[teravor]

    > Because we own the infra, the per-token price reflects the cost of running the model, not the cost of renting someone else's plus their margin.

is not compatible with

    > Pricing on Telnyx: $2.70/1M input tokens, $13.50/1M output tokens, $0.27/1M cached input tokens.

since you asserted something false and bizarre, how about telling us what is your markup?

[alexeldeib]

Why are those incompatible? Pricing is 10% under moonshot, and pure infra providers also want money

[gruez]

I don't get it, what's the contradiction supposed to be?

[rubslopes]

They are selling it even cheaper than Moonshot AI. Why are you so sure they are lying?

[teravor]

    We estimate that the true blended price per million tokens for running Opus 4.7 on agentic tasks at $0.99 despite the sticker price being $5/$25 per MTok.

https://newsletter.semianalysis.com/p/ai-value-capture-the-s...

according to Semianalysis those prices would be far above actual costs.

[petu]

They're not saying anything about Anthropic serving costs in that quote, just calculating what MTok price is for running agents. Next sentence after your quote:

> Agentic workloads have extremely high input-to-output ratios (our Claude Code usage has a ratio of about 300:1) and high cache hit rates (90%+). Because cached input tokens only cost $0.50/MTok, most of the tokens end up in the cheapest tier.

90% cache hit input blend: 0.9 * $0.5 + 0.1 * $5 = $0.95 per MTok.

300:1 input/output blend: (300 * $0.95 + $25) / 301 = $1.03 per MTok.

They don't say exact cache hit rate they calculated for ("90%+"), so close enough.

[Grimblewald]

Please correct me if I'm incorrect, but it seems to me those numbers are describing a fairly different situation to this one. Anthropic serving their own model to their own users at that scale gets cache hit rates and machine utilisation that someone standing up another company's 2.8T model in four regions isn't going to get, and this thing needs 64 accelerators minimum before it will run at all, so a rack sitting mostly idle through a quiet hour costs the same as a busy one. The margin figure quoted is also just the price against the cost of producing the tokens, it doesn't have buying the hardware in it, or depreciation, or maintenance, or the money they'd have made renting those machines out instead, which with rental prices up 40% since October isn't nothing.

I largely agree things are overpriced, I just don't think that article is the right basis for saying it about this one.

[tpm]

If they are indeed far above actual costs, then surely price discovery will be done by the overall market in due course.

[FergusArgyll]

**reflects** the cost...

Not the **literal** cost