> I even got a warning on my OpenAI account.

This idea of software threatening the user with consequences is totally wild and dystopian. Fellow developers, what kind of world have be built? This is insanity. Imagine if my hammer told me, "Hey, you shouldn't use me on screws--only nails. Do it again and I'll self-destruct!" WTF people, stop making this kind of software!

We have an enterprise cursor account so I can try all the mainstream models. Using composer 2 on our own code which I obviously have the source code for I couldn't get it to turn on a debug flag to bypass license checks while I was troubleshooting something. Infuriating. It was like that old Patrick from SpongeBob meme.

I don't understand why we would turn the models into law enforcement officers. Things that are illegal are still illegal and we have professionals to deal with crimes. I don't need Google to be the arbiter of truth and justice. It's already bad enough trying to get accountability from law enforcement and they work for us.

> I even got a warning on my OpenAI account.

This is kind of terrifying to me, regularly. No real manner of recourse to normal people without a following, potential exclusion from real fundamental tooling. Imagine OpenAI goes on to buy 20 companies and now you cant use Figma, Next, whatever just because you once tripped some very foggy line somehow. Not just OpenAI but the entire ecosystem is so... hard to read.

I was asking Gemini about a quote from catch 22 and it kept dying mid stream saying it cant talk about it, god knows why, it had no violent or sexual content -- though that is in the book. I could imagine it dinging my whole workspace account just because ... shrug?...

I know ideally the future is local, but I don't know how real that is for most people at least in the next few years with practical costs and power usage except I guess through a M* processor if you're in that ecosystem.

Buying it now to test this out, I’ve been looking for a model that doesn’t treat me like a child lol

Speaking of this: is anyone working on binary to source decompiler models? Seems like a no brainer and I could see it working exceptionally well especially if they were fine tuned for each language. So if you can tell it’s a Go binary use a Go model, etc.

>> I even got a warning on my OpenAI account.

I was using GPT 5.5 through Cursor recently, and it found what it thought to be a security-related issue. I read the code, didn't see what it was seeing, and said "Run the chain of operations against my local server and provide proof of the exploit."

It thought for a few seconds, then I got a message in the chat window UI saying OpenAI flagged the request as unsafe, and suggested I use a "safer prompt."

Definitely soured me on the model. Whatever guardrails they are putting are too hamfisted and stupid.

> even got a warning on my OpenAI account

Edit: https://chatgpt.com/cyber

Silicon Valley has do to dirty tricks now. Next phase is they win....

"A Dark-Money Campaign Is Paying Influencers to Frame Chinese AI as a Threat" - https://www.wired.com/story/super-pac-backed-by-openai-and-p...

Are you kidding? Ask this question and see what answer you get: What famous photo depicts a man standing in front of a line of tanks?

I've been having the same experience. Tasks like "go through this entire module and pedantically make it match my preferred styleguide exactly" were not worth a couple dollars with frontier models. It's nice to be able to put deepseek flash on stupid, unnecessary or highly speculative tasks without thinking about the cost.

I find a lot of the inefficiency also comes from the model just randomly poking around and grepping all the time which is the fault of the harness. I ended up building a Prolog based MCP where I use tree-sitter to parse the code into a graph, and then the model can just ask questions like 'what are all the functions connected to this function'. So, in case you're trying to focus on what a particular endpoint is doing, you can trivially and predictably trace the whole subgraphs of calls.

https://github.com/yogthos/chiasmus

> would have burned somewhere between $9 to $13 easily with not much benefit

With not much benefit compared to DeepSeek v4 Pro @ 9 cents (1/100th of the price) or did neither offer any benefit?

Even taking into account the fact that they are billing at 75% discount it's still quite cheaper

How did you use it? OpenRouter, or provider directly?

Only similarity it has to Opus 4.6 is the 4 in the name. I do not understand these dishonest comparisons. OOS models are vool, cheap and promising for a future -- but why are we pretending they are better than they are?

This is very false DS4 is super cheap. I would advise to begin by reading their release paper. https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro/blob/main...

They introduce very novel methods to improve long context efficiency and attention. HCA & mCH. It requires only 27% of flops for inference and 10% for KV cache than v3.2. This makes it super efficient. Think of this. For flops, we can now serve more than 3x the amount with the same number of compute, and you would need 30% of prior KV cache.

Furthermore, this release is a PREVIEW, DeepSeek is the real open labs and they not only cook up quite a bit with every single release, but they publish and share it. I'm running this locally.

Let me tell you how "CHEAP" this is. With v3.2 I would run out of GPU ram, spill into system ram with 256k context. It ran quite alright and I was happy with my 7tk/sec. With this, I'm 100% in GPU ram with full 1million token, run more than 2x fast while getting better results.

This is super cheap. moonshot has made it clear that they are starved for GPUs and that's why. If they had GPU capacity like we do in US and subsidized the models like we do here, they would be giving it away for free!

Sure that can happen but it hasn’t been my experience. I just spent a whole day using it for some pretty hefty refactors, many rounds of back-and-forths, thousands of lines of code changes, reviews, investigations, many subagents running parallel tasks, the works. Total cost $0.95, altogether.

I had attempted this with Opus 4.6 in the past and it burned through the $10 budget I’d given it before it returned from my initial prompt.

Even if it’s heavily discounted, it would still have cost me single digits for a complete solution vs double-digits for exactly nothing.

Might be a dumb question but do you have to read the files in the same order in new sessions to ensure the correct prefix for the cache?

The cool thing about open-weights model is that you are free to use alternative providers that won't phone home to the original model creators.

I see 6 alternative providers listed on Openrouter for DeepSeek V4 Pro for example.

> I'm surprised that people here don't care at all about these models openly training on your data

You can use zero data retention and zero training providers for most open weights. See OpenRouter and OpenCode Go/Zen for examples.

This is actually one of the big selling points behind open weights - neither China nor the US get your data.

I am personally okay helping them as long as they publish the models and dont keep them closed. And I dont trust the settings where providers say they wont train on it.

Because they give it away for free and offer APIs at very acceptable rates. Not that hard to figure out, Robin Hood stealing our data tax back comes to mind.

At this point, that's kind of the reason I use open-weight models through the official providers when I can now.

There's some use cases I won't use a hosted model for, and will only do self hosted.

Otherwise, if they're going to keep releasing open-weight models, I'm going to keep giving them data.

I am fine with them training on my open source code (which is pretty bad but not the point, because they're providing the service for free). I will be super pissed if I pay for enterprise and they train on it though. I believe this is the opinion of majority programmers.

If the data is opensource on github, then in my opinion it should be fair game.

My policy is that I don't allow agents to access all code. Some of it is shielded behind bind mounts. Maybe this is a pathetic, artisanal (or ego-driven), reaction of mine to the inevitable. I allow them to work on about 90% of the code (most codebases fully), with some code being considered too valuable to expose to the vendor. When data is involved, LLMs only get to see anonymized data.

This cute policy of mine won't affect anything though. The more we use the models, the more the models will replace this kind of work. Centralisation of power is inevitable; in Medival Europe, we used to have state & church ruling. In modern times but before the internet, it was probably state and banks. Maybe with ongoing digitization (bank offices disappearing) making banks less costly to operate; combined with with bank bailouts, maybe govenments will fully nationalize or at least banks will consolidate.

Then the AI companies will consolidate with the internet information and communication companies (Google/Meta for the US, and Alibaba/Tencent for China). Maybe we'll end up with a few de-facto governmental megacorps that rule in tandem and close cooperation with the formal government, who might handle mostly infra, utilities and the army. The megacorp would control narrative more and take more of a paternal role (educating and protecting the citizens, normally handled by formal governments).

Does this make sense?

You definitely have a bone to pick. Chinese researchers usually have given the world the most cheap and consistent high quality research around LLMs. They don't pretend, they do the work and release the goodies. Mostly so cheap, every one in the world has a chance to use close to frontier models. Why would you respond with "Anger"?

You let us know what your real complaint is about and let's not feign indignation at open models and research.

AWS Bedrock has DeepSeek models running on their infrastructure. That should be enough to prevent training on user data (there's a markup compared to DeepSeek's pricing though).

And unfortunately AWS doesn't have prepaid billing, so you can't just give the internet access to your API key without getting FinDDoS'd.

What do you mean specifically? Data passed through OpenRouter? Or that they too indiscriminately ingest data all over the web? If the former, I assume it's just that anyone still using them just doesn't care where the data comes from. If the latter, well, it seems like every day there's some news on some new model from somewhere, and it takes dedication to complain every time. There's also the factor that I believe DeepSeek is more open with the model, while others keep it entirely proprietary, which feels fairer and (personally) is also less offensive.

If they give me the resulting model in the end, they can train on my data all they want. Hell, I'll send them more of it.

Two factors. First is anti-americanism (or at least anti-american-capitalism).

But the more important one is the social contract. Github came far before LLM era. The branding around it is being the storage of open source projects and many users want to it stay away from AI hype. You won't expect LLM providers to stay away from AI hype (duh) so it's less an issue for them.

As opposed to?

Do you really think OpenAI, Anthropic or any other entity in the same business respects your data?

The Chinese AI companies who release open weights actually deserve whatever input you give them. They are the reason why there is competition and not duopolies in the domain.

IMO there are two things that make me optimistic that we won’t see a big rug pull where price-to-capability ratio skyrockets relative to today:

* As you’ve noted, people keep finding ways of slamming more intelligence into smaller models, meaning that a given hardware spec delivers more model capability over time.

* Hardware will continue to improve and supply will catch up to demand, meaning that a dollar will deliver more hardware spec over time.

I hope that one day we’ll look back on the current model of “accessing AI through provider APIs” the same way we now look back on “everyone connecting to the company mainframe.”

By the time the dust settles I wouldn't be surprised if personal interactive usage couldn't even be had for under $200. I can't fit my modelling of the serving costs of these things to any public reporting, even the more bearish examples

You can use Tinfoil for inference, which lets you use the model in the cloud while getting similar privacy as running locally: https://tinfoil.sh/inference.

Disclaimer I'm the cofounder. This works by running the model inside a secure enclave (using NVIDIA confidential computing) and verifying the open source code running inside the enclave matches the runtime attestation. The docs walk you through the verification process: https://docs.tinfoil.sh/verification/verification-in-tinfoil

Thanks for sharing your experience, I’m looking to try it out.

Which provider are you using for inference? Opencode or the DeepSeek api?

Do you have a link to this?

The article quotes the full price.

obviously everyone subsidizes for user acquisition - after all people need to be coaxed to test your model, claude code subscriptions come to me one.

DeepSeek pro is 65/86% cheaper (i/o tokens) in subsidized pro vs pro and 91/97% cheaper with current subsidies.

Flash vs Sonnet 4.6 is 95/98%

Yeah even the Chinese open models have a problem that inference costs for these aren't that cheap. The only way out for the AI bubble collapse is simply more efficient hardware at lower costs and infrastructure setup downtime.

I am using flash, and it's so good. 150M tokens at $2.

I’ve found that if I turn off auto mode, I get much more usage from the $100/mo plan.

What would be the non-subsidized price for a V4 api? Can it be priced 3x cheaper than bigger models? In Openrouter, this 1600B param model costs 0.4$. Whereas Kimi 2.6, 1000B params is 0.7; GLM 5.1, 754B params is 1.0$.

Someone on Twitter got >200M tokens for around $10 at the current pricing level

Did you do this test through OpenRouter?

> "actually no", "hang on", "wait that makes no sense"

Claude does the same thing, claude code just hides the thinking now

Before CC and Codex removed thinking/verbose and hid most of it, both do that .

You can just use it through Claude Code, so you get to keep the system prompt and tooling you are used to.

3rd party models are a drop-in replacement with `ANTHROPIC_BASE_URL` in Claude Code, something people seem to miss right now. And contrary to what Anthropic might like to have you think, you don't need Opus 4.7 to run the harness to get similar performance.

https://api-docs.deepseek.com/quick_start/agent_integrations...

I feel the reasoning might be tuned for hard questions and not agentic work. I feel it overthinks, good for a very hard question, not for small incremental agentic steps. In theory, disabling thinking and using really well formed instruction, forcing it to still emit a bunch of tokens each step prior to taking action, could help. Only one way to find out though.

> It tried to build something simple and while it got the job done the thinking displayed did not fill me with confidence. It was pages and pages of "actually no", "hang on", "wait that makes no sense". It was like the model was having a breakdown.

It has been probanly trained to assess its own "thoughts" regularly and outputs those for the assesment results. I wouldn't worry much about the reasoning text contents, and it's nice to have them in contrast to the closed model "summaries", so it's easier to see what's going on.

Opus 4.6 and GPT 5.4 do the same thing through GH Copilot and Bedrock. I get plenty of "Actually the simplest solution is ..., wait no, actually I should do ..., the best fix is ..."

Eh, you're seeing raw thinking tokens. With Claude <x> 4, and I think GPT-5 series, you are no longer seeing real thinking tokens, but "summarized" tokens that are probably highly different to the raw thinking.

use hide_thinking in opencode to get the claude experience :p

I see similar things using GLM 5.1 in pi.

I had to turn off thinking traces because it was just giving me anxiety looking at it.

> Bear in mind open code was also new to me so I could be just seeing thinking where I usually don't

Well there's your problem.

Edit: I remember seeing similar things with ChatGPT or Codex, although I can't remember in which context.

One shot canvas and svg images or animations are also just something that at this scale shouldn't be an issue at all, even Qwen running locally on 24gb cards can do impressive ones.

Don't understand why this test gets any attention, I mean other than the pelicans which isn't a good test, theres no meat in this article.

Relevant: https://news.ycombinator.com/item?id=47839493

It also seems like all of the models have converged on very similar images.

Isn't it widely speculated that these are distilled from current frontier models? Distillation is far less compute intensive than primary training. That said, if distillation produces something almost as good for a fraction of the cost, Jensen's point may stand.

These were trained on NVIDIA gpus. It is running inference on Huawei.

It's too late already, that ship has long sailed. China has the know how in software and hardware. They don't need American tech, they just want it because it's convenient.

The embargo won't backfire, because any delay of China's development was worth it to the US. The situation was never, "China wasn't developing AI chips, now it is", it was always, "China IS developing their own AI chips, let's just slow them down as much as we can."

Have you given GLM 5.1 or Kimi K2.6 a shot for coding? They outperform Deepseek v4 pro.

Keep in mind that DeepSeek has a max thinking mode of its own in the API.

I tried it for two tasks using Claude Code, on max effort.

1. Web platform, asking it to analyse a feature to create reports, and coming up with better solution and better UX. it did great, I would say on par with Sonnet 4.6 or even opus considering the thinking and explanation

2. Mac app with some basic functionality, it did well from functional perspective but then I used Opus 4.7 to evaluate and suggest improvements, where I noticed it missed many vital points in design system and usability.

I think it’s a leap, I haven’t used a model this capable that is not OpenAI or Anthropic

That's what an evaluation dataset is for, create your own and you can bench a model in a few hours to see if it fits your needs.

DeepSeek V4 Pro has about 25GB worth of active parameters, so if you can fit the whole ~870GB weights + cache in RAM your tok/s is bounded above by 25GB divided into your system memory bandwidth in GB/s. If you can't fit your whole model in RAM you'll be bottlenecked to some degree by storage bandwidth which is in the single or low double digits in GB/s.

Mind you, it's an absolutely sensible setup either way if you are just testing a few queries and are willing to run them unattended/overnight. Especially since the KV-cache size is apparently really low (~10GB is said to be typical) so you get a lot of batching potential even in consumer setups, which amortizes the cost of fetching weights.

Because V4 doesn't even beat Kimi K2.6 and GLM 5.1, which have been out longer. It's only talked about as much as it is because it's Deepseek and R1 was the first open source reasoning model. V4 isn't even multimodal (unlike Kimi) and the 1M context doesn't seem to perform particularly well.

They were and are still great for coding. They were not trained for agentic workflow and coding harness.

Huh? R1 was one of the earliest openly available MoE and reasoning models, that's definitely not "hype". People tried to do reasoning before by asking the model to "think it through step by step" but that was a hack. The later V3.1 and V3.2 releases AIUI unified reasoning/non-reasoning use under a single model.

Is there real evidence that the volume was meaningful for distillation vs say extensive benchmarking and testing?

It’s certain all the labs use each others APIs extensively for testing - what’s the actual evidence that Deepseek was at significantly higher scale etc.?

How immoral of those LLM developers. The rest of the field does such a good job of crediting their inputs.

And the credit of OpenAI is to Google?

https://arxiv.org/abs/1706.03762

Aw man, I'm going to shed a tear, the poor AI companies that stole books, works of art, writings any anything they could get their grubby hands on while happily telling everyone that their jobs are over by the exabyte are getting their precious little tokens stolen by big evil chinese LLMs :(

It's morally right to fuck over Anthropic (and OpenAI, or any other lab). Works generated by AI are not copyrightable anyways, and their terms of service have zero legal value.