Model Selection

Choosing the right AI models for each stage of analysis.

Available Models

Aparture supports models from three providers:

Anthropic (Claude)

Claude Opus 4.1 - Most capable

• Best for: PDF analysis, complex reasoning
• Speed: Slow
• Cost: Highest
• Context: 200K tokens
• Vision: Excellent

Claude Sonnet 4.5 - Balanced

• Best for: Abstract scoring, general use
• Speed: Medium
• Cost: Medium
• Context: 200K tokens
• Vision: Good

Claude Haiku 4.5 - Fast, intelligent & cheap

• Best for: Quick filtering, batch processing, coding
• Speed: Fast
• Cost: Low
• Context: 200K tokens (1M available on Developer Platform)
• Vision: Good

OpenAI (ChatGPT)

GPT-5 - High quality

• Best for: Any stage, comprehensive analysis
• Speed: Medium
• Cost: High
• Context: 272K tokens
• Vision: Excellent

GPT-5 Mini - Balanced

• Best for: Abstract scoring, moderate complexity
• Speed: Fast
• Cost: Medium
• Context: 272K tokens
• Vision: Good

GPT-5 Nano - Efficient

• Best for: Quick filtering, simple tasks
• Speed: Very fast
• Cost: Low
• Context: 272K tokens
• Vision: Basic

Google (Gemini)

Gemini 2.5 Pro - Premium

• Best for: Complex analysis, reasoning
• Speed: Medium
• Cost: High
• Context: 1M tokens
• Vision: Excellent

Gemini 2.5 Flash - Efficient

• Best for: Most stages, good balance
• Speed: Fast
• Cost: Medium
• Context: 1M tokens
• Vision: Good

Gemini 2.5 Flash-Lite - Budget

• Best for: Quick filtering, high volume
• Speed: Very fast
• Cost: Very low
• Context: 1M tokens
• Vision: Basic

Context window

All models have sufficient context for Aparture's needs. Paper abstracts are typically <500 tokens, and PDFs rarely exceed 50K tokens.

Recommended Configurations

Budget Configuration

Goal: Minimize costs while maintaining acceptable quality

Quick Filter: Claude Haiku 4.5
Abstract Scoring: Gemini Flash
PDF Analysis: Claude Sonnet 4.5
NotebookLM: Claude Opus 4.1 (fixed)

Daily cost (30 papers, 10 PDFs):

• Quick Filter: ~$0.05
• Abstract Scoring: ~$0.20
• PDF Analysis: ~$1.50
• NotebookLM: ~$0.30
• Total: ~$2.05/day

Pros:

• Very affordable for daily use
• Reasonable quality
• Fast processing

Cons:

• Lower scoring accuracy
• Less detailed PDF analyses
• May miss subtle relevance

Balanced Configuration (Recommended)

Goal: Best quality-to-cost ratio

Quick Filter: Claude Haiku 4.5
Abstract Scoring: Claude Sonnet 4.5
PDF Analysis: Claude Opus 4.1
NotebookLM: Claude Opus 4.1 (fixed)

Daily cost (30 papers, 10 PDFs):

• Quick Filter: ~$0.05
• Abstract Scoring: ~$0.60
• PDF Analysis: ~$3.00
• NotebookLM: ~$0.30
• Total: ~$3.95/day

Pros:

• Excellent scoring accuracy
• High-quality PDF analyses
• Reliable relevance detection

Cons:

• Moderate cost
• Slower than budget config

Premium Configuration

Goal: Maximum quality, cost secondary

Quick Filter: Claude Sonnet 4.5
Abstract Scoring: Claude Opus 4.1
PDF Analysis: Claude Opus 4.1
NotebookLM: Claude Opus 4.1 (fixed)

Daily cost (30 papers, 10 PDFs):

• Quick Filter: ~$0.15
• Abstract Scoring: ~$2.00
• PDF Analysis: ~$3.00
• NotebookLM: ~$0.30
• Total: ~$5.45/day

Pros:

• Best possible quality
• Nuanced understanding
• Detailed justifications

Cons:

• Expensive for daily use
• Slower processing
• Diminishing returns on some stages

Speed-Optimized Configuration

Goal: Fastest possible analysis

Quick Filter: Gemini Flash-Lite
Abstract Scoring: Gemini Flash
PDF Analysis: GPT-5 Mini
NotebookLM: Claude Opus 4.1 (fixed)

Daily cost (30 papers, 10 PDFs):

• Quick Filter: ~$0.02
• Abstract Scoring: ~$0.15
• PDF Analysis: ~$1.80
• NotebookLM: ~$0.30
• Total: ~$2.27/day

Analysis time: ~50% faster than balanced

Pros:

• Very fast completion
• Low cost
• Good enough for many use cases

Cons:

• Lower quality scoring
• Less detailed analyses

Stage-Specific Recommendations

Quick Filter (Stage 1)

Purpose: Fast YES/NO/MAYBE classification

Best models:

1. Claude Haiku 4.5 - Best balance of speed and accuracy
2. Gemini Flash-Lite - Fastest, cheapest
3. GPT-5 Nano - Good alternative

Don't use:

• ❌ Opus/GPT-5 - Overkill for simple task
• ❌ Sonnet/GPT-5 Mini - Unnecessary cost

Why it matters:

• Processes all papers (high volume)
• Simple binary decision
• Cost adds up quickly

Abstract Scoring (Stage 2)

Purpose: Detailed relevance scoring (0-10)

Best models:

1. Claude Sonnet 4.5 - Best balance (recommended)
2. Claude Opus 4.1 - Highest quality
3. GPT-5 - Excellent alternative

Good alternatives:

• Gemini Flash - Budget option
• GPT-5 Mini - Fast and decent

Don't use:

• ❌ Haiku/Nano/Flash-Lite - Too simple for nuanced scoring

Why it matters:

• Core filtering step
• Determines which papers get deep analysis
• Quality directly impacts results

PDF Analysis (Stage 3)

Purpose: Deep analysis of full papers

Best models:

1. Claude Opus 4.1 - Best vision + reasoning (recommended)
2. GPT-5 - Excellent alternative
3. Gemini 2.5 Pro - Good for long papers

Acceptable alternatives:

• Claude Sonnet 4.5 - Budget option
• GPT-5 Mini - Fast budget option

Don't use:

• ❌ Haiku/Nano/Flash-Lite - Insufficient for deep analysis

Why it matters:

• Most expensive stage (long context)
• Creates value-add summaries
• Vision capability important for figures/equations

Post-Processing (Optional)

Purpose: Re-score papers for consistency

Best models:

• Same as Abstract Scoring
• Recommended: Claude Sonnet 4.5 or Opus 4.1

Note: This stage compares papers side-by-side, so use a high-quality model.

NotebookLM Document Generation

Fixed: Claude Opus 4.1

This stage requires the highest quality model to create well-structured, podcast-optimized documents. The model is not configurable.

Cost Analysis

Cost Breakdown (Typical Daily Run)

Assumptions:

• 30 papers fetched
• Quick Filter enabled (20 pass to scoring)
• 10 papers selected for PDF analysis
• Average PDF size: 2.5 MB (~40K tokens)

Budget Config: $2.05/day

Quick Filter (Haiku):     $0.05
Abstract Scoring (Flash): $0.20
PDF Analysis (Sonnet):    $1.50
NotebookLM (Opus):        $0.30

Balanced Config: $3.95/day

Quick Filter (Haiku):     $0.05
Abstract Scoring (Sonnet):$0.60
PDF Analysis (Opus):      $3.00
NotebookLM (Opus):        $0.30

Premium Config: $5.45/day

Quick Filter (Sonnet):    $0.15
Abstract Scoring (Opus):  $2.00
PDF Analysis (Opus):      $3.00
NotebookLM (Opus):        $0.30

Annual Costs

Configuration	Daily	Monthly	Annual
Budget	$2.05	$61.50	$738
Balanced	$3.95	$118.50	$1,422
Premium	$5.45	$163.50	$1,962

Quick Filter saves money

Enabling Quick Filter can reduce abstract scoring costs by 40-60% by pre-filtering irrelevant papers.

Quality Considerations

Scoring Consistency

Problem: Different batches may get different scores

Impact: Paper with score 7 in batch 1 might score 6 in batch 2

Solutions:

1. Enable Post-Processing - Re-scores papers for consistency
2. Use higher-quality models - Opus/GPT-5 more consistent than Haiku/Nano
3. Smaller batches - More API calls but better consistency

Model rankings (consistency):

1. Claude Opus 4.1 (most consistent)
2. GPT-5
3. Claude Sonnet 4.5
4. Gemini Pro
5. Others (less consistent)

Justification Quality

What makes good justifications:

• Specific references to paper content
• Clear reasoning for score
• Comparison to research criteria
• Mentions of key contributions

Model rankings (justification quality):

1. Claude Opus 4.1 (most detailed)
2. GPT-5
3. Claude Sonnet 4.5
4. Gemini Pro
5. Others (more generic)

Vision Capability

Importance: Papers with complex figures, equations, diagrams

Vision rankings:

1. Claude Opus 4.1 (excellent)
2. GPT-5 (excellent)
3. Gemini Pro (very good)
4. Claude Sonnet 4.5 (good)
5. Others (basic)

When vision matters:

• Machine learning (architecture diagrams)
• Mathematics (equations, proofs)
• Experimental sciences (plots, data)

When vision matters less:

• Theory papers (text-heavy)
• Reviews (narrative)
• Position papers (conceptual)

Switching Models

Testing Different Models

Recommended approach:

1. Run minimal test with different configurations
2. Compare score distributions and justifications
3. Check cost vs. quality trade-off
4. Select based on your priorities

Example test:

# Test 1: Budget config
npm run test:minimal
mv reports/minimal.md reports/test_budget.md

# Test 2: Balanced config (change settings in UI)
npm run test:minimal
mv reports/minimal.md reports/test_balanced.md

# Compare results
diff reports/test_budget.md reports/test_balanced.md

When to Change Models

Upgrade to better models:

• Too many false negatives (missing relevant papers)
• Scores seem inconsistent
• Justifications too generic
• PDF analyses lacking depth

Downgrade to cheaper models:

• Costs too high for daily use
• Current quality exceeds needs
• Fast turnaround more important
• High paper volume with good filtering

API Key Requirements

Minimum (single provider):

• Anthropic (Claude) API key only
• Use Claude models for all stages

Recommended (two providers):

• Anthropic + OpenAI or Google
• Flexibility to optimize each stage

Maximum (all three):

• Anthropic + OpenAI + Google
• Full flexibility, redundancy if one is down

Model availability

Aparture gracefully handles missing API keys by showing only available models in dropdowns.

Provider Comparison

Anthropic (Claude)

Strengths:

• Best vision capability
• Excellent reasoning
• Strong consistency
• Great for research tasks

Weaknesses:

• Slower than competitors
• Higher cost (Opus)
• Rate limits on free tier

Best for: PDF analysis, high-quality scoring

OpenAI (ChatGPT)

Strengths:

• Fast response times
• Good quality across models
• Reliable API
• Strong vision (GPT-5)

Weaknesses:

• Less detailed than Claude Opus
• Higher cost than Gemini
• Stricter content policies

Best for: Balanced speed and quality

Google (Gemini)

Strengths:

• Largest context window (1M)
• Very fast (Flash models)
• Lowest cost (Flash-Lite)
• Good for high volume

Weaknesses:

• Less consistent than Claude/GPT
• Justifications more generic
• Newer, less proven

Best for: Budget configurations, quick filtering

Next Steps

• Understand multi-stage analysis →
• Test different configurations →
• Learn about CLI automation →