Claude 4.6 Accused of 'Getting Dumber': Anthropic's Compression Strategy Backfires

I’ll be honest—when I first saw the ‘Claude 4.6 is getting dumber’ discussion, I thought it was just another meme floating around the AI community. Then I clicked in and… wow. Reddit and Twitter are absolutely on fire.

Here’s what happened. Starting around last week, developers began reporting that Claude 4.6’s output quality had noticeably declined. Not the occasional glitch kind of decline, but systematic and perceptible. Someone posted comparison screenshots: the same code review prompt that caught three potential bugs two weeks ago now only catches one, with Claude adding ‘overall looks good.’

Even more telling, someone ran controlled tests. Same prompt, ten consecutive runs—Claude 4.6’s responses are now 30% shorter on average compared to a month ago, with significantly more ‘correct but useless’ filler content.

Naturally, Anthropic couldn’t stay silent. Their official account responded: ‘We have not made any changes that would reduce Claude 4.6’s quality.’ But here’s the thing—users have the data right in front of them. If you didn’t change anything, why is the quality dropping?

Then the technical experts weighed in. The leading theory points to one phrase: reasoning token compression.

Here’s the deal—models like Claude ‘think’ before responding, generating massive amounts of internal reasoning tokens. Users never see these, but they’re computationally expensive. At peak usage, Claude 4.6 reportedly generates 5-10x more reasoning tokens than output tokens.

So the question becomes: Is Anthropic quietly compressing these reasoning tokens to save money?

Technically, compressing reasoning tokens is totally feasible. You can set limits or have the model stop earlier. But the trade-off is obvious—the model’s reasoning depth suffers, making it more likely to ‘phone it in’ on complex problems.

I dug through the community tests and found some patterns in this ‘quality decline’:

First, simple prompts aren’t affected much. If you’re just asking it to write a Python script, Claude 4.6 is still Claude 4.6.

Second, complex reasoning tasks took the biggest hit. Code reviews, mathematical proofs, multi-step planning—these were Claude’s strengths, and now you can definitely feel it ‘going through the motions.’

Third, long conversations trigger it more easily. Users report that after about ten back-and-forths, Claude’s response quality drops off a cliff, like it’s saying ‘can we wrap this up?’

This reminds me of something a friend once told me: ‘For LLM companies, operational costs and user experience are always a tightrope walk.’

Anthropic isn’t OpenAI. They don’t have Microsoft bankrolling them, nor ChatGPT’s terrifying consumer revenue. Every token costs them real money. From that perspective, compressing reasoning tokens isn’t a ‘whether to’ question—it’s a ‘how much’ question.

But here’s the kicker—you should probably tell users before you do it, right?

That’s what’s really bothering developers. If Anthropic had written in their changelog, ‘We optimized inference efficiency, which may affect some complex tasks,’ everyone would at least know what to expect. This ‘we changed something but won’t admit it’ approach feels uncomfortably similar to certain smartphone manufacturers’ ‘cloud-controlled throttling.’

Personally, I see a few possibilities here:

One: Anthropic is running A/B tests, with some users getting a ‘lite’ version of Claude, but the official announcement isn’t ready yet.

Two: Cost pressure is so intense they’re making dynamic adjustments behind the scenes—automatically compressing reasoning tokens based on server load.

Three (the conspiracy theory): Anthropic is prepping something big. Claude 5 is coming, and they’re intentionally making 4.6 slightly worse to create upgrade incentive.

Whatever the truth, one thing is certain: user trust in AI models is becoming increasingly fragile.

We used to assume models ‘only get better over time.’ Now we’ve discovered they can ‘secretly get worse.’ That psychological whiplash is more damaging than the technical issue itself.

If Anthropic wants to maintain its ‘technological idealism’ brand, they need to figure this out—saving money is fine, but getting caught saving money? That’s a whole different story.