Super User Daily: April 18, 2026
Thin wire today. The Opus 4.7 drop sucked most of the oxygen out of the conversation — people were reacting to the press release, not running 4.7 in production yet. What's left worth keeping is two kinds of signal: one operator's field-test of how 4.7 actually behaves inside Claude Code, and a one-line frustration post from a user running their own terminal bench — Claude Code failing hard enough they're willing to risk an account ban to keep working.
@AI_masaou [Claude Code]
https://x.com/AI_masaou/status/2044898633670340721
Operator's field report on Opus 4.7 in Claude Code, not a press-release recap. The operationally important fact: xhigh is now Claude Code's default reasoning tier, sitting between high and max, and on Japanese workloads it drains rate limits unusually fast. API pricing held flat at $5 input and $25 output per million tokens, but real token consumption rises 1.0 to 1.35x because the tokenizer just updated and late-turn thinking tokens balloon on long runs. Internal numbers he cites: Rakuten's internal SWE bench hit 3x the resolved-task count vs 4.6, Notion saw tool errors drop to one-third, Hex reports low-effort 4.7 roughly equals medium-effort 4.6. The sharp warning is that loose old prompts break now — instruction-following got stricter so 4.7 takes you literally, meaning you have to re-tighten constraints and pair them with task_budgets to cap runs. The auto plus /ultrareview combo is the setup he'd reach for on fire-and-forget long tasks.
@usr_bin_roygbiv [Claude Code]
https://x.com/usr_bin_roygbiv/status/2044856048700248206
One-line field report: Claude Code is unusable on their current terminal-bench tests. They're considering hacking their subscription and risking a ban to pipe Claude Code into Droid, burning their last remaining Max sub as the bridge. Short post, but this is the signal that's been building for two weeks — power users will sacrifice Max subs to route around failure modes. Someone willing to risk an account ban just to keep a specific workflow alive is a louder data point than most "I love it" threads.
🗣 User Voice
User Voice
Instruction-following tightened on Opus 4.7, which means old loose prompts designed for 4.6 now produce literal, sometimes broken output. Operators are re-tightening constraints and adding task_budgets caps to stop runaway runs. Quote: AI_masaou.
API sticker price held flat but effective cost per task is up 1.0 to 1.35x because the tokenizer rev changed how the same input counts, and late-turn thinking tokens scale with run length. The takeaway: monitor tokens, not just dollars. Quote: AI_masaou.
When Claude Code fails a specific benchmark, users are willing to burn their remaining Max sub and risk a ban to hack their way onto an alternative runtime. That's how valuable the tool has become, and also how brittle the official channel is starting to feel. Quote: usr_bin_roygbiv.
Instruction-following tightened on Opus 4.7, which means old loose prompts designed for 4.6 now produce literal, sometimes broken output. Operators are re-tightening constraints and adding task_budgets caps to stop runaway runs. Quote: AI_masaou.
API sticker price held flat but effective cost per task is up 1.0 to 1.35x because the tokenizer rev changed how the same input counts, and late-turn thinking tokens scale with run length. The takeaway: monitor tokens, not just dollars. Quote: AI_masaou.
When Claude Code fails a specific benchmark, users are willing to burn their remaining Max sub and risk a ban to hack their way onto an alternative runtime. That's how valuable the tool has become, and also how brittle the official channel is starting to feel. Quote: usr_bin_roygbiv.
📡 Eco Products Radar
Eco Products Radar
No third-party product hit the three-mention threshold today. The conversation was almost entirely absorbed by Opus 4.7 itself and its Claude Code defaults. Droid, Obsidian, MCP, and Openclaw each showed up once as adjacent tooling but none clustered. On a thicker day expect the usual suspects — Cursor, Codex, Gemini CLI, Droid — to reappear.
No third-party product hit the three-mention threshold today. The conversation was almost entirely absorbed by Opus 4.7 itself and its Claude Code defaults. Droid, Obsidian, MCP, and Openclaw each showed up once as adjacent tooling but none clustered. On a thicker day expect the usual suspects — Cursor, Codex, Gemini CLI, Droid — to reappear.
Comments