Anthropic's conversation share fell 22 points while its downloads rose 58%

A four-month slide

The Anthropic family's share of AI-model conversation on Hacker News and Lobsters peaked at 70.1% in March. By August, a partial month at 22 days, it stood at 48.5%. The fall ran from April through August on the all-tagged basis. On the AI-model-restricted basis, where only mentions of named model families are counted, the drop went from 68.7% in April to 48.5%.

The OpenAI family moved the other way. Its share rose from 22.1% in March to 35.1% in August. The gap between the two families, 48.0 points in March, compressed to 13.4 by August.

The decline is not concentrated in one feed. On hn_new, the highest-volume source, Anthropic family mentions fell from 1,105 in June to 614 in August, down 44%. On hn_top, mentions fell from 63 to 37, down 41%. Lobsters held flat at roughly one to three mentions per month.

Two triggers, one window

The slide began with a single post. On June 30, a Hacker News item titled "Claude Code is steganographically marking requests" reached 2,425 points, the highest-scoring Anthropic-tagged story in the study window. It opened a run of critical coverage. A Zig maintainer accused Anthropic of "blowing smoke" (1,473 points, July 13). A report that Claude Code sends 33,000 tokens before reading a prompt scored 681 (July 12). A judge approved a $1.5 billion settlement over pirated books (555 points, July 21). On July 3, a story about Alibaba banning Claude Code in the workplace over backdoor risks scored 329.

Nine days after the steganography post, OpenAI launched GPT-5.6. The announcement reached 1,546 points on July 9, preceded by GPT-Live at 730 on July 8. A follow-up on price-performance scored 599 on July 30. The two cycles overlapped in mid-July. Anthropic's share, already falling, did not recover.

A third force

DeepSeek entered the frame in late July. Its share of AI-model conversation rose from 0.4% in March to 7.5% in August. Weekly mentions over the last eight weeks ran 7, 10, 7, 12, 24, 26, 45, 21. The week of August 10, with 45 DeepSeek mentions, approached Claude Code's 51 in the same period.

The trigger was a release cycle. DeepSeek-V4-Flash Update scored 735 points on July 31. DeepSeek V4 Pro 0813 scored 1,031 on August 12. DeepSeek Harness developer preview scored 729 on August 13. DeepSeek now ranks third among AI-model families by conversation share, ahead of Gemini, Llama, Mistral, Kimi, and Qwen individually.

The divergence

Conversation fell. Adoption rose. Between June 12 and August 22, Anthropic's PyPI downloads climbed from 31.4 million per week to 49.6 million, up 58.2%. Its npm downloads rose from 23.6 million to 28.4 million per week, up 20.4%. Claude Code's npm downloads went from 11.1 million to 18.2 million per week, up 64.3%, with 62.1% of that growth in the last 28 days. Those 28 days are the same period as the steepest conversation decline.

OpenAI's adoption also rose. PyPI went up 54.3% (70.7M to 109.2M per week); npm went up 13.4% (25.5M to 29.0M). But OpenAI's conversation and adoption moved in the same direction. Only Anthropic shows the split: developers are installing Claude Code at a record pace while the discourse around it has turned skeptical.