Why not just run Claude Fable 5 on every message — isn't quality more important than cost?

Quality is more important than cost, yes — but quality doesn't actually go up if you use a more capable model on a task the cheaper model already solves correctly. 95% of Telegram messages are «obviously clean» or «obviously spam» — deterministic filters and small models classify them at the same accuracy as frontier models, in milliseconds, for a fraction of a cent. Routing those messages through Fable 5 burns money without improving outcomes. Frontier capability matters on the hard 5%, and that's where we use it.

What's the actual cost per message at Varta scale?

It depends on which tier handles the message. Tier 1 (deterministic filters, regex, account-age signals) costs essentially zero per message — microseconds of CPU time. Tier 2 (sender history lookup, cross-group reputation query) costs a fraction of a millicent — a database read. Tier 3 (small LLM triage on borderline messages) costs around $0.0005 per call. Tier 4 (frontier model on the genuinely ambiguous cases) costs $0.01-0.05 per call. Average cost across all messages, weighted by tier frequency: well under $0.001 per message. Running Fable 5 on every message would average around $0.01-0.03 per message — 10-30× more.

Won't the cheaper tiers miss spam that only Fable 5 could catch?

They'd miss spam if we routed messages to the wrong tier. The pipeline is designed so that every tier knows when it's unsure — and unsure messages escalate to the next tier automatically. Tier 1 doesn't try to handle ambiguous cases; it confidently classifies the easy ones and passes the rest up. Same for Tiers 2 and 3. By the time a message reaches Fable 5, it's a genuinely hard case where the smaller models all reported low confidence. The frontier model handles those well — and only those.

Does the pipeline approach work for use cases other than spam moderation?

Yes — it's the general pattern for any production AI system at scale. Customer support classification, content moderation, document categorization, sentiment analysis, anomaly detection. The principle is the same: most inputs are easy and don't need the most capable model; the hard inputs are where the capability matters. Building the routing layer — knowing when to escalate — is the actual engineering work. The cost savings compound; teams that skip this end up either burning runway or capping their AI usage in ways that hurt the product.

Should You Use Claude Fable 5 for Telegram Moderation?

Anthropic released Claude Fable 5 and Claude Mythos 5 today — the next generation of intelligence for the hardest knowledge work and coding problems.

If you build with LLMs, you've probably already started reading the release notes. If you build AI moderation for Telegram groups, you're going to face a temptation in the next few weeks: just route every message through Fable 5. The model is more capable. The classifications will be better. Done.

Don't. Here's the math, and the architecture we built instead.

Claude Fable 5 and the moderation temptation

The temptation is structural. Every time a more capable frontier model ships, teams in the «throw the best LLM at every problem» school of architecture get a free upgrade — their existing pipeline calls a smarter model and outputs better classifications. It feels like progress.

For moderation specifically, the temptation is sharper. Spam in 2026 is genuinely sophisticated — Cyrillic-Latin substitution tricks, image-embedded URLs, slow-funnel conversations across multiple messages, admin-impersonator profile clones. A more capable model handles edge cases better. The argument writes itself.

The problem is that the argument is correct about quality and wrong about economics. At Telegram-community scale, «throw the best LLM at every problem» becomes a per-month bill that breaks the unit economics of free-tier moderation.

The cost math of running frontier on every message

Let me make this concrete with production-realistic numbers.

A typical Telegram community Varta protects processes maybe 30-100 messages per day in active periods. The 48-community network handles around 2,000-5,000 messages per day across all of them combined. Each message averages around 50-200 tokens after we strip metadata.

If we ran every message through a frontier model — call it $15 per million output tokens at current Sonnet-class pricing, more for the truly top-tier — the per-message cost lands around $0.005-0.03. At 5,000 messages a day, that's $750-4,500 a month. Across 48 protected communities.

For a free-tier moderation tool, those numbers don't work. Either you pass the cost to users (defeating the purpose of «free until first spam catch»), or you eat it (defeating the purpose of running a sustainable business). The economics of moderation-as-product require that most messages be classified for fractions of a cent each.

Meanwhile — and this is the part the naive approach misses — most messages don't actually need a frontier model to classify correctly. The clean ones are obviously clean. The crude spam is obviously crude. A frontier model gives the same verdict on those as a deterministic rule does, for orders of magnitude more cost. The capability differential only matters on the genuinely ambiguous cases.

The multi-stage pipeline

What we built instead: four tiers, escalating in capability and cost. Each tier handles what it can confidently handle and passes the rest up.

Tier	What handles it	Cost / msg	Volume share
1	Deterministic filters + account signals	~$0	60–70%
2	Sender history + cross-group reputation	< $0.0001	15–20%
3	Small LLM triage (cheap model)	~$0.0005	10–15%
4	Frontier model (Claude Fable 5)	$0.01–0.05	3–7%
Weighted average across all tiers		< $0.001	100%

Tier 1 — Deterministic filters and account signals. Catches: blatant promotional content, known-bad URL patterns, accounts that fail platform-level checks. Confidence: high, on the cases it acts on. Doesn't try to handle ambiguity — passes anything unclear to Tier 2.

Tier 2 — Sender history and cross-group reputation lookup. Catches: known spammers across the network (the cross-group reputation signal), repeat offenders, accounts flagged in other Varta-protected communities. Confident classifications act, ambiguous ones escalate.

Tier 3 — Small-model triage. Catches: routine spam patterns that need linguistic context but not heavy reasoning. Cyrillic-Latin substitution tricks, basic phishing attempts, low-effort cross-language scams. Confidence threshold is calibrated — borderline outputs go to Tier 4.

Tier 4 — Frontier model on the hard cases. Catches: genuinely ambiguous cases. Slow-funnel scams across multiple messages. Sophisticated admin impersonators. Borderline marketplace listings. Image content that requires vision + reasoning. Specifically for cases where lesser models reported low confidence.

For the 48-community network's ~5,000 daily messages, total LLM bill is in the low hundreds of dollars per month — not thousands.

The quality of moderation decisions doesn't drop because the escalation logic is calibrated. Every tier knows when it's unsure. The 3-7% of messages that reach Tier 4 get the same frontier capability they'd get in a naive «route everything to Fable 5» setup. The 93-97% of messages that resolve at lower tiers get the same correct classification for a fraction of the cost.

Where Claude Fable 5 actually belongs in the stack

Fable 5's strengths — frontier reasoning, complex contextual judgment, multi-step inference — are exactly what Tier 4 needs. Putting it there isn't a downgrade; it's the right placement.

Concretely, here's where Fable 5 starts being used in the Varta stack:

Slow-funnel conversation detection across 3-7 messages from the same sender (requires reasoning across context windows the smaller models can't hold well).
Admin-impersonator pattern detection where the impersonator's behavior is subtle (legitimate-seeming first messages, payment script in DMs only after building rapport).
Marketplace borderline listings where the question is «is this a legitimate seller or a scam funnel» — requires evaluating photos, prices, descriptions, and account context together.
Multilingual edge cases where the message mixes languages, idioms don't translate cleanly, and intent has to be inferred from cultural context.
Verdict-level reasoning for the live classifier demo on our landing page — when someone pastes a message to see how Varta would handle it, we want the frontier model's reasoning because the explanation is the product.

The frontier model earns its cost in those cases. It would earn the same cost on cases the smaller tiers also classify correctly — and earn nothing for the user, because the verdict doesn't change.

The broader principle for AI builders

The lesson generalizes beyond moderation.

The engineering question isn't «what's the best model for this task?» It's «what's the cheapest stage that can answer with sufficient confidence?»

Plus the corollary: «how do I know when this stage is unsure?»

The first question is about capability matching. The second is about calibration. Most teams skip the calibration step — they pick a model, run everything through it, ship — and end up either burning runway on tokens that don't improve outcomes or capping AI usage in ways that hurt the product. The teams that build the routing layer get both: same quality, fraction of the cost.

For moderation specifically, here's the rough pattern I'd recommend if you're building this from scratch:

Start with deterministic filters for the obvious cases — they're cheap and they handle most of your volume.
Add a sender history layer that lets known patterns short-circuit. Cross-group reputation, account-age signals, behavior history.
Use a cheap LLM as your default classifier for ambiguous cases. Calibrate it to know when it's unsure.
Reserve the frontier model for the hardest cases — the ones where the small model explicitly said «I don't know» or returned low confidence.
Log everything. Watch which messages each tier handles. Adjust thresholds when you see misrouting.

The architecture isn't novel. It's how production systems have been built for decades — escalation from cheap heuristics to expensive computation. What's new in 2026 is that the «expensive computation» tier is a frontier LLM, and the cost asymmetry between tiers is enormous.

Claude Fable 5 is a tool. It's a sharp one. Used in the right place in your stack, it makes the hard cases tractable. Used in every place, it burns money on cases that didn't need it.

If you're building anything AI-driven at scale, the multi-stage pipeline is worth your engineering time. The savings compound month over month. The quality doesn't drop. And when Fable 6 ships next year, your stack absorbs the upgrade at Tier 4 without ballooning your costs at Tiers 1-3.

That's the engineering reality behind moderation that costs $0.001 per message instead of $0.01. Same intelligence. Routed where intelligence is actually needed.

Varta is the Trust Layer for Telegram — AI moderation in 33 languages, built on a multi-stage pipeline that keeps frontier models for the hard cases. Free forever plan with basic keyword protection; the 5-day full-AI trial starts only when Varta catches your first spam. Add Varta for free →