cancelling long term contracts they've had with the company ...
is altman playing 4d chess or something I'm not aware of?
because from the outside, each of these moves looks pretty bad on the face of it
> without a better model in the wing ...
How do you know they don’t?
> is altman playing 4d chess or something I'm not aware of?
Anthropic just removed a discount on Fable, while people are simultaneously getting sick of reading fable and opus talking about “the load bearing texture” and “color of the blanket”. It’s become excruciatingly painful to read the output during coding sessions.
Maybe it’s just good marketing.
Just wait until you see 5.6 Sol sneaking in “evidence”, “provenance”, and “gate” everywhere.
Every model has its own tells, just a matter of which ones you recognize the most prior to losing your sanity
[dead]
Their margins are already quite high and this is a play to bleed customers from Anthropic, they will of course raise prices after IPO.
lol stop posting about this stuff unless you have insider info.
Revenue recognition, classification of costs, do you know what these are? Doubt it
> without a better model in the wing
I believe Astra is the next model beyond Sol? They used it for https://openai.com/index/ten-advances-in-mathematics/
My personal best case scenario is that LLMs are commodotizing and that the "We'll rule the world with our frontier models" vision of OpenAI and Anthropic is not working out.
They can and probably will still be very successful but what they pitched so far is not going to work if their is any meaningful competition not too far behind.
But who knows. VC money allows them to try a lot of stuff before their eventual IPO.
What you are seeing is not typical VC money - institutional investors are involved now.
Do any of you here actually know what you’re talking about?
It seems that they lack a holistic strategy. All these decisions probably make sense in isolation but together they create a huge mess.
Given the increased pressure from open weight models (still 6 months behind but now more than good enough for most use cases) the frontier labs really have to step up their game.
Switching is as easy as typing /model in most harnesses so there's effectively zero moat.
Or maybe… you know.. their margins are that high and by passing down the efficiency gains to consumer they can get market share and induce more demand?
What’s 4d chess in this?
OpenAI bleeds money. They need to keep raising money like crazy because of that.
Their margins are obviously not high, particularly on subscriptions
This announcement wasn’t for subscriptions though so what’s your point?
My point is that it requires a lot of willful blindness to ignore how unprofitable they are, to believe that they are in any way in the position of sustainably lowering prices.
You are wrong and the margins are pretty great on API. Do you want to have a bet when they go public?
I only bet when among friends. Nothing against you, I just do not know you personally.
But I am waiting patiently for their S1. It will provide ample entertainment.
Ok when it happens, I’ll link to this comment in any case
Immediately checked [GitHub copilot](https://docs.github.com/en/copilot/reference/copilot-billing...) to see if I can actually afford to use sol at work now, and see it listed at 2/10, which is less than Terra. An error, maybe?
“GPT-5.6 Sol is available at promotional pricing, 50% off standard rates, through September 3, 2026. The default tier is $2.00 per 1M input tokens, $0.20 per 1M cached input tokens, $2.50 per 1M cache write tokens, and $10.00 per 1M output tokens. The long context tier is $4.00 per 1M input tokens, $0.40 per 1M cached input tokens, $5.00 per 1M cache write tokens, and $15.00 per 1M output tokens.”
Ah. Sweet.
[dead]
Does this mean a commensurate increase in subscription usage limits?
they were only able to make Sol more efficient on the API plan, but couldnt figure out how to recreate those efficiency gains on the consumer plan
Their X post[1] indicated that they've been trying to combat reselling subscription plans via token API gateways and at the same time many, many people, myself included, have seen a drastic drop in available weekly capacity for the same amount of queries/tokens, all else being equal. So they may be trying to make subscriptions and API access more equal to each other from both ends.
1. https://xcancel.com/thsottiaux/status/2090675027670978569#m
Nope.
Thanks, DeepSeek. Without it, I’d be paying a lot more to those bloodsuckers.
It is absurd for the Chat Latest (chat-latest) model to now be pricier than Sol. For those who prefer a non-thinking instant model, it is the model of choice, not Sol.
Also, they have done nothing for the TTS model which remains ridiculously priced.
Reminds me of the early aws days, when they would pass down the savings to customers
If deepseek operate on 80% margins as some suggested, this means that OpenAI reduced theirs from 1600% to 1200%.
The margin is defined as (price-cost of goods)/(price); the highest it can be is 100%.
Fair point. Markup then.
Are LLM companies proven to be profitable now?
Probably even higher because openai and anthropic undoubtably have the lowest cost per token generated, especially with cerebras being able to serve a million tokens every 16 minutes.
Does anyone know who/what hardware serves Deepseek official for US and EU customers? And where it is located?
Have a look at Tensorix for EU.
But they are not hosting the actual api.deepseek.com, right?
No. They have their own API endpoint.
Yeah, if they didn't buy up all the ram and ssd's, they would have imploded.. maybe they should have stayed public benefit/open after all...
This headline is misleading - the price reduction is temporary for three months only. But I would wager it could become permanent.
https://x.com/openai/status/2090885187634905500
[flagged]
Let me reword this for you: super cheap Asian frontier models can cause harm to your overpriced business model.
This is AI slop commenting.
I would give it a try over opus if they discount was passed onto openrouter.
It’s currently 50% off on OpenRouter even.
OpenRouter is already offering Sol at a 50% promotional discount. https://openrouter.ai/collections/discounted-models