The inertia is legal and financial. People are paying Anthropic through AWS accounts because the simple reason of not dealing making new contract and legal agreements is enough of reason of the inertia.
But, eventually, I’m quite sure that AWS will also provide open models with those contracts without any inertia. Copilot is already offering Kimi.
My company has a deal with Devin and they provide new models all the time, and open models are becoming the most used ones by our internal metrics, especially because the company is very worried about cost.
Say you have 20% of usecases that require the more expensive model — but in 80% you could just use Llama instead of Sonnet (eg, for basic queries of a document). That saves 80% of that 80%, or 65% of your total bill!
That is the kind of “swap” that’s likely to occur in automated tooling as pricing pressure kicks in — “can you save 65% on our AI bill by switching Bedrock over in 80% of uses?”
Bedrock is really out of date with the models it offers, to the extent that I'm not sure they even have plans to update what's on there now they have the deal with Anthropic. They're still offering Qwen 3, not even 3.5 and certainly not 3.6. GLM 5 is the newest z.AI model they have, when it's 5.2 that would be the one to worry Sonnet.
There are some ok models on there (Qwen 3 Coder Next is usable and fast, for instance) but the lack of updates in a fast-moving field makes it something I don't want to recommend to my org.
Maybe for a fantasy of legal liability of output produced. I haven't heard of any LLM corpo being held liable for any output they generate. Even NYT lawsuit is going nowhere for 3 years in courts already, despite being the most grounded.
Copilot insures your company up to x if someone sues our company. I guess most companies do this. That’s why we can’t use anything else except copilot.
Watch out. Those claims are often made based on KL-divergence over some arbitrary corpus, not performance in the real world or benchmarks.
I’ve found that I need to go a couple steps past whatever quantizations are good enough in the KL-divergence testing to get good performance in real tasks with long context. So when Q4 is claimed to be lossless I end up with Q5 or Q6 for actual long-context tasks.
