Claude Sonnet 4.6

The 89 comments (top 25 by upvotes shown, all HackerNews) come from an agentic-coding-heavy audience, and the direct product reports split into capability praise and cost/skepticism signals. A team at 'mocha' running a coding-agent harness on 3-10 minute web-app tasks reports Sonnet 4.6 'feels like a fundamentally different model than Sonnet 4.5... much closer to the Opus series in agentic behavior and autonomy' — in zero-shot app-building it ran 3-4x longer without intervention and produced functional apps 'on par in terms of quality' with Opus. Another user asked it to build 'a working web based email client from scratch in rust which can interact with a JMAP based mail server. It did. It took about 20 minutes,' with a first pass containing a few bugs (polling instead of streaming) that were fixed via prompting. Counterweight: an independent eval reproduced 'an increase in output token amount of roughly 15-45% compared to 4.5,' concentrated in task-inference/evaluation prompts — a real verbosity/cost signal for per-token users. Older-model skepticism persists: a user reports Sonnet 4.5 'pretended to do some algebra' on the bat-and-ball question, arguing the LLM has no semantic understanding of its output. Much of the highest-upvoted discussion is macro rather than product-specific: labor displacement ('one engineer... do the work of 3'), SaaS consolidation risk, self-hosted agent CI/CD setups (Gitea/GitLab, Docker worktrees), and a concrete prompt-injection warning — LLM agents reading web content (e.g., a bank website or check memo) reintroduce in-band-signaling vulnerabilities. Net: developers are shipping real software with 4.6 and calling the agentic jump genuine, while cost-aware users measure it talking more to do the same job.

No aggregate ratings were found for this product during the last harvest.