> ## Documentation Index
> Fetch the complete documentation index at: https://docs.highailabs.com/llms.txt
> Use this file to discover all available pages before exploring further.
# Platform Cost Tracking
> Comprehensive cost analysis for all external APIs, AI models, compute, and storage across the High IQ / TIWIH platform — with per-unit math, scale projections, and vendor pricing links.
## Overview
This page tracks the cost of every external service that powers the High IQ platform. The biggest cost center is the **Strain Research Pipeline** (Trigger.dev v2) — a 7-stage AI pipeline that researches, enriches, and generates images for cannabis strains. The **Strain Music Pipeline** adds AI-generated songs per strain. Other platform costs (auth, hosting, database) are documented at the bottom.
**Self-serve model (March 2026):** The pipeline no longer runs in batch. When a user visits a strain page with no research, they tap a button to trigger on-demand research via direct mode. Results appear in \~5-10 minutes. This means **all strain research runs at direct-mode pricing**.
**Data snapshot (July 3, 2026):** **2,525 strains researched**, **2,750 in the unprocessed queue**, and **\~25,000 unmatched competitor strains tracked** (Leafly/AllBud sitemaps). Pipeline throughput has been **≈0 strains/week recently** — a restart decision is pending. At direct-mode pricing ($0.57/strain), the theoretical cost to process the current queue is **~$1,568\*\* for research + **\~$239** for music = **~$1,807 total**.
***
## Strain Research Pipeline (Trigger.dev v2)
The pipeline chains 7 stages, each with its own external API calls. Here's exactly what happens per strain and what it costs.
### Pipeline Architecture
Neural search for web articles about the strain. Returns 25 results with extracted text snippets.
| Detail | Value |
| ------------------- | -------------------------------- |
| API | Exa (`search()`) |
| Calls per strain | **1** |
| Results requested | 25 |
| Search type | `auto` (hybrid neural + keyword) |
| Max chars/result | 5,000 |
| Trigger.dev machine | small-1x |
Filters search results to URLs that actually mention the strain name. Uses 11 pattern variants (slugs, no-spaces, accent-normalized, spelling variants). **No external API calls.**
| Detail | Value |
| -------------- | ----------------------------- |
| API | None (local pattern matching) |
| Typical output | 5-15 relevant URLs |
Batch-scrapes all filtered URLs to extract markdown content.
| Detail | Value |
| ------------------- | ------------------------------------ |
| API | Firecrawl (`asyncBatchScrapeUrls()`) |
| Calls per strain | **1 batch call** |
| URLs per batch | 5-15 (avg \~10) |
| Credits consumed | **1 credit per URL** (\~10/strain) |
| Format | Markdown only, main content |
| Trigger.dev machine | medium-2x |
Structured data extraction from each scraped article — terpenes, effects, aromas, lineage, awards, cannabinoids.
| Detail | Value |
| --------------------- | ------------------------------------ |
| API | Google Gemini |
| Model | `gemini-3.1-flash-lite` |
| Calls per strain | **\~10** (one per article, parallel) |
| Input tokens/article | \~12,500 (50k char limit) |
| Output tokens/article | \~2,000 (structured JSON) |
| Total tokens/strain | \~125k input + \~20k output |
| Trigger.dev machine | medium-2x (concurrency: 40) |
Three-pass FactFinder + Copywriter + Image Prompt generation. This is the most expensive stage.
**5 API calls per strain:**
| Pass | Model | Max Tokens | Purpose |
| ------------------ | --------------------------- | ---------- | ----------------------------------------------- |
| FactFinder Core | `claude-haiku-4-5-20251001` | 4,096 | Strain type, lineage, cannabinoids |
| FactFinder Profile | `claude-haiku-4-5-20251001` | 4,096 | Terpenes, effects, aromas |
| FactFinder Meta | `claude-haiku-4-5-20251001` | 4,096 | Medical uses, awards, conflicts |
| Copywriter | `claude-sonnet-4-6` | 8,192 | Marketing descriptions, Professor High insights |
| Image Prompts | `claude-sonnet-4-6` | 4,096 | Hero + mascot image generation prompts |
The 3 FactFinder passes run in **parallel** (`Promise.all()`). Copywriter and Image Prompts run sequentially with 6-second checkpointed waits between them.
| Detail | Value |
| ------------------- | --------------------------- |
| Total Haiku tokens | \~30k input + \~9k output |
| Total Sonnet tokens | \~20k input + \~9k output |
| Trigger.dev machine | medium-2x (concurrency: 15) |
Generates 2 AI images per strain at 2K resolution (2048×2048px).
| Detail | Value |
| ------------------- | -------------------------------------------------------- |
| API | Google Gemini |
| Model | `gemini-3.1-flash-image-preview` |
| Images per strain | **2** (hero + mascot) |
| Resolution | **2K** (2048×2048px, 1680 output tokens/image) |
| Storage | Supabase (`strain-images/` and `mascot-images/` buckets) |
| Trigger.dev machine | large-1x (4 vCPU, 8GB RAM for base64) |
Atomic dual-table write via `upsert_strain_dual_table` RPC. Writes to both `strains_master` (JSONB) and `strains_v2` (columns) in a single transaction.
| Detail | Value |
| ------ | ------------------------- |
| API | Supabase RPC |
| Cost | Included in Supabase plan |
***
### Per-Strain Cost: Direct Mode (Self-Serve)
**This is the production pricing model.** Since we switched to self-serve (on-demand per user visit), all strains are processed in direct mode.
| Stage | Service | Units | Cost |
| ---------------- | ------------------------------------- | --------------------- | --------------------- |
| 01 Search | Exa (1 search, 25 results) | 1 API call | **\$0.007** |
| 02 Filter | Local pattern matching | — | \$0.000 |
| 03 Scrape | Firecrawl (\~10 URLs) | \~10 credits | **\$0.010** |
| 04 Extract | Gemini 3.1 Flash Lite (\~10 articles) | \~125k in + \~20k out | **\$0.061** |
| 05 FactFinder | Claude Haiku 4.5 (×3 passes) | \~30k in + \~9k out | **\$0.075** |
| 05 Copywriter | Claude Sonnet 4.6 | \~15k in + \~6k out | **\$0.135** |
| 05 Image Prompts | Claude Sonnet 4.6 | \~5k in + \~3k out | **\$0.060** |
| 06 Imaging | Gemini 3.1 Flash Image (2 × 2K) | 2 images | **\$0.202** |
| 07 Save | Supabase RPC | — | \$0.000 |
| Compute | Trigger.dev (8 runs, \~105s active) | — | **\$0.021** |
| **Total** | | | **\~\$0.57 / strain** |
#### Cost by Provider
| Provider | Cost/Strain | Share |
| ------------------ | ----------- | ----- |
| Anthropic (Claude) | \$0.270 | 47% |
| Google (Gemini) | \$0.263 | 46% |
| Trigger.dev | \$0.021 | 4% |
| Firecrawl | \$0.010 | 2% |
| Exa | \$0.007 | 1% |
#### Cost by Stage
| Stage | Cost/Strain | Share |
| ------------------------- | ----------- | ----- |
| 06 Imaging (Gemini) | \$0.202 | 35% |
| 05 Copywriter (Sonnet) | \$0.135 | 24% |
| 05 FactFinder (Haiku ×3) | \$0.075 | 13% |
| 04 Extract (Gemini) | \$0.061 | 11% |
| 05 Image Prompts (Sonnet) | \$0.060 | 11% |
| Compute (Trigger.dev) | \$0.021 | 4% |
| 03 Scrape (Firecrawl) | \$0.010 | 2% |
| 01 Search (Exa) | \$0.007 | 1% |
***
### Per-Strain Cost: Batch Mode (Reference)
Batch mode is no longer the primary processing model but remains available for bulk backfills or special operations. Batch discounts apply to **Stage 05** (Anthropic Batch API, 50% off) and **Stage 06** (Gemini Batch API, 50% off). Other stages are unchanged.
| Stage | Direct | Batch | Savings |
| ------------------------- | ----------- | ----------- | ------- |
| 01 Search (Exa) | \$0.007 | \$0.007 | — |
| 02 Filter (Local) | \$0.000 | \$0.000 | — |
| 03 Scrape (Firecrawl) | \$0.010 | \$0.010 | — |
| 04 Extract (Gemini) | \$0.061 | \$0.061 | — |
| 05 FactFinder (Haiku ×3) | \$0.075 | **\$0.038** | 50% |
| 05 Copywriter (Sonnet) | \$0.135 | **\$0.068** | 50% |
| 05 Image Prompts (Sonnet) | \$0.060 | **\$0.030** | 50% |
| 06 Imaging (Gemini 2×2K) | \$0.202 | **\$0.101** | 50% |
| 07 Save (Supabase) | \$0.000 | \$0.000 | — |
| Compute (Trigger.dev) | \$0.021 | \$0.018 | \~14% |
| **Total** | **\$0.571** | **\$0.333** | **42%** |
***
## Strain Music Pipeline
Auto-fires after research completes. Generates a unique AI song per strain using Google Lyria 3 Pro. Always runs in direct mode (no Lyria batch API exists).
Fetches strain data and downloads hero/mascot images for multimodal input. **No AI call.**
AI generates a creative, strain-specific Lyria 3 prompt from the strain's terpene profile, effects, aromas, and Professor High insight.
| Detail | Value |
| ------ | ------------------------ |
| Model | `gemini-3.1-flash-lite` |
| Tokens | \~2k input + \~2k output |
Generates a full 3-minute song with lyrics. Takes the text prompt + hero image + mascot image as multimodal input.
| Detail | Value |
| ----------- | ------------------------------- |
| Model | `lyria-3-pro-preview` |
| Price | **\$0.08 per song** (fixed) |
| Output | MP3 audio (\~4MB) + lyrics text |
| Concurrency | 3 (rate-limited) |
Uploads MP3 to Supabase Storage (`strain-music` bucket) and updates `strains_v2` with audio metadata. **No AI call.**
### Per-Strain Music Cost
| Component | Cost |
| --------------------------------------- | ---------------------- |
| Gemini 3.1 Flash Lite (prompt building) | \$0.001 |
| Lyria 3 Pro (full song) | **\$0.080** |
| Trigger.dev (5 runs, \~80s active) | \$0.006 |
| Supabase Storage (\~4MB) | \~\$0.000 |
| **Total** | **\~\$0.087 / strain** |
***
## Combined Per-Strain Total
| Pipeline | Direct Mode | Batch Mode |
| ------------------- | ----------- | ----------- |
| Research (7 stages) | \$0.571 | \$0.333 |
| Music (4 stages) | \$0.087 | \$0.087 |
| **Combined** | **\$0.658** | **\$0.420** |
***
## Scale Projections
### Self-Serve Model (Direct Mode — Production)
All strains researched on-demand at direct-mode pricing ($0.57 research + $0.087 music).
| Strains | Cost |
| ------------------------- | ----------- |
| 10 | \$5.71 |
| 50 | \$28.55 |
| 100 | \$57.10 |
| 500 | \$285.50 |
| 1,000 | \$571 |
| **2,750 (current queue)** | **\$1,570** |
| 5,000 | \$2,855 |
| 10,000 | \$5,710 |
| Strains | Cost |
| ------------------------- | ----------- |
| 10 | \$6.58 |
| 50 | \$32.90 |
| 100 | \$65.80 |
| 500 | \$329 |
| 1,000 | \$658 |
| **2,750 (current queue)** | **\$1,810** |
| 5,000 | \$3,290 |
| 10,000 | \$6,580 |
### Batch Mode (Reference — Bulk Backfill)
If batch mode were used for a large backfill operation ($0.33 research + $0.087 music).
| Strains | Research Only | Research + Music |
| ------------------------- | ------------- | ---------------- |
| 1,000 | \$333 | \$420 |
| **2,750 (current queue)** | **\$916** | **\$1,155** |
| 5,000 | \$1,665 | \$2,100 |
| 10,000 | \$3,330 | \$4,200 |
### Self-Serve Usage Estimates
Since the pipeline is now self-serve (triggered by user visits), actual costs depend on user traffic:
| Daily Researches | Monthly Cost (Research) | Monthly Cost (Research + Music) |
| ---------------- | ----------------------- | ------------------------------- |
| 5/day | \$85.50 | \$98.70 |
| 10/day | \$171 | \$197.40 |
| 25/day | \$427.50 | \$493.50 |
| 50/day | \$855 | \$987 |
| 100/day | \$1,710 | \$1,974 |
***
## Monthly Fixed Costs (Subscriptions)
These are paid regardless of usage volume.
| Service | Plan | Monthly Cost | What's Included |
| -------------------------- | ---------------- | -------------- | -------------------------------------------- |
| Trigger.dev | Pro | **\$50** | \$50 compute credit (\~2,380 strains worth) |
| Firecrawl | Standard | **\$99** | 100k scrape credits (\~10,000 strains worth) |
| Exa | Pay-per-use | \$0 base | First 1,000 searches/month free |
| Anthropic | Pay-per-use | \$0 base | — |
| Google AI (Gemini + Lyria) | Pay-per-use | \$0 base | — |
| Supabase | Pro (shared) | Already paying | Included with platform |
| Vercel | Pro (3 projects) | Already paying | API + Website + Mobile Web hosting |
| Convex | Pro | Already paying | User data, real-time sync |
| Clerk | Pro | Already paying | Authentication |
| **Total fixed** | | **\~\$149/mo** | (pipeline-specific) |
### Credit Budgeting
* **100,000 credits/month** included in Standard plan
* At \~10 URLs/strain: supports **\~10,000 strains/month**
* Extra credits if needed: $57/30,000 = **$1.90 per 1,000 credits\*\*
* At self-serve pace (25/day): \~750 strains/month — well within included credits
* **\$50 included usage/month** on Pro plan
* At \$0.021/strain compute: covers **\~2,380 strains/month**
* At self-serve pace (25/day): \~750 strains/month × $0.021 = **$15.75/month\*\* — well within budget
* Beyond included: billed at per-second machine rates
* First **1,000 searches/month free** on paid plans
* Beyond free tier: **\$7 per 1,000 searches**
* At self-serve pace (25/day): \~750/month — covered by free tier
***
## Vendor Pricing Reference
Direct links to current pricing documentation for every external service used.
### AI Models
| Vendor | Pricing Page | Models We Use |
| ------------------ | --------------------------------------------------------------------------------------- | ------------------------------- |
| Anthropic (Claude) | [anthropic.com/pricing](https://docs.anthropic.com/en/docs/about-claude/pricing) | Haiku 4.5, Sonnet 4.6 |
| Google Gemini | [ai.google.dev/pricing](https://ai.google.dev/gemini-api/docs/pricing) | 3.1 Flash Lite, 3.1 Flash Image |
| Google Lyria 3 | [ai.google.dev/pricing](https://ai.google.dev/gemini-api/docs/pricing) (bottom of page) | Lyria 3 Pro (\$0.08/song) |
### Search & Scraping
| Vendor | Pricing Page | What We Use |
| --------- | ---------------------------------------------------------- | ------------------------ |
| Exa | [exa.ai/pricing](https://exa.ai/pricing) | Search endpoint (\$7/1k) |
| Firecrawl | [firecrawl.dev/pricing](https://www.firecrawl.dev/pricing) | Scrape (1 credit/page) |
### Infrastructure
| Vendor | Pricing Page | What We Use |
| ----------- | ---------------------------------------------------- | ------------------------------- |
| Trigger.dev | [trigger.dev/pricing](https://trigger.dev/pricing) | Pro plan, compute + invocations |
| Vercel | [vercel.com/pricing](https://vercel.com/pricing) | Pro plan, 3 projects |
| Supabase | [supabase.com/pricing](https://supabase.com/pricing) | Pro plan, DB + Storage |
| Convex | [convex.dev/pricing](https://convex.dev/pricing) | Pro plan, user data |
| Clerk | [clerk.com/pricing](https://clerk.com/pricing) | Pro plan, auth |
***
## Unit Pricing Tables
Detailed per-token / per-unit pricing for the specific models and tiers we use.
### Anthropic Claude (Direct vs Batch)
| Model | Input (Direct) | Output (Direct) | Input (Batch) | Output (Batch) |
| --------------------------------------- | -------------- | --------------- | ------------- | -------------- |
| Haiku 4.5 (`claude-haiku-4-5-20251001`) | \$1.00/MTok | \$5.00/MTok | \$0.50/MTok | \$2.50/MTok |
| Sonnet 4.6 (`claude-sonnet-4-6`) | \$3.00/MTok | \$15.00/MTok | \$1.50/MTok | \$7.50/MTok |
**MTok** = million tokens. Batch API gives a flat 50% discount on both input and output.
### Google Gemini (Standard vs Batch)
| Model | Input (Std) | Output (Std) | Input (Batch) | Output (Batch) |
| ------------------------------ | ----------- | ------------ | ------------- | -------------- |
| 3.1 Flash Lite (text) | \$0.25/MTok | \$1.50/MTok | \$0.125/MTok | \$0.75/MTok |
| 3.1 Flash Image (text input) | \$0.50/MTok | \$3.00/MTok | \$0.25/MTok | \$1.50/MTok |
| 3.1 Flash Image (per 2K image) | — | **\$0.101** | — | **\$0.050** |
### Trigger.dev Compute (Pro Plan)
| Machine | vCPU | RAM | Cost/sec | Cost/hour |
| --------- | ---- | ------- | ----------- | --------- |
| micro | 0.25 | 0.25 GB | \$0.0000169 | \$0.061 |
| small-1x | 0.5 | 0.5 GB | \$0.0000338 | \$0.122 |
| medium-2x | 2 | 4 GB | \$0.0001700 | \$0.612 |
| large-1x | 4 | 8 GB | \$0.0003400 | \$1.224 |
Run invocation: \*\*$0.000025/run** ($0.25 per 10,000 runs). Waits longer than 5 seconds are checkpointed and **free** (no compute charges).
***
## Cost Math: Worked Example
Here's the full math for one strain researched in direct mode, showing exactly how each number is calculated.
* 1 search call
* Exa pricing: \$7 per 1,000 searches
* Cost: 1 / 1,000 × $7 = **$0.007\*\*
* \~10 URLs scraped (1 credit each)
* Standard plan: 100k credits for $99 = $0.00099/credit
* Cost: 10 × $0.00099 = **~$0.010\*\*
* \~10 articles × 12,500 input tokens = 125,000 input tokens
* \~10 articles × 2,000 output tokens = 20,000 output tokens
* Input: 125,000 / 1,000,000 × $0.25 = $0.03125
* Output: 20,000 / 1,000,000 × $1.50 = $0.03000
* **Total: \~\$0.061**
* 3 passes × \~10k input tokens = 30,000 input tokens (Haiku 4.5)
* 3 passes × \~3k output tokens = 9,000 output tokens
* Input: 30,000 / 1,000,000 × $1.00 = $0.030
* Output: 9,000 / 1,000,000 × $5.00 = $0.045
* **Total: \$0.075**
* \~15,000 input tokens (Sonnet 4.6)
* \~6,000 output tokens (max 8,192)
* Input: 15,000 / 1,000,000 × $3.00 = $0.045
* Output: 6,000 / 1,000,000 × $15.00 = $0.090
* **Total: \$0.135**
* \~5,000 input tokens (Sonnet 4.6)
* \~3,000 output tokens (max 4,096)
* Input: 5,000 / 1,000,000 × $3.00 = $0.015
* Output: 3,000 / 1,000,000 × $15.00 = $0.045
* **Total: \$0.060**
* 2 images at 2K resolution (2048×2048px)
* Each 2K image = 1,680 output tokens
* Per image: 1,680 / 1,000,000 × $60.00 = $0.1008
* Text input prompts (\~4k tokens): 4,000 / 1,000,000 × $0.50 = $0.002
* **Total: 2 × $0.101 + $0.002 = \~\$0.202**
* 8 task runs (orchestrator + 7 stages)
* Run invocations: 8 × $0.000025 = $0.0002
* Compute (checkpointed waits are free):
* Orchestrator: \~3s × small-1x = \$0.0001
* Stage 01: \~8s × small-1x = \$0.0003
* Stage 02: \~3s × micro = \$0.0001
* Stage 03: \~8s × medium-2x = \$0.0014
* Stage 04: \~30s × medium-2x = \$0.0051
* Stage 05: \~20s × medium-2x = \$0.0034
* Stage 06: \~30s × large-1x = \$0.0102
* Stage 07: \~3s × small-1x = \$0.0001
* **Total: \~\$0.021**
* Gemini prompt builder: \~2k in + \~2k out at Flash Lite pricing = \~\$0.001
* Lyria 3 Pro: fixed **\$0.080** per full song
* Trigger.dev compute (5 runs, \~80s): \~\$0.006
* **Total: \~\$0.087**
***
## Potential Cost Optimizations
Ranked by impact, for future consideration:
| Optimization | Savings/Strain | Impact | Trade-off |
| ------------------------------------- | -------------- | ---------- | --------------------------------- |
| Switch to 1K images (1024px) | \~\$0.068 | 12% | Lower image quality |
| Use Haiku for Image Prompts | \~\$0.040 | 7% | Potentially less creative prompts |
| Use Gemini Batch for extraction | \~\$0.030 | 5% | Adds latency (bad for self-serve) |
| Reduce search to 15 results | \~\$0.020 | 4% | Fewer source articles |
| Use Lyria 3 Clip (30s) instead of Pro | \~\$0.040 | Music: 46% | Short clips instead of full songs |
Most batch-mode optimizations **don't apply** to the self-serve model since users expect results in 5-10 minutes. The Anthropic Batch API takes hours and is incompatible with on-demand research.
***
## Other Platform Costs
Services that are already paid and shared across the platform (not per-strain).
| Service | Plan | Monthly | Purpose |
| ---------- | ---------------- | ------- | -------------------------------- |
| Vercel | Pro (3 projects) | \~\$60 | API, Website, Mobile Web hosting |
| Supabase | Pro | \~\$25 | PostgreSQL, Storage, Auth |
| Convex | Pro | \~\$25 | Real-time user data sync |
| Clerk | Pro | \~\$25 | Authentication |
| Mintlify | Starter | \~\$0 | Documentation site |
| Sentry | Team | \~\$26 | Error monitoring |
| EAS (Expo) | Production | \~\$99 | iOS builds and submissions |
These are approximate monthly costs. Actual amounts may vary based on usage tiers and billing cycles. Check each vendor dashboard for current billing.
***
## Version History
| Date | Change |
| -------------- | ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
| March 29, 2026 | Initial cost analysis — all models verified from code, pricing from vendor docs |
| July 3, 2026 | Data snapshot corrected: 2,525 strains researched, 2,750 in unprocessed queue, \~25,000 unmatched competitor strains tracked; pipeline throughput ≈0/week recently (restart decision pending). Backlog cost projections updated to the current queue size. |