DeepSeek is somewhere between Sonnet and Opus in capability, for much lower price.
Their $0.87 per million output tokens is one of the few API offerings that is probably subsidized (the break-even price is probably around $2 judging by[1])
I think Anthropic and OpenAI's margins will erode some over time. But I think they are very profitable now on the API prices they are charging, and their margins will remain healthy.
Put it like this: AWS is a very expensive way of getting compute, and there are numerous competitors that are much cheaper.And yet AWS is very profitable.
[1] https://openrouter.ai/deepseek/deepseek-v4-pro#providers
I use it as something that sits right below Sonnet for daily tasks that can be run in loops with validation checks, perfect. Anything more advanced...it just doesn't cut it.
Same experience with z-ai glm 5.1 btw... don't know what system prompt magic antrophic set in front of their Sonnet 4.6, but it does execute tasks efficiently.
claiming any of these deepseek models are close to opus is a bit optimistic
For software development, I notice no difference. For all the work I did using AI, DeepSeek is as capable as claude.
I tried DeepSeek and MiMo through both Claude Code and Pi.dev. I found Claude Code a slightly better harness, but I still have both installed and switch around frequently.
As for "open source AI", what exactly are you talking about? Is it open models?
I tried some through Ollama, but I don't have the hardware to run decent models.
Anyway I'm rambling, what Microsoft is doing with AI in enterprise is basically what they did with Teams and similar systems. They provide a platform for it which is good enough that your organisation is going to want it rather than deal with multiple vendors. Not for tech organisations, but for every other enterprise organisation it'll be so much easier to just go this way. I imagine that Anthropic is getting some sort of payment from Microsoft for Cowork, but what Cowork shows is that Microsoft can be completely model agnostic and still sell "top" AI. Especially because they've set cost on a fixed rate that I'm sure they'll increase by 25% every year.
Or do things like the fact that you need some sort of special Agent 365 license for your sysadmins to manage the admin.microsoft part of Copilot which has to do with security policies... Ask me if it was fun doing that agent by agent... It's frankly the most Microsoft thing I've ever seen.
The cost of hardware still needs to dramatically drop for open-weight models to be viable for local usage. Even with the release of things like Nvidia DGX Spark and Ryzen AI Halo, you'd likely want a few of them to run agents in parallel.
Sure, you can use cloud hosted variants of models like DeepSeek etc at API rates, but subscriptions still come out on top for bulk usage. GPT is already tightly integrated into peoples workflows, has wide adoption, has good tooling for developers, etc.
Plus there's nothing stopping them from competing on a price level if they really feel the need. It just means they might burn more cash in the short term.
It's more efficient to do the opposite on a constrained platform. Run agents in parallel using a single model, then round-robin among models for cross-checking purposes. (The makers of local inference engines are dropping the ball by not making batched inference a first-class citizen of that workflow. It's not just useful for vLLM and SGlang.)
But a lot of other people are going to happily shell out for Opus 4.8 at way higher prices.
The API GLM 5.2 launched last night with several providers on OpenRouter. I had a short conversation, didn't get to test much, but initial impressions "the vibes were good".
(For tax purposes for most businesses it doesn't matter whether an expense is CORS or not.)
There are arguments for and against this - awareness is an issue for startups, but most large companies continue to market. It is true that it is fairly easy to change the amount a company spends on marketing.
Either way, as the parent says - provided the investors understand it there is nothing weird about doing it this way.
(To explain: It's basically saying they are spending $X now to get them as a customer which will yield a lifetime value which is some multiple of $X. That's a very valid thing to do - see the whole field of cohort analysis for SAAS)
If the marketing and promotions and discounts stopped tomorrow, that revenue will drop.
Cost of revenues Research and development Sales and marketing General and administrative
Take it up with the SEC I guess?
[1] https://s206.q4cdn.com/479360582/files/doc_financials/2026/q...
But for all practical intents and purposes, you can't just fire all your sales people and shut off your marketing firehose, and stop giving discounts, and expect revenue to remain steady in either the short, or the long-term.
You can shut down your R&D, and it will remain steady, at least in the short term.
I am clearly mincing terms that mean specific things, and I'm very sorry to do so. I am simply trying to communicate that generally speaking, OpenAI is currently spending ~$1.05 to make a dollar, and is also on top of that, spending ~$3.00 in R&D.
Now, I do think that it's likely that they can increase their prices to close that 5-cent gap.
On top of that, oil wells decline, slowly but surely, just like... customers churn.
If you spend $100 mill on sales and marketing and get zero customers, I'd agree it should expensed. If you get a bunch of customers, it's hard to argue against the capex treatment idea.
They were first on a few things but the tell is the coherence of the thinking traces of Claude. You have to put a loss on that. GPT 5 series thinking traces are creepy, Gemini thinking traces are disturbing. They both represent forced discontinuities on the policy gradient.
Claude is good at tool use because it's gigantic and well-labeled, but the reason you pay the premium is for a thinking partner, not a tool user.
Claude Code is the cancer that will kill the patient, Boris is the the Kardashian version of Karpathy, with less business sense.
I am not sure I understand.
The API business throws a massive model that by definition can't be inferred efficiently because nothing can across 4 different compute substates, at a problem that DSv4 nails at or near 100% while leaving most of the actual unique value of Claude on the table.
Claude should be in your house and car and your kid's classroom and shit.
Having it write tail -n5?
That's because Anthropic's A-Team is Meta's C-Team. Hell, I fired some of their stars myself.
Humans.
The legibility of thinking traces has gone down, but mostly this nets out to spilling more XML because their website is a vibecoded heap
That's not necessarily true.
It depends on two things that are impossible to derive from the leaked numbers:
1 - How much of their compute costs are subsidized.
2 - How much of that R&D chunk can actually be reduced for the company to continue working. What goes into "R&D"? Is that only things such as training for new models? This is impossible to determine from those numbers alone.
If those numbers are true, they would not be "close to profitability". They would be profitable, period.
This does not explain why they need to raise money like crazy. It would be possible to train new models at a more sustainable pace without the need to jack up prices at all, only with the difference in between "Revenue minus Cost of Revenue".
They also wouldn't need to be on a mad race to IPO and dump this into the public market.
To be frank, the more I look at those numbers, the more I think this is completely unsustainable.
It feels like cost of revenue should account for training, but I'm not an accountant so who knows?
Obviously you need new training rounds over time. Knowledge is not static. New things that are created would need to be part of later models.
How do you account for that? Do you account for some sort of model depreciation?
A lot of things are very nebulous in those leaked numbers.
How is "cost of revenue" considered? If Microsoft, Oracle and such provides computing at a loss to Open AI, that is clearly not sustainable, and it is a way to pretend the numbers are better than they actually are prior to an IPO. This might even be a source of pressure for an IPO. As losses accumulate in the ecosystem, they need to dump it into the public so they can bring the numbers to reality.
Also, I am not sure if they own any compute (for example, the Stargate datacenters). If they do, do they lump the costs of building those data centers all in "R&D"? That would be one hell of a way to pretend that inference is cheaper, good luck having inference without those datacenters.
The more I look at this, the more it looks like "OpenAI is profitable when you pretend they have no costs".
> How much of that R&D chunk can actually be reduced for the company to continue working.
Worth noting that almost all engineering (including software engineering) is always included in R&D.
Development is important! Most software (and indeed engineering!) companies couldn't continue if they stopped R&D.
Ford couldn't continue as a company without R&D.
Yes. But is all R&D model training? That was the actual question.
Everybody that is presuming "inference is profitable" here is just naively deducting "cost of revenue" from "revenue" and running with that.
We know OpenAI is forecasting $25-30B revenue for 2026. They will be very close to breaking even at those number.
Given Anthropic has forecast more revenue than OpenAI and we know has spent less on R&D (cite their desperate scramble for compute capacity!) the rumours of them being profitable this year seem very credible.