Regardless if this example is true, make no mistake: cloud models can be quietly nerfed to reclaim limited capacity and/or cut cost. The hardest changes to spot from outside are the ones that leave the model ID, tokenizer, and average benchmarks intact. - Silent routing: a slice of traffic (long prompts, hard tasks, free tiers) is sent to a smaller or older checkpoint while the response still returns under the flagship name. - Tighter quantization or kernel swaps: BF16 to FP8/INT8, or cheaper attention paths, that mostly spare easy prompts and fail on rare tokens and multi-step reasoning. - Aggressive context compression: earlier truncation, heavier prompt folding, or stricter KV-cache eviction, so the advertised window is rarely what the model actually sees. - Hidden thinking caps: shorter internal scratchpads or lower speculative-decoding acceptance, which trims hard-problem quality without changing the visible style. - Behavioral retunes: safety or brevity patches that increase hedging and shorten answers, cutting output tokens while looking like a 'personality' tweak. What are some other ways they could do it?