Core References
AI Model Reference Guide
Compare AI model families (GPT, Claude, Gemini, DeepSeek), understand temperature settings, and choose models for reasoning, writing, and analysis tasks
Reference page for: AI model selection and configuration. Other pages link here for model-family characteristics and temperature guidance.
Overview: Model Families, Not Versions
Warning
AI models evolve quickly. By the time you read this, newer versions likely exist.
This guide focuses on model families and providers rather than specific version numbers. Version numbers below were current as of March 2026, but check provider websites for the latest.
The capabilities and characteristics stay relatively stable within families, though specific benchmarks change constantly. Family characteristics (OpenAI = reliable, Anthropic = nuanced, Google = context kings) are subjective generalizations from limited testing. Try the ones that fit your work — model personalities vary by task, and your experience may differ.
What matters for research:
- Reasoning depth: Can it handle complex theoretical frameworks?
- Context capacity: How many papers can it process at once?
- Writing quality: Does it produce academic-grade prose?
- Cost efficiency: What's the price-performance ratio?
Which AI Model Should I Use?
Best Choices:
- Claude Opus 4.6 (Frame-breaking, sees beyond local optima)
- GPT-5.4 (Reliable fixes, consistent reasoning)
- Gemini 3.1 Pro (Large context + strong reasoning with rigorous prompts)
Use Cases: Theory building, critical analysis, synthesis, methodology design
Pro Tip: Start with cheaper models for initial exploration, then use premium models when you're confident in your prompts.
Full comparison: AI Model Reference Guide
Model Families by Provider
OpenAI GPT Family
- Current: GPT-5.4 series (Pro, Thinking, Instant)
- Context: 400K-1M tokens
- Strength: Reliable, consistent, excellent writing
- Best for: General research tasks, final writing, systematic coding
Anthropic Claude Family
- Current: Opus 4.6, Sonnet 4.6, Haiku 4.5
- Context: 200K-1M tokens
- Strength: Deep reasoning, nuanced understanding, academic style
- Best for: Theory development, qualitative analysis, complex arguments
Google Gemini Family
- Current: 3.1 Pro, 3.1 Flash, 3.1 Flash-Lite
- Context: 1M tokens
- Strength: Massive context, free tier, fast, cost-effective Flash-Lite
- Best for: Large literature sets, exploratory analysis, volume processing
Free Access Options
Google AI Studio: Your Free Backup Plan
Why This Matters: Google AI Studio provides free daily access to Gemini models, ensuring you can continue your research even if you exceed API credits.
| Model | Rate Limit | Context Window | Best For |
|---|---|---|---|
| Gemini 3.1 Flash | ~500 requests/day | 1M tokens | High-volume literature processing |
| Gemini 3.1 Pro | ~100 requests/day | 1M tokens | Complex theoretical analysis |
| Gemini Embeddings | Generous limits | N/A | Document similarity, semantic search |
Info
Free tier limits change frequently — Google reduced quotas significantly in late 2025 before partially restoring them. Check current rate limits for the latest numbers. Limits reset daily at midnight Pacific Time.
Getting Started with Google AI Studio
- Create Account: Visit aistudio.google.com
- Generate API Key: Go to aistudio.google.com/app/apikey
- Add to Cherry Studio: Settings → API Keys → Add Provider → Google Gemini
- Test Connection: Verify your free daily limits are active
When to Use Free Options
- Literature exploration: Use Gemini Flash for processing large paper collections
- Backup strategy: When API credits are running low
- Experimentation: Try different approaches without cost concerns
- Learning: Understand model differences before using premium credits
Important Notes
- Free access available worldwide (some regions may vary)
- Same high-quality models as paid versions
- Perfect for systematic review tasks requiring large context windows
Understanding AI Configuration Settings
Temperature Settings
Default recommendation: high temperature (1.0–1.5)
Most research tasks benefit from creative, exploratory thinking. We usually start high rather than low.
HIGH Temperature (1.0–1.5): Where we usually start
- Use for: Most research tasks — theory synthesis, exploration, analysis, writing
- Why: Produces more varied perspectives and unexpected connections
- We find these work well: GPT-5.4 (1.0–1.2), DeepSeek (1.0–1.3), Qwen (1.0–1.4)
- Example: "Show me unexpected connections between these frameworks"
MEDIUM Temperature (0.6–0.8): For model-specific quirks
- Use for: Bringing out model personality, exploratory synthesis
- Why: Some models produce their most interesting work at medium temps
- Sweet spots:
- Kimi K2.5 at 0.6–0.7: Developer-recommended, unlocks creative side
- Gemini 3.1 Pro at 0.7–0.8: Quirky insights, interesting angles
- GLM at 0.7–0.9: Creative multilingual connections
- Example: "Give me fresh perspectives I haven't considered"
LOW Temperature (0.1–0.3): Only when needed
- Use for: Deterministic tasks only (citations, final formatting, systematic coding)
- Why: Reduces variation; produces repetitive, conservative outputs
- When: You need the same answer every time
- Example: "Extract author names from this citation — nothing else"
Warning
Our experience: The old advice of "start at 0.1 for precision" tends to flatten the AI's range. Research is creative work, so we usually let the models explore. Your needs might differ, but experiment with higher temperatures before defaulting to low.
For a deep dive into advanced reasoning, see our guide on Mastering Sequential Thinking with MCP.
Strategic Model Usage for Research
Discovery & Exploration
Sample Widely - Build Understanding
- All models: Try everything to understand what works for your research style
- Focus: Finding the right tool for each type of task
- Approach: Small tasks, broad exploration, document preferences
Common Discovery Tasks:
- Initial literature scanning
- Research question refinement
- Methodology exploration
- Theoretical framework discovery
When Cost Considerations Matter
High Token Consumption Scenarios
These are where strategic model selection saves significant money:
-
Automated Literature Processing
- Processing 100+ papers automatically
- Multiple extraction passes
- Strategy: Develop workflow with DeepSeek, deploy with premium if needed
-
Iterative Development
- Refining complex prompts through many cycles
- Testing workflow logic extensively
- Strategy: Iterate with efficient models, finalize with best
-
Large-Scale Analysis
- Systematic coding of hundreds of documents
- Cross-referencing massive literature sets
- Strategy: Prototype small-scale, then choose model based on quality needs
Practical Guidelines
Context Window Strategy
- Under 50 pages: Any model works fine
- 50-200 pages: Use 200K+ models (Claude, GPT-5.4)
- 200+ pages: Use 1M+ models (Gemini 3.1 Pro)
Quality Assurance (Always Important)
- Verify citations: All models can hallucinate references
- Cross-check critical analysis: Use multiple models for important insights
- Use reasoning modes: For complex theoretical questions
- Document model choices: Track what works best for different tasks
Trying it yourself
The tables above tell you what's available. The companion page is the protocol for finding what fits your research — a four-phase experiment (sample, match, calibrate temperatures, build a personal strategy) you run against your own materials: AI Model Discovery Protocol.