[user3939382]

Subscription is ~$200/month. When I tried running on tokens it was ~$500/day for the same usage pattern.

My guess is my usage is atypically high, but even if the average equivalent is $100-250/day you can get a sense of how much Anthropic is subsidizing subscriptions as a loss leader to lock market share. IMHO this is a doomed strategy since open models will get into a long tail of parity and they’ll ultimately essentially be in the business of commodity compute.

[theptip]

I think you need to consider how much dev mindshare you get from the loss-leader. Claude has grown at an insane rate. They brought vibe-coding mainstream.

At some point they might decide they have enough demand and inertia from enterprise to reduce the subsidy. But to say “it’s doomed” really misses the fact that it has already been immensely successful.

[visarga]

We can switch in a heartbeat to a competitor, what kept us to Claude was that it had better models for a while. Closing Fable off means they are squarely inferior now.

For me Opus 4.8 was a slow model with a strong habit of talking down to me in an obnoxious way that would not be possible to prompt away. GPT 5.5 is now my main driver for serious work.

[miki123211]

In my experience, all Anthropic models tend to produce hard-to-understand word soup. They're like the kind of person who likes to use fancy words just to show you how much smarter they are.

That doesn't mean those models are stupid or just pretending. I do think that Fable is the best we can get right now. Nevertheless, I prefer OpenAI models for easier tasks where the primary output is text and explanations, not code.

[AbstractH24]

I’m still rattled by all the DoD stuff in the spring. At which point I deleted my ChatGPT/OpenAI account.

But Anthropic’s games with Fable and false humility is getting a bit old. And increasingly it seems like Agent Orange is gonna implode rather than team up with someone like Sam Altman to form America’s Third Reich. Particularly post-midterms.

Nonetheless, wish there was a third option on par with them two. Maybe there is I need to investigate.

[theptip]

Agreed that switching costs to Codex are quite low (not zero due to harness, prompt, and skill differences), but only because OpenAI is running the same loss-leader strategy.

[yehosef]

I use them all https://github.com/yehosef/llm-tools

[verall]

I think it's for the insane amount of data they get to collect. Most people probably don't turn off the "help improve other models" while corporations paying by the token have blanket no-train policies applied to everyone.

[theptip]

Yeah that is definitely an angle too.

[KronisLV]

Using Opus 4.8 on their Max 5x plan (108 EUR with VAT) would have cost me around 4600 EUR at API rates: https://news.ycombinator.com/item?id=48684539

I also compared them against GLM Coding Plan (cause everyone was positive of the model): https://blog.kronis.dev/blog/z-ai-s-glm-5-2-is-a-great-model... and Anthropic still gives you a bit more tokens despite having some of the presumably most expensive to train and run models.

[andxor]

What makes you think the API costs are "the true price" and anywhere near their inference costs? Also what percentage of users do you think actually maxes out their subscription?

I suspect they're subsidizing a lot less than people think.

[rootatixww3]

there are many sources saying that Anthropic has 80% margin (profit) on API tokens

[alwillis]

> I suspect they're subsidizing a lot less than people think.

Agreed.

[edg5000]

We can look at openrouter GLM 5.2 prices I think to get a rough idea about the pure inference cost (with margin). They are probably running on simmilar hardware. Although scale helps Anthropic probably.

[solumunus]

Well they’ve claimed operational profitability on that, and probably aren’t lying as they’re about to IPO…

[miki123211]

Anthropic and OpenAI are really two different businesses in one company.

One business is serving great models via the API. I think that one is indeed going to become commoditized. The other is making consumer / developer focused products like Chat GPT, Claude Code or Codex.

Anthropic is leaning hard into making Claude Code work with things like enterprise compliance policies. I wouldn't be surprised if at least one of these companies basically ends up as a SaaS fronting many different models for different capability points, some of which are open.

[oezi]

I track the subscription value against API rates and you get between 13-20k at current rates. When Fable launched I got 32k for a short time. 500 USD per day is thus very little.