"OpenAI was plugging Claude into its own internal tools using special developer access (APIs)"
Unless it's actually some internal Claude API which OpenAI were using with an OpenAI benchmarking tool, this sounds like a hyped-up way for Wired to phrase it.
Almost like: `Woah man, OpenAI HACKED Claude's own AI mainframe until Sonnet slammed down the firewall man!` ;D Seriously though, why phrase API use of Claude as "special developer access"?
I suppose that it's reasonable to disagree on what is reasonable for safety benchmarking, e.g: where you draw a line and say, "hey, that's stealing" vs "they were able to find safety weak spots in their model". I wonder what the best labs are like at efficiently hunting for weak areas!
Funnily enough I think Anthropic have banned a lot of people from their API, myself included - and all I did was see if it could read a letter I got, and they never responded to my query to sort it out! But what does it matter if people can just use OpenRouter?
I agree, it does sound like they're hyping it up. But maybe the author was confused. In the API ecosystem, there are special APIs that some customers get access to that the normal riff-raff do not. If someone called those "special developer access" I don't think it'd be wrong
> Seriously though, why phrase API use of Claude as "special developer access"?
Isn't that precisely what an API is? Normal users do not use the API. Other programs written by developers use it to access Claude from their app. That's like asking why is an SDK phrased as a special kit for developers to build software that works with something they wish to integrate into their app
If I'm an OpenAI employee, and I use Claude Code via the API, I'm not doing some hacker-fu, I'm just using a tool a company released for the purpose they released it.
I understand that they were technically "using it to train models", which, given OpenAI's stance, I don't have much sympathy for, but it's not some "special developer hackery" that this is making it sound like.
Because it's not "special developer access". It's just "normal developer access". The phrasing gives an impression they accessed something other users cannot.
It would be normal standard English to assume that special modifies the word access. That would make the sentence semantically be the same as “special access, specifically the type of access used by developers”
Compare with a sentence like “the elevator has special firefighter buttons” which does not mean that only some special type of firefighter uses the button.
I think it's pretty reasonable to assume that if someone bothers to say "special developer access" rather than just " developer access", there must be some difference between the two. There's clearly not any reason that " developer access" wouldn't be sufficient to describe using APIs though, so it's hard not to read the word "special" as being at least redundant if not actively misleading.
In content intended for an audience of developers, it's reasonable to assume "special developer access" means access for special developers. If the audience is the general public, it would be sensible to interpret it as "special access for developers," in contrast to the normal sort of access most other people use.
Counter-point: as a wordsmith, it's incumbent on the article's author to make their point in an unambiguous way. In your example, rather than write "the elevator has special firefighter buttons", the author could choose to write "the elevator has buttons which are only available to firefighters". Or alternatively, "the elevator has buttons which are only available to certain firefighters".
The amount of care the author puts into their phrasing determines whether their point comes across as intended, or not. The average magazine reader can likely figure out that there's no such thing as "special" firefighters with "privileged" access to elevator buttons that other firefighters lack. They may not have the programming knowledge to do likewise with "developer access", even if they are reading a magazine like "Wired".
It's Newswriting 101-level stuff.-
In your example, "firefighter buttons" is a noun phrase which refers to a particular type of button. "Special" applies to the whole of "firefighter buttons," not just to "firefighter" and not just to "buttons." The same would apply for "special developer access."
> the elevator has special firefighter buttons
If you said that to anyone they'd assume there are non standard buttons beyond the normal "call" / "fire" buttons. Special changes the meaning in both sentences.
Firefighter buttons are meant to only be used in very rare special occasions (emergencies) so "special" is just emphasis, whereas developer access is a completely normal way to use the product and thus "special" suggests additional significance. Sure not everyone uses the product as developers, but then not everyone uses the 18th floor button either.
Developer access isn’t normal to a lay audience. That’s my point. To a lay audience developers are special computer expert people who do completely different things than they do.
From the perspective of a non technical reader developer access isn’t normal, it’s special.
The HN audience doesn’t see that. But the phrase isn’t confusing to normal people.
>That's like asking why is an SDK phrased as a special kit
It's Software Developer Kit, not Special Developer Kit ;-)
Most people reading Wired probably don't know what an API is.
Normal users use the API constantly, they just don’t realize it.
Isn’t half the schtick of LLMs making software development available for the layman?
> According to Anthropic’s commercial terms of service, customers are barred from using the service to “build a competing product or service, including to train competing AI models”
That's... quite a license term. I'm a big fan of tools that come with no restrictions on their use in their licenses. I think I'll stick with them.
These anti-competitive clauses are becoming standard across all major AI providers - Google, Microsoft, and Meta have similar terms. The industry is converging on a licensing model that essentially creates walled gardens for model development.
You guys are tripping. EULAs have had anti-competition, anti-benchmarking, anti-reverse engineering and anti-disparagement clauses since the late 90s.
These unknown companies called Microsoft, Oracle, Salesforce, Apple, Adobe, … et al have all had these controversies at various points.
Yeah, if I remember correctly iTunes had a clause it couldn’t be used for nuclear development.
Not sure what Apple lawyers were imagining but I guess barring Irani scientist from syncing their iPods with uranium refiner schematics set back their programme for decades.
I think Apple had it in all their software. It's a good stance and easy to ridicule by taking iTunes as an example.
That is hilarious if true.
Blame ITAR.
I am not a fan of Apple or Oracle, but you are not contractually prevented from competing with them if you use Macs or Oracle Cloud to build software.
I wouldn't suggest building on Oracle's property as you drink its milkshake, but the ToS and EULAs don't restrict competition.
Oracle licenses 100% restrict reverse engineering it's product to build a competing once, which is probably the closest to what these AI giants are trying to restrict.
Oracle db products are not meant to build databases, unlike LLM code generator which are meant to build any kind of software, so the restriction sounds a bit different.
Imagine if Oracle was adding a restrictions on what you are allowed to build with Java, that would be a more similar comparison IMO.
Yeah but did you know you also can't publish benchmarks?
E.x. if you make a product that works on multiple databases, you can't show the performance difference between them.
"Everyone else is doing it" doesn't make it right.
It also makes it dangerous to become dependent on these services. What if at some point in the future, your provider decides that something you make competes with something they make, and they cut off your access?
When that provider's ToS allows them full access to the inputs/outputs you're sending through their system, there is a strong incentive to build something competitive with you once you're proven profitable enough.
I don't know how companies currently navigate that.
Does this mean you can’t make a potential competitor to Claude Code using Claude Code, though?
Oh I wonder if that applies to me? I've been using claude to do experiments with using SNN's for language models. Doubt anything will come of it... has mostly just been a fun learning experience, but it is technically a "competing product" (one that doesn't work yet)
If you release it, it will be a competing product, experiments are just research.
But he’s building it right now, and the building part is what’s illegal. It’s a very gray area.
Also Twitter TOS when accessing firehose was that you could not recreate a Twitter client.
Same with Discord, for example. In fact, in another instance, my account got disabled for having used it for bots.
For years it was a license violation to use Microsoft development tools to build a word processor or spreadsheet. It was also a violation of your Oracle license to publish benchmark results comparing Oracle to other databases.
If you compete with a vendor, or give aid and comfort to their competitors, do not expect the vendor to play nice with you, or even keep you on as a customer.
>For years it was a license violation to use Microsoft development tools to build a word processor or spreadsheet.
source?
You have to go pretty far back, it was in the Visual C++ 6.0 EULA, for instance (for lack of a better link):
https://proact.eu/wp-content/uploads/2020/07/Visual-Basic-En...
It wasn't a blanket prohibition, but a restriction on some parts of the documentation and redistributable components. Definitely was weird to see that in the EULA for a toolchain. This was removed later on, though I forget if it's because they changed their mind or removed the components.
Doesn’t the ban on benchmarking Oracle still stand today?
Given the law firm in question which just happens to develop an RDBMS, I wouldn't want to find out
Besides, lol, who cares how fast a model-T can go when there are much nicer forms of transportation that don't actively hate you
Hmm so "because you split spending between us and a competitor, we'll force you to give the competitor the whole share instead!"
Certainly a mindset befitting microsoft and Oracle, if I ever saw one.
I can understand the benchmark issue. It often happens when someone benchmarks something, it’s biased or wrong in some way.
I don’t believe it should be legal, but I see why they would be butt-hurt
Well, Open AI had been whining about DeepSeek back in the day, so it is fair in a way.
Can't we say it's "fair use"? They do whatever they want saying it's "fair use", I don't see why I couldn't.
Exactly this. Strange that this comment got downwoted. AI companies are scrapping the entire internet disregarding copyright and pirating books. Without it, models will be useless.
Would something like that hold up in court?
they can choose who they do & don't want to do business with
But who's going to enforce this for them? And would they even find out if the service is otherwise available to the general public?
Law does not work like that.
- Contracts can have unenforceable terms that can be declared null and void by a court, any decision not to renew the contract in future would have no bearing on the current one.
- there are plenty of restrictions on when/ whether you can turn down business for example FRAND contracts or patents don’t allow you choose to not work with a competitor and so on.
People always say "this wouldn't hold up in court" and "the law doesn't work like that" when it comes to contract, but in reality, contracts can mostly contain whatever you want.
I see no reason why Anthropic can't arbitrarily ban OpenAI, regardless of my opinion on the decision. Anthropic hasn't "patented" access to the Claude API; there are no antitrust concerns that I can see; etc.
Nobody was asking if Anthropic can ban OpenAI. I believe they were asking if the contract that can ban using the output to train an AI would hold up in court.
And no, it isn't clear to me that this contract term would hold up in court as Anthopic doesn't have copyright ownership in the AI output. I don't believe you can enforce copyright related contracts without copyright ownership.
I could be wrong of course, but I find it odd this topic comes up from time to time but apparently nobody has a blog post by a lawyer or similar to share on this issue.
> And no, it isn't clear to me that this contract term would hold up in court as Anthopic doesn't have copyright ownership in the AI output
They don't need copyright ownership of the AI output to make part of the the conditions for using their software running on their servers (API access) an agreement not to use it training a competing AI model.
There would have to be a law prohibiting that term, either in general or for a company in the specific circumstances Anthropic is in. (The “specific circumstances” thing is seen, e.g., when a term is otherwise permitted but used but a firm that is also a monopoly in a relevant market as a way of either defending or leveraging that monopoly, and thus it becomes illegal in that specific case.)
Dumbest thing they could do, why would you cut off insight into what your competitors are doing?
Because they don't blatantly read people's prompts. They have a confidential inference architecture.
They don't target and analyze specific user or organizations - that would be fairly nefarious.
The only exception would be if there are flags for trust and safety. https://support.anthropic.com/en/articles/8325621-i-would-li...
Good luck with that! Most of the relevant model providers include similar terms (Grok, OpenAI, Anthropic, Mistal, basically everyone with the exception of some open model providers).
You're like 50% of the way there...
The article does not say anything substantial, but just some opposite viewpoints.
1/ Openai's technical staff were using Claude Code (API and not the max plans).
2/ Anthropic's spokesperson says API access for benchmarking and evals will be available to Openai.
3/ Openai said it's using the APIs for benchmarking.
I guess model benchmarking is fine, but tool benchmarking is not. Either openai were trying to see their product works better than Claude code (each with their own proprietary models) on certain benchmarks and that is something Anthropic revoked. How they caught it is far more remarkable. It's one thing to use sonnet 4 to solve a problem on Livebench, its slightly different to do it via the harness where Anthropic never published any results themselves. Not saying this is a right stance, but this seems to be the stance.
Feels like something a Jepsen or such should be doing instead of competitors trying to clock each other directly. I can see why they would feel uncomfortable about this situation.
>Nulty says that Anthropic will “continue to ensure OpenAI has API access for the purposes of benchmarking and safety evaluations as is standard practice across the industry.”
This is, ultimately, a great PR move by Anthropic. 'We are so good Open Ai uses us over their own'
They know full well OpenAI can just sign up again, but not under an official OpenAi account.
It's not easy for a multi-billon dollar company to hide and evade ban. If they are found doing so, they can be easily dragged in courts.
>OpenAI's own technical staff were also using our coding tools ahead of the launch of GPT-5,
This smells more like a cheap PR potshot. Using an API for training vs developers using it for coding are very different interpretations of "build a competing product"
I would expect every AI company to use other companies' models in "its own internal tools using special developer access (APIs)", at the they least for evals.
If anything, this is a bad look for Anthropic.
OpenAI Services Agreement: "Customer will not [...] use Output to develop artificial intelligence models that compete with OpenAI’s products and services"
Live by the sword, die by the sword.
Didn't a whole bunch of AI companies make the news that they refuse to respect X law in AI training. So far, X has been:
* copyright law
* trademark law
* defamation law (ChatGTP often reports wrong facts about specific people, products, companies, ... most seriously claiming someone was guilty of murder. Getting ChatGPT to say obviously wrong things about products is trivial)
* contract law (bypassing scraping restrictions they had agreed as a compay beforehand)
* harassment (ChatGPT made pictures depicting specific individuals doing ... well you can guess where this is going. Everything you can imagine. Women, of course. Minors. Politics. Company politics ...)
So far, they seem to have gotten away with everything.
Most of these things are extremely valuable for humanity and therefore I think it's a very good thing that we have had a light touch approach to it so far in the west.
e.g. ignoring that I find that openai, Google, and anthropic in particular do take harassment and defamation extremely seriously (it takes serious effort to get chatgpt to write a bob saget joke let alone something seriously illegal), if they were bound by "normal" law it would be a sepulchral dalliance with safety-ism that would probably kill the industry OR just enthrone (probably) Google and Microsoft as the winners forever.
> defamation law
Not sure if you're serious... you think OpenAI should be held responsible for everything their LLM ever said? You can't make a token generator unless the tokens generated always happen to represent factual sentences?
Given that they publish everything their AI says? That that's in fact the business model? (in other words, they publish everything their AI says for money) Quite frankly, yes.
If I told people you are a murderer, for money, I'd expect to be sued and I'd expect to be convicted.
But you aren't an LLM (that we know of I suppose)
I am a company though. Well, that's how I work. So why should it be different for me vs OpenAI or anyone else?
Nope - they do not charge you for truth
If you had a disclaimer (like OpenAI) does, probably different story
Disclaimers do not shield you from the majority of tort law. There's a reason South Park makes fun of such disclaimers.
> trademark law
Presumably an AI should know about the trademarks, they are part of the world too. There is no point shielding LLMs from trademarks in the wild. A text editor can also infringe trademarks, depending how you use it. AI is taking its direction from prompts, humans are driving it.
Does openai team use claude to vibe code ? Their models qre too stupid to code.
Claude's AI models aren't great for any kind of coding other than vibe coding either. :(
So it begins!
Let's hope.
Wow, that's why Copilot has been acting funny on Vscode. Code quality dropped and requests keep failing. But I'm still seeing Claude models in the selector. Can't read this article because of paywall. Are they exaggerating?
no, that's just because claude has been having persistent brown-outs for a while now. Microsoft and Anthropic messed up by bailing out of full-tilt AI hyperscaling a fair bit. (i.e., Microsoft cancelled a bunch of AI datacenter projects earlier this year)
Who will pay me for my AI chat histories?
Seriously, make a browser extension that people can turn on and off (no need to be dishonest here), and pay people to upload their AI chats, and possibly all the other content they view.
If Reddit wont let you scrape, pay people to automatically upload the Reddit comments they view normally.
If Claude cuts you off, pay people to automatically upload their Claude conversations.
Am I crazy, am I hastening dystopia?
There is an A16z company that does exactly this, called yupp.ai. They need genuine labelling/feedback in return, but you get to either spend credits on expensive APIs or cash out. Likewise, openrouter has free endpoints from some providers who will retain your sessions for training.
> Who will pay me for my AI chat histories?
All the chatbots with free access do that, they pay you by running your arbitrary computations on their servers.
Than I would simply use AI to generate chat histories and get paid (:
That is not a problem if the price paid is lower than what generating synthetic data of similar size will cost .
Great point. Verifying the synthetic data also has a cost, I wonder if it is cheaper than generating it?
They could probably pay you based on loss / some similar metric during training.
Two things. First, no one wants your AI chat histories. They want to interact with the LLM themselves. Second, their business models break down when they can't steal content to train on. Paying for training data on a large scale is out of the question.
I've done a lot of post training and data collection for post-training
I think if you're not OpenAI/Anthropic sized (in which case you can do better) you're not going to get much value out of it
It's hard to usefully post-train on wildly varied inputs, and post-training is all most people can afford.
There's too much noise to improve things unless you do a bunch of cleaning and filtering that's also somewhat expensive.
If you constrain the task (for example, use past generations from your own product) you get much further along though.
I've thought about building a Chrome plugin to do something useful for ChatGPT web users doing a task relevant to what my product does, then letting them opt into sharing their logs.
That's probably a bit more tenable for most users since they're getting value, and if your extension can do something like produce prompts for ChatGPT, you'll get data that actually overlaps with what you're doing.
This article says absolutely nothing and appears to be an ad for anthropic
Do you have adblocker on?