The Anthropic Pricing Paradox

Remember when everyone was mad about Claude Opus pricing? The small usage limits on Pro, the $5 per million tokens on the API, the general feeling that Anthropic was charging too much. People complained about it for months.

Then Fable 5 came out. $10 per million input, $50 per million output. Double the price. And the complaints just vanished.

I am not exaggerating. People who spent weeks saying Opus was too expensive suddenly had nothing to say about a model that cost twice as much. Because Fable was really good. Good enough that the price stopped mattering. Good enough that everyone forgot they were mad about the exact same thing a few weeks earlier.

Here is what I think. More than half the people using Fable did not have a project that actually needed it. They just wanted to use the smartest model.

Opus 4.8 was $5 in and $25 out. People hated the price. Fable 5 is $10 in and $50 out. Same people loved it. Opus 5 is $5 in and $25 out. Same as Opus 4.8. Now those same people say it is a bargain.

The number did not change. $5 and $25 was unacceptable in May. Now it is a steal in July. The only difference is that Fable sits above it and makes Opus 5 look reasonable. That is a textbook pricing anchor, and Anthropic pulled it off perfectly.

Scroll through LinkedIn or Instagram right now. Every other post is about how Opus 5 is incredible value. “Almost as smart as Fable at half the price.” “Best deal on the market.” People actually believe they are getting a discount.

Here is the thing. Kimi K3 from Moonshot costs $3 in and $15 out. On the independent Artificial Analysis Index it scores 57.1. Fable scores 59.9. Opus 5 is somewhere in between. K3 has 2.8 trillion parameters, open weights, a million token context, and costs about a third of what Fable charges. But you are not seeing LinkedIn posts calling Kimi K3 a great deal, because these people are not in that world. They learned Claude. They built their workflows on it. The loyalty stopped being about value a long time ago. It is branding now.

And most of the people celebrating Opus 5 and Fable 5 are building pretty basic stuff. Todo apps. SaaS wrappers. Landing pages with an AI chat feature. The kind of projects GPT-4 could handle two years ago. But they want the frontier model because using the frontier model is the point. It makes them feel like they are on the edge.

I have been running DeepSeek V4 Pro and V4 Flash in agent loops and the gap is smaller than the pricing suggests. The scaffolding does a lot of the work. Context management, error recovery, tool use patterns, those matter more than people admit. But that is a less exciting story than “I used the most powerful model ever created,” so nobody wants to talk about it.

Opus 5 is a good model. I am not saying it is bad. The whole thing is about framing. Anthropic put Fable 5 above it at double the price, and suddenly the old pricing felt new again. It worked on basically everyone.

Kimi K3 is open weight. You can download it, run it yourself, or pay $3/$15 on the API. It matches Fable on most independent benchmarks. That is the actual value story nobody is telling.

Use whatever works. But maybe think about whether you need Fable 5 for your CRUD app, and whether Opus 5 is actually cheap or just cheap compared to the most expensive option. The numbers did not move. Just the reference point.