If you're still on one of the older Claude enterprise contracts, just wait.
The old deals we see were around $200 a seat and that got you access plus usage behind rate limits (not even close to what that usage would cost at API rates). New ones are $20 for access only and every token gets billed at full API rates.
The $180 you save on the seat doesn't come back as usage. Expect your renewal to come in at multiples of what you're paying now.
The cost of delivering this stuff is brutal and it's still being heavily subsidized in many cases.
A lot of our customers have contracts expiring soon. If yours is coming up how are you handling it?
I’m actually surprised-about-the-surprise around AI subs continuously having their quotas lowered. The frontier providers are running the subs at a massive subsidized loss. What should be more surprising to everyone is that they’re not completely gone yet because that’s the trajectory. Enterprise agreements already have zero usage per seat included.
Historically subscription-based products balanced out via covariant user usage. That’s not the case at all with the OpenAI or Anthropic subs. A few companies are trying to build financially sustainable subs but you’ll notice their limits are even lower (they have to be).
Enjoy the party while it lasts but it’s a certainty that it ends soon.
We're in San Francisco at The AI Conference this week. If you're here, let's talk. We'd love to compare notes on what we're seeing across the industry and hearing from customers about how their AI consumption is changing.
Same themes keep coming up. Unlocking model choice, actually knowing what everyone in the company is running, and spending less on inference at the end of the day. If you're working through the same questions, happy to share what we've learned and help where we can.
The most useful conversations we're having with enterprise customers aren't about model access. They're about what it takes to put AI to work across a business. Cost attribution. Identity. Security. Visibility. New models without new work.
We're @AIConference. Demos at BOOTH 131.
We’ve been seeing exactly what we anticipated play out over the past year. Frontier model providers are doing everything they can to create lock in, preparing for the complete end to all subsidized inference. The last thing they want is to become commoditized, which is the opposite of what companies around the world need to have occur. It’s incredibly important to have a strategy which appropriately leverages as many providers and models as possible, responsibility, effectively and efficiently. Leaning on the providers themselves for solutions is playing right into their hands.
https://t.co/e7uFF1qXeV
Well, that was a long time to keep quiet.
Actualyze is out of stealth as of this morning. A year of building, a $7M seed, and a lot of conversations where I had to bite my tongue about what we were actually working on.
I wrote up the full story on our blog: what a year with enterprise platform and security teams taught us, and what we built because of it.
https://t.co/5gF2uiCbbZ
There's a lot more I want to say about this space, and now I finally can. More coming.
If you're at Ai4 in Vegas this week, come find us at booth 313.
It’s time to demystify Mythos.
Mythos is not magic. It’s not a doomsday device. It’s the first of many models that can automate cyber tasks (just like coding).
OpenAI’s GPT-5.5-cyber can now do the same. And all the frontier models (including those from China) will be there within approximately 6 months.
It’s important to recognize that these models do not create vulnerabilities; they discover them. The bugs are already in the code. Using AI to discover and patch them will actually harden these systems.
The leap from pre-AI cyber to post-AI cyber means that there will be a big upgrade cycle. After that, however, the market is likely to reach a new equilibrium between AI-powered cyber-offense and AI-powered cyber-defense.
Obviously it’s important that cyber defenders get access before cyber attackers. That process is already underway but needs to happen quickly (see point above about Chinese models).
Unlike Mythos, GPT-5.5-cyber appears not to be token constrained so it may be the first cyber model that defenders actually get to use.
@webdevcody It’s much more literal about following instructions. I think a lot of the initial complaints are because fork lifting existing agent instructions and steering will not produce a direct “upgrade”. The model actually does what you ask it to now, without excessive convincing.
@thdxr This is how I run Opencode today. It’s a better model (imo) given the architecture you already have, rather than defaulting to a 1:1 server:client process.
@thdxr We’re still in the infancy for AI adoption. Even across software engineering, let alone literally everything else bound to be changed or dramatically disrupted with AI.
This is easily the most transformative technological shift since the Industrial Revolution. It’s not a bubble
Bill Gates is setting up factories to manufacture 7 leading vaccine candidates before we know which is best & safest; we can test the vaccines in parallel, and then throw away all but the factory for the best vaccine. May save many months.
Just extraordinary.
I’m seeing so many articles that provide no value aside from invoking fear. Stay informed but steer clear of any doomsday news. Anything providing actionable info is valuable. Otherwise it’s information that contributes to fear for factors out of ones control. #coronavirus
Open-label pilot trial found that hydroxychloroquine (antimalarial drug also used for lupus) has shown great promise in the treatment of COVID-19. Addition of the antibiotic azithromycin w/hydroxychloroquine was even better. More RCTs to confirm underway. https://t.co/5qfdITkc4i
@CBSThisMorning@eyepv6@drtaranarula Be precise. 20% of those who were sick enough to seek care AND qualified for testing. Approximately 80% of people who contract this have minor symptoms or are asymptomatic. Phrases like “in this country” imply we know of every case when the opposite is true; we know of a fraction
The statistics being reported around #covid-19 are incredibly misleading, as they’re based on confirmed cases. A case can only be confirmed with testing and testing in the US remains highly restricted. Point is we have accurate numerators with entirely unreliable denominators.
https://t.co/FOQSUXONKN
Air traffic control recording of Kobe Bryant crash.
Start from the beginning, Kobe Bryant's helicopter is cleared for "special VFR" (almost always a bad idea, just saying...). He reads back the clearance at around 2:20 in the audio.
@kelseyhightower DevOps is a set of successful, not-new practices given a marketing name the industry could embrace and repeat. Fundamentally it’s what true “senior system administrators” were doing 15 years ago; sensibly automating every layer of lifecycle management, considering day-two.