Source: Twitter/X via agent-reach (OpenCLI backend), 2026-07-06. 71 unique tweets across 6 queries. Design principle (from @brian_armstrong, ♥6.1k): keep spend flat with better defaults, routing, and caching — not friction and alerts. So this dashboard is a visibility/routing tool, not a nag.
| # | Pain point | Evidence (reach) | Feature that answers it |
|---|---|---|---|
| 1 | Silent context bloat — Claude Code resends the whole session (system prompt + tools + history) every message; you pay it all as input and never see it | "$20 to type 'hey'" = 847k tokens on one msg, 1.4k upvotes; @theo effort tips ♥8.3k | Context-weight meter per session + est. cost-per-message |
| 2 | Silent Opus misrouting — Fable classifiers fall back to Opus 4.8 on routine coding, and Opus is where the money goes | @bridgemindai "$321 session, 75% routed to Opus" ♥3.6k | Per-model spend split (Fable/Opus/Sonnet/Haiku) |
| 3 | The image-OCR cost trick is viral — turn context into PNGs, cut input ~70% (this is pxpipe) | @IntCyberDigest ♥4.8k, @Polymarket ♥2.7k | pxpipe savings hero — live % + tokens saved |
| 4 | Fable → credit billing urgency — "you have Fable for a few days"; $10/M in, $50/M out after | @shadcn /improve ♥6.3k, @PrajwalTomar, @eliana | Fable Jul-7 countdown + credit-cost estimator |
| 5 | Effort level = furnace — xhigh/max burn quota with worse output | @theo ♥8.3k ("only use high; xhigh is token hungry") | Effort-level surface + usage burn-down |
| 6 | Caching is the real lever | @brian_armstrong ♥6.1k | Cache-efficiency (cache-read vs fresh input) |
Rates used for cost estimates (from the tweets / public pricing, $/M tokens):
- Fable 5: in 10, out 50 (credit billing)
- Opus 4.8: in 15, out 75
- Sonnet 5: in 3, out 15
- Haiku 4.5: in 0.80, out 4
- Cache read ≈ 10% of input rate; cache write ≈ 125% of input rate.