> Point being my real API usage here must be very unsustainable!

I think it’s funny that everyone anchors to the API pricing as the real cost.

Most likely is that their API costs are printing profits. They can sell the subscription plans at a slight loss because it gets more people like you hooked on GPT models at home and suggesting them at work, where the real money is made.

I think their subscription plans go mostly unused when averaged across all subscribers, too. Some customers are getting great deals by maxing out 100% every week, but most probably use much less.

The reset game is an addictive challenge that gets the hardcore users more hooked on their products because you feel pressured to use it as much as you can before the next unpredictable surprise reset lands.

▲

slopinthebag 17 hours ago | parent [-]

Idk, Kimi K3 is not that much cheaper than Opus and it’s probably a smaller model. We’ll know for certain when commodity providers price it, but I expect it to be the same price.

▲

manquer 17 hours ago | parent | next [-]

It really depends on caching[1] and active parameters, and some dark arts really only possible at Hyperscalers they also have better access to 288GB VRAM B300s[2] that really makes a difference in cache space.

K3 may be bit smaller/ similar in total parameter count than Opus, but Opus (and GPT) definitely has become a lot more efficient in the last 6 months or so, hence more or less forced upgrades. I expect the active parameter count and cache performance is quite different.

Opus is still almost twice as expensive (if tokens were equal) at $5/$15 compared to K3 at $3/$15. Tokens are not equal though, Anthropic's tokenizer is much less dense than other frontier lab's so the actual price difference is like 3x.

Kimi uses more reasoning tokens and is generally more inefficient with its usage, that doesn't impact token cost economics for the provider though . It does for us as buyers thus the need to evaluating by Cost per task rather than unit pricing.

On pure tokens/$ - there is definitely room for a price war if operators start going by pure unit costs. Both probably want(ed) to have good enough numbers in preparing for the IPO.

---

[1] the pricing kind of reflect this already - 10x diff for uncached input.

[2] Moonshot does not have access latest gen GPUs so their unit economics is likely hampered a bit for high parameter models.

▲

Palmik 13 hours ago | parent | prev [-]

Commodity providers aren't a good indicator. They have margins too.

Remember they ~doubled the price going from GLM 5 to GLM 5.2, despite same [1] cost of inference.

[1] GLM 5.2 is actually slightly more efficient, thanks to baked in indexer cache.

	▲	slopinthebag 13 hours ago \| parent [-]
		These models are more fungible than you think, so these prices tend towards the cost to serve. On openrouter GLM 5.2 is ~3x cheaper than 5.0.