Claude 4.7 Costs 50% More—But Reddit Says 'Give Us Back 4.6'
After Opus 4.7 dropped, the liveliest place wasn’t Anthropic’s website—it was Reddit.
The titles wrote themselves: “Give us back 4.6,” “4.7 took my money and gave me regression,” “50% price increase for worse performance—what’s Anthropic doing?”
Honestly, the complaints have data behind them.
Multiple developer tests show Opus 4.7 doesn’t significantly outperform 4.6 on complex tasks—and in some scenarios actually regresses. Price went up 50%, experience didn’t improve proportionally. The math doesn’t work.
I saw someone post on Maimai (a Chinese professional network) a while back: “The worst thing about subscription services isn’t price increases—it’s when users feel they’re not getting value.”
That quote perfectly describes Anthropic’s current predicament.
What’s Anthropic’s Pricing Logic?
To answer this, we need to understand how AI companies actually set prices.
Mainstream AI model pricing is token-based. Anthropic’s Opus line targets high-end users, priced significantly above Sonnet and Haiku. The 4.6 to 4.7 upgrade reportedly increased inference costs substantially—not training costs, inference costs.
What does that mean?
Every time a model “thinks” and “responds,” it consumes compute. Higher inference costs mean higher prices.
But here’s the gap: users don’t feel the direct connection between “your GPU bills” and “what I’m paying.” They care about one thing: can this model solve my problem? Fast enough? Accurately enough?
On both those dimensions, 4.7 didn’t deliver for everyone.
“Model Degradation” Is the Real Issue
Let me be direct.
When users say a model “got dumber,” that’s not a pricing problem. It’s something deeper.
Claude 4.6 got criticized not for being insufficiently capable, but for getting “well-behaved.” The boldness, the willingness to take risks, to give aggressive solutions—gone. Replaced by conservatism, safe answers, playing it safe.
This isn’t capability regression. It’s behavior adjustment—maybe safety concerns, compliance requirements, or just side effects from the alignment process.
Users aren’t stupid. They feel it.
So when 4.7 launched, everyone expected: “Bring back the early 4.6 intelligence.”
The result?
4.7 improved, but not enough. Definitely not “worth the price increase” enough.
My Take
I think Anthropic is stuck in an “impossible triangle”:
Capable version → Users complain it’s too “wild”/unsafe → Adjust alignment → Becomes conservative → Users say “it got dumb” → New version fixes it → Capable version → …
Around and around.
Finding balance between “safety,” “capability,” and “cost” might be AI companies’ eternal challenge.
As for whether 4.7 is worth buying—my advice: wait a week or two for comprehensive benchmarks. If there’s clear improvement, jump in. If not, 4.6’s price is more attractive.
The “wait and see” crowd never loses.
Reflection Question
Should AI model pricing be tied to “user-perceived value” or “actual compute costs”? How do you balance these two?