GPT 5.6 Sol 20% price reduction

86 points by izakfr 18 hours ago on hackernews | 73 comments

If deepseek operate on 80% margins as some suggested, this means that OpenAI reduced theirs from 1600% to 1200%.

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.

downrightmike | 16 hours ago

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...

chvid | 16 hours ago

Does anyone know who/what hardware serves Deepseek official for US and EU customers? And where it is located?

BlackRabbit1 | 15 hours ago

Have a look at Tensorix for EU.

chvid | 15 hours ago

But they are not hosting the actual api.deepseek.com, right?

: BlackRabbit1 | 12 hours ago
No. They have their own API endpoint.

jsnell | 15 hours ago

The margin is defined as (price-cost of goods)/(price); the highest it can be is 100%.

: ReptileMan | 13 hours ago
Fair point. Markup then.

aitchnyu | 9 hours ago

Are LLM companies proven to be profitable now?

virgildotcodes | 16 hours ago

Does this mean a commensurate increase in subscription usage limits?

: Sabinus | 16 hours ago
Nope.
: mrtesthah | 16 hours ago
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
: rjtc | 15 hours ago
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

millsau | 16 hours ago

I would give it a try over opus if they discount was passed onto openrouter.

: Cu3PO42 | 16 hours ago
It’s currently 50% off on OpenRouter even.
: applfanboysbgon | 16 hours ago
OpenRouter is already offering Sol at a 50% promotional discount. https://openrouter.ai/collections/discounted-models

gentlewater | 16 hours ago

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?

seb2026 | 15 hours ago

“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.”

: gentlewater | 15 hours ago
Ah. Sweet.

johnnyApplePRNG | 16 hours ago

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

laichzeit0 | 15 hours ago

> 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/

LaurensBER | 15 hours ago

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.

teruakohatu | 15 hours ago

> 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.

: atmonostorm | 14 hours ago
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

Vespasian | 15 hours ago

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.

: euue | 9 hours ago
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?

simianwords | 15 hours ago

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?

surgical_fire | 12 hours ago

OpenAI bleeds money. They need to keep raising money like crazy because of that.

Their margins are obviously not high, particularly on subscriptions

simianwords | 12 hours ago

This announcement wasn’t for subscriptions though so what’s your point?

surgical_fire | 12 hours ago

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.

simianwords | 11 hours ago

You are wrong and the margins are pretty great on API. Do you want to have a bet when they go public?

surgical_fire | 7 hours ago

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.

: simianwords | 7 hours ago
Ok when it happens, I’ll link to this comment in any case

solenoid0937 | 15 hours ago

Their margins are already quite high and this is a play to bleed customers from Anthropic, they will of course raise prices after IPO.

: euue | 9 hours ago
lol stop posting about this stuff unless you have insider info.
Revenue recognition, classification of costs, do you know what these are? Doubt it

JSR_FDED | 16 hours ago

Or put differently, when lobbying doesn’t make you competitive you have to lower your prices.

simianwords | 15 hours ago

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.

postepowanieadm | 15 hours ago

...and failed attempts of regulatory capture.

: simianwords | 15 hours ago
So true..

spwa4 | 12 hours ago

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.

: simianwords | 12 hours ago
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.

gr_norm | 15 hours ago

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.

resters | 15 hours ago

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).

nsoonhui | 15 hours ago

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.

seanmcdirmid | 14 hours ago

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.

irthomasthomas | 14 hours ago

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.

hiq | 9 hours ago

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?

: irthomasthomas | 8 hours ago
I really don't know. It was built before prompt caching was common, and the switching cost was much lower.

nsoonhui | 13 hours ago

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.

: seanmcdirmid | 6 hours ago
What was the tax? You were using v4 flash right? What thinking setting were you using?
I’m pretty happy with it, it’s about as good as Gemini flash 3.5, the only other model I have deep experience with.

weiran | 9 hours ago

The only way the Chinese models financially make sense is if you use their subs or host it yourself. DSV4 is especially token hungry.

OutOfHere | 15 hours ago

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.

m00dy | 15 hours ago

Thanks, DeepSeek. Without it, I’d be paying a lot more to those bloodsuckers.

returnInfinity | 15 hours ago

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

brokencode | 15 hours ago

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.

: SR2Z | 6 hours ago
Management is going to force one model or another because that's how procurement works in most places.
For better or worse that really favors Anthropic because they were the first to land these contracts.

on_the_train | 15 hours ago

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.

: t098i3 | 15 hours ago
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?)

Wowfunhappy | 15 hours ago

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.

bluegatty | 15 hours ago

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.

: gatio | 14 hours ago
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.

ThePhysicist | 15 hours ago

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.

louiskottmann | 15 hours ago

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.

mcintyre1994 | 15 hours ago

If developers are choosing Claude when given unlimited access to both, then any executive who tries to force them to Codex is really stupid.

: gatio | 14 hours ago
"when given unlimited access to both" is carrying a lot there.
One may as well say "given unlimited budget".
: returnInfinity | 4 hours ago
yes in an ideal world, in business world its all wine dine, contracts, costs savings.
perhaps all the companies should dump teams and use slack.

lukeify | 15 hours ago

> 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.

simianwords | 15 hours ago

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

znpy | 14 hours ago

Reminds me of the early aws days, when they would pass down the savings to customers

albatross79 | 11 hours ago

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.

jms703 | 7 hours ago

Fix headline. Reduction is for 3 months only.

iJohnDoe | 5 hours ago

I'm not a cost expert for all the models. My usage is low enough not worry about it. However, I noticed today how expensive GPT 5.5 is. I realized GPT 5.6 was already cheaper before the price reduction and it's the newest model. I figured the newest shiny model should be pricier.

Are the older models supposed to get cheaper over time or do they stay expensive to encourage people to stop using the old stuff?