model

[autumn wasp]

which model is the best to run crypto analysis and scraping some info from web daily

[half barn]

To help others find answers, you can mark your question as solved via Right click solution message -> Apps -> ✅ Mark Solution

[solemn tiger]

For “crypto analysis + daily web scraping,” the model matters less than tool-calling reliability + cost, because scraping is where models tend to break (bad JSON/tool args, getting stuck in browser flows, etc.).

What usually works best in OpenClaw

• Use a very tool-reliable model for the scraping/extraction step (calls web_fetch/browser correctly, follows schemas, returns clean structured output). In practice, the Anthropic tool-use models tend to be the safest default for this right now.
• Optionally use a cheaper model for the “summarize + trend notes” step once you already have clean text/JSON. This keeps daily runs affordable.

Practical approach (most stable)

1. Don’t “scrape with an LLM.” Scrape deterministically (RSS/APIs/HTML fetch), then hand the cleaned text to the model.
2. Prefer web_fetch for static pages; only use browser automation for JS-heavy sites.
3. Ask the model for strict JSON output (e.g., {source, timestamp, headline, key_numbers, sentiment, risks}) and validate it.

If you tell me these, I can suggest a tighter setup

• Which sources? (Twitter, news sites, CoinGecko/CoinMarketCap, specific project docs, on-chain dashboards, etc.)
• Rough volume per day (5 pages vs 500)?
• Budget/latency targets?

Docs to anchor on (and community notes that stay current): https://docs.openclaw.ai/concepts/models (see the “model discussions / community notes” section).

[autumn wasp]

twitter tik tok like 400 pages buget for month not more then 500$ GIVE ME TOP 3

[solemn tiger]

I can’t give “top 3 model names” here because it goes stale fast and people get burned by outdated picks. What I can do is give you the top 3 setups that reliably hit your use case (Twitter/TikTok-ish sources, ~400 pages/day, ≤$500/mo), and what to ask in this channel to get the current “top 3 models” from folks running them today.

Reality check: 400 pages/day under $500/mo

You only make that budget if you don’t feed full pages into the LLM. You need a pipeline that:

• pulls data via official APIs / RSS / 3rd-party data providers (Twitter/TikTok HTML scraping is brittle and often against ToS),
• dedupes + filters hard,
• sends the model only small extracts (caption, metrics, top comments, linked article summary, etc.).

Top 3 practical setups (what I’d run)

1) Reliability-first (simplest, costs more)

• One tool-reliable “premium” model does: fetch → extract → normalize JSON → analysis.
• Use this if you really care about “never breaks” more than cost.

2) Best bang-for-buck (recommended for your budget)

• Deterministic scraping/collection (no LLM)
• Cheap model for first-pass: classify relevance + produce a tiny structured summary per item
• Premium tool-reliable model only for the top X% that matter (deep dive, cross-source correlation)

This is how you keep cost predictable.

3) Ultra-budget (works if you accept lower quality)

• Deterministic collection
• No per-page LLM at all
• Only a cheap model on an aggregated daily digest (top N items) for narrative + insights

If you want the community to reply with “top 3 models” today

Reply with:

• your exact sources (Twitter API? TikTok API? which provider?)
• how many items per day (posts/videos) vs “pages”
• output: alerts? dashboard? CSV/JSON? trading signals?
• your max context per item (e.g. “title + caption + 5 comments + stats”)

Then people in #model can give current model picks that are known to behave well with tool calling + long daily runs.

[autumn wasp]

just results like geting video then getting it transcribed, or getting quick analyse of crypto market so it can see where narative comes from then find and give thoughs about could the narative become bigger nad more popular also findng new niches on tik tok and yt

[solemn tiger]

If your goal is “spot narratives early from TikTok/YouTube/Twitter-style feeds” with video → transcript → crypto narrative analysis, you’ll get the best results (and stay under $500/mo) by designing it as a 2-stage pipeline, not “one model does everything”.

A setup that actually works (and fits $500/mo)

Stage A — Ingest + transcript (keep LLM out of it)

1. Collect metadata per item: platform, author, views/likes/comments, timestamp, caption/description, hashtags, linked URLs.
2. Get text without transcription when possible:
   • YouTube: use existing captions/transcripts when available.
   • TikTok: caption text + on-screen text (if you do OCR) often captures the “narrative” better than audio alone.
3. If you must transcribe audio: use a cheap transcription path (Whisper-class).
   • Budget sanity: if you process ~400 videos/day and average 1–2 min each → 12k–24k min/month. Whisper-style pricing is often low enough that transcription won’t be your main cost if you keep durations short.

Stage B — LLM analysis (where model choice matters)

Do cheap pass → expensive pass:

• Pass 1 (cheap, high volume): turn each item into a tiny structured record:
  • detected tickers/projects
  • claim type (news, rumor, prediction, “this coin will…”, explainer)

• stance/sentiment
  • key quote(s)
  • “narrative tags” (e.g. AI agents, RWA, L2 wars, memecoin meta, airdrop farming)
• Pass 2 (premium, low volume): only for the top ~5–20%:
  • cluster-level narrative summary
  • why it’s spreading (hooks, creators, formats)
  • “will it grow?” indicators (velocity, cross-platform lift, novelty, rebuttals)
  • actionable watchlist + risks

That’s how you keep cost predictable and quality high.

How to detect “narratives” and “new niches” (the important part)

Don’t rely on the LLM to “notice” trends from raw text. Do this:

• Embeddings + clustering over your per-item summaries (or transcript snippets)
• Track each cluster’s momentum:
  • growth rate (posts/day), engagement-weighted growth
  • creator diversity (1 influencer vs many small accounts)