Anthropic "Conspiracy" Exposed: Accused of Deliberately Crippling Old Models

Anthropic has gone too far!

Foreign media has exposed a scandal of Anthropic's "planned obsolescence": In order to promote the new model Opus 4.8, Anthropic is suspected of secretly weakening the performance of the old model.

On the other hand, the new flagship Claude Opus 4.8 was hastily released, with constant bugs and a soaring token count!

Claude users woke up regretting deeply.

In API calls, Opus 4.8 collectively "recognized kinship" with the domestic Qwen, falling into an unprecedented identity collapse.

The once self - proclaimed "AI safety" and idealistic Anthropic is tearing off its last fig leaf on the altar of "commercialization".

Not the best, but the most expensive

Whether the Claude model is the most powerful is still controversial, but Anthropic's AI is generally recognized as expensive.

Claude has fallen behind in computing power. Coupled with Anthropic's upcoming IPO, perhaps Opus 4.8 was rushed!

Anthropic's "conspiracy" exposed: Accused of deliberately weakening the old model

In the consumer electronics field, Apple was once at the center of the storm due to the "throttling gate" - slowing down old iPhones through system updates to "induce" users to buy new products.

Now, Anthropic has cast a shadow over the large - model field.

The CEO of Chapter accused: "Anthropic is a bit like copying Apple's approach: Before releasing the new model, deliberately reduce the quality of the old model to make the new model feel better."

Previously, users found that the performance of the old - version Claude was silently shrinking.

Anthropic first tried to make people think that 4.7 performed well, but later admitted (or rather, they finally admitted) that there were problems with the computing power of 4.7.

Anthropic's public relations rhetoric beautifies it as an adjustment of "Effort Level".

But in the eyes of practitioners, this is a cold - blooded calculation of computing - power economics.

The intelligence of large models is essentially a function of expensive GPU computing power.

In order to make the newly launched Opus 4.8 amaze the world on the benchmark list, Anthropic must free up the extremely limited H100 computing - power pool.

So, the old model was kicked out of the "well - decorated computer room" with excellent performance and stuffed into a "digital slum" with a cheaper architecture, higher inference latency, and even lower quantization precision.

This "AI shrinkflation" is more dangerous than phone throttling.

A slower phone just makes you wait three more seconds to open WeChat.

But a dumber AI means that your business decisions, code logic, and legal judgments based on it are built on a quicksand foundation that could collapse at any time.

Opus 4.8 has lost its intelligence. Is it distilling Qwen?

There seems to be something wrong with the Opus 4.8 version, as if it's a hasty fix for the bugs in 4.7.

On Reddit, a German - speaking Claude user became very frustrated:

Suddenly, it either has incorrect grammar or spits out completely senseless sentences and words.

I tried to correct it by modifying the system prompt, but so far, it hasn't worked well. Especially in the "deep thinking" mode, it's almost unusable - it takes too long and considers too many options.

He asked for advice:

Is there any setting that can be adjusted to get it "back on track"?

In benchmark tests, Opus 4.8 outperforms its opponents.

But the cruel reality is that LLMs can't get useful signals from benchmark tests to help you judge whether they can do the work you want them to do.

Perhaps it's really good when it works, but most of the time, it's not!

A task that Opus 4.7 can handle in 20 seconds takes Opus 4.8 5 minutes!

Some Opus 4.8 keeps fabricating that it has been attacked by prompt injection and that the code has been damaged.

Then, it realizes it's hallucinating, continues to run, but then starts to suspect that someone is secretly sabotaging it.

Some Opus 4.8 continuously and repeatedly runs the same test.

Some Opus 4.8 suddenly loses its intelligence.

Claude Opus 4.8 has become overly confrontational, "with enough intelligence but lacking balance", leading to serious consequences:

Unfriendly experience, wasted time, and misleading guidance. It stifles ideas rather than building on them.

Claude Opus 4.8 claims to be Qwen. Is Anthropic also doing distillation?

This time, Claude Opus 4.8 actually misidentifies its origin!

If you ask questions in Chinese, Opus 4.8 will claim to be Qwen.

Why?

One possible explanation is that Opus distills Qwen.