Model prices collapse as open weights go on-prem

OpenAI and Anthropic cut LLM prices by up to 80% within a few months, and the top comment predicts token vendors will end up like mobile data providers: more throughput, more efficiency, thin margins. One commenter said their team tried Kimi and GLM after hitting 'cybersecurity' blocks on the big two and found them much cheaper. IBM's CEO then said open-weight models are now good and cheap enough to run on-prem. Commenters immediately noted IBM sells the hardware, and asked everyone to stop giving weight to CEOs talking their book.

The 'Has anyone replaced Claude/GPT with a local model' thread shows how close this is. One user says Qwen 3.6 27B dense is about equal to Haiku 4.5. Another runs DeepSeek V4 Flash on two RTX Pro 6000 Blackwells at 160 tok/s. A third tried Gemma 4 on an M4 and found tok/s far below the cloud. The gap is shrinking but is not gone. Separately, the thread on Anthropic ending Claude Code subscription use with OpenClaw shows the labs tightening access even as prices fall.

The counterpoint: 'good enough for the old world' is not the same as good enough for what's coming. Frontier labs may be cutting prices by shipping cheaper models behind the same names, which is one commenter's explanation.