mdcontext 0.1.0 → 0.2.0

package/research/llm-summarization/openai-2026.md

@@ -0,0 +1,473 @@
+# OpenAI Models & Pricing Research - 2026
+
+**Research Date:** January 26, 2026
+**Focus:** Code understanding, summarization, and analysis capabilities
+
+---
+
+## Executive Summary
+
+OpenAI has significantly expanded its model lineup in 2025-2026, with major releases including GPT-5 (August 2025), GPT-5.2 (January 2026), and specialized coding models. The company has also reduced pricing substantially, notably an 80% price drop for o3 reasoning models. For code summarization and understanding tasks, GPT-5.2-Codex represents the state-of-the-art, while GPT-4o-mini and GPT-5-mini offer cost-effective alternatives.
+
+---
+
+## Current Model Lineup (2026)
+
+### 1. GPT-5 Series - Flagship Models
+
+#### GPT-5.2 (Latest - January 2026)
+**Best for:** Coding, agentic tasks, document summarization, complex reasoning
+
+- **Pricing:** $1.75/1M input tokens, $14.00/1M output tokens
+- **Context Window:** Up to 400K tokens
+- **Knowledge Cutoff:** August 2025
+- **Key Features:**
+  - Designed for deeper work and complex tasks
+  - Excels at coding, summarizing long documents, file Q&A
+  - Step-by-step math and logic reasoning
+  - Flagship model for coding and agentic tasks across industries
+  - Stronger multimodal performance than predecessors
+
+#### GPT-5.2-Pro
+- **Pricing:** $21/1M input tokens, $168/1M output tokens
+- **Context Window:** Up to 400K tokens
+- **Key Features:**
+  - Smartest and most trustworthy option for difficult questions
+  - Fewer major errors in testing
+  - Stronger performance in complex domains like programming
+  - Supports extended reasoning for top-quality execution
+
+#### GPT-5 (Standard)
+- **Pricing:** $1.25/1M input tokens, $10.00/1M output tokens
+- **Context Window:** Up to 400K tokens
+- **Release:** August 7, 2025
+- **Key Features:**
+  - 74.9% accuracy on SWE-bench Verified
+  - 88% on Aider Polyglot benchmarks
+  - State-of-the-art for real-world coding tasks
+
+#### GPT-5 Variants
+- **GPT-5 Pro Global:** $15/1M input, $120/1M output
+- **GPT-5-mini:** $0.25/1M input, $2.00/1M output
+- **GPT-5-nano:** $0.05/1M input, $0.40/1M output
+
+### 2. GPT-5.2-Codex - Specialized Coding Model
+
+**Best for:** Complex software engineering, large refactors, code migrations, agentic coding
+
+- **Pricing:** $1.25/1M input tokens, $10.00/1M output tokens (estimated based on GPT-5 Codex)
+- **Note:** Specific GPT-5.2-Codex API pricing not yet publicly finalized
+- **Context Window:** Up to 400K tokens
+- **Key Features:**
+  - Most advanced agentic coding model for complex, real-world software engineering
+  - Improvements in context compaction for long-horizon work
+  - Stronger performance on large code changes (refactors, migrations)
+  - Improved Windows environment support
+  - **Significantly stronger cybersecurity capabilities**
+  - State-of-the-art on SWE-Bench Pro and Terminal-Bench 2.0
+  - Better at working in large repositories over extended sessions
+  - Can reliably complete complex tasks without losing context
+
+**Availability:**
+- Currently available in ChatGPT Codex surfaces for paid users
+- API access rolling out in coming weeks (as of January 2026)
+
+**ChatGPT Subscription Access:**
+- **Plus Plan:** Up to 160 messages with GPT-5.2 every 3 hours
+- **Pro Plan:** $200/month with unlimited messages (subject to fair use)
+
+### 3. GPT-4 Series - Previous Generation
+
+#### GPT-4.1
+- **Pricing:** $2.00/1M input tokens, $8.00/1M output tokens
+- **Context Window:** Up to 1.0M tokens (largest context window)
+- **Release:** April 14, 2025
+
+#### GPT-4.1-mini
+- **Pricing:** $0.40/1M input, $1.60/1M output tokens
+- **Context Window:** 128K tokens
+
+#### GPT-4o
+- **Pricing:** $2.50/1M input tokens, $10.00/1M output tokens
+- **Context Window:** Up to 128K tokens
+- **Key Features:**
+  - Text generation, summarization, knowledge-based Q&A
+  - Reasoning, complex math, coding capabilities
+  - Strong multimodal performance
+
+#### GPT-4o-mini (Budget Option)
+- **Pricing:** $0.150/1M input tokens, $0.600/1M output tokens
+- **Context Window:** Up to 128K tokens
+- **Key Features:**
+  - Most cost-effective option
+  - Suitable for high-volume applications
+  - Good for straightforward code analysis tasks
+
+### 4. Reasoning Models (o-series)
+
+#### o3 (Latest Reasoning Model - 2026)
+**Best for:** Complex reasoning, mathematical proofs, advanced problem-solving
+
+- **Pricing:** $2.00/1M input tokens, $8.00/1M output tokens
+- **Price Reduction:** 80% cheaper than previous rates
+- **Cached Input Discount:** Additional $0.50/1M tokens discount
+- **Flex Mode:** $5/1M input, $20/1M output (synchronous processing)
+- **Key Features:**
+  - Extended reasoning capabilities
+  - Suitable for complex logic and mathematical tasks
+
+#### o3-pro
+- **Availability:** ChatGPT Pro ($200/month) and Team users only
+- **Pricing:** Significantly higher API costs (reflects extended compute time)
+- **Note:** Best for most demanding reasoning tasks
+
+#### o4-mini
+- **Pricing:** Not specified, positioned as cost-effective
+- **Key Features:**
+  - Remarkable reasoning performance at fraction of cost
+  - Cost-effective for high-volume applications
+
+#### o1 (Previous Generation - Historical Reference)
+- **o1-preview:** $15/1M input, $60/1M output tokens
+- **o1-mini:** $1.10/1M input, $4.40/1M output tokens
+
+---
+
+## Batch API Pricing
+
+For non-urgent workloads processed within 24 hours:
+
+- **50% discount on all models**
+- Example: GPT-5 drops to $0.625/$5.00 per million tokens (vs $1.25/$10.00 standard)
+- Ideal for bulk code analysis and summarization tasks
+
+---
+
+## Rate Limits & API Tiers
+
+### Rate Limit Metrics
+OpenAI measures limits in five ways:
+- **RPM:** Requests per minute
+- **RPD:** Requests per day
+- **TPM:** Tokens per minute
+- **TPD:** Tokens per day
+- **IPM:** Images per minute
+
+### 2026 Rate Limit Updates
+
+**Significant Increases:**
+- GPT-5 Tier 1: Increased from ~30,000 TPM to **500,000 TPM**
+- GPT-5 Tier 1: Approximately **1,000 RPM**
+- Higher tiers provide substantially more capacity
+
+### Usage Tiers
+- Users automatically graduate to higher tiers as API spending increases
+- Most models available on Tiers 1-5 (subject to organization verification)
+- Check current limits at: `platform.openai.com/settings/organization/limits`
+
+### Best Practices for Rate Limits
+
+1. **Error Handling:**
+   - Watch for `429: Too Many Requests` errors
+   - Implement `RateLimitError` exception handling
+
+2. **Exponential Backoff:**
+   - Use retries with randomized, increasing wait periods
+   - Most effective strategy for handling rate limits
+
+3. **Request Throttling:**
+   - Reference OpenAI's `api_request_parallel_processor.py` script
+   - Batch multiple small requests together
+
+4. **Optimization:**
+   - Set `max_tokens` parameter realistically
+   - Cache responses for frequently asked questions
+   - Use Batch API for non-urgent workloads
+
+---
+
+## Recommendations for Code Summarization Use Cases
+
+### Best Model Choices by Use Case
+
+#### 1. Production Code Summarization at Scale
+**Recommended:** GPT-4o-mini or GPT-5-mini
+- **Why:** Extremely cost-effective ($0.15/$0.60 or $0.25/$2.00 per million tokens)
+- **Use for:**
+  - Summarizing function docstrings
+  - Generating README files
+  - Inline code comments
+  - High-volume batch processing
+- **Cost at 10M tokens/day:**
+  - GPT-4o-mini: ~$7.50/day
+  - GPT-5-mini: ~$22.50/day
+
+#### 2. Complex Code Understanding & Analysis
+**Recommended:** GPT-5.2 or GPT-5
+- **Why:** Superior reasoning and code comprehension
+- **Use for:**
+  - Architecture documentation
+  - Complex dependency analysis
+  - Security vulnerability assessment
+  - Technical debt analysis
+- **Cost at 1M tokens/day:**
+  - GPT-5.2: ~$15.75/day
+  - GPT-5: ~$11.25/day
+
+#### 3. Agentic Code Refactoring & Large Migrations
+**Recommended:** GPT-5.2-Codex
+- **Why:** Specifically optimized for multi-step coding tasks
+- **Use for:**
+  - Large-scale refactors
+  - Framework migrations
+  - Codebase modernization
+  - Security-critical code changes
+- **Key Advantage:** Maintains context over long sessions
+- **Cost:** ~$11.25/day at 1M tokens (estimated)
+
+#### 4. Deep Reasoning About Code Logic
+**Recommended:** o3
+- **Why:** Best reasoning capabilities, now 80% cheaper
+- **Use for:**
+  - Algorithm correctness verification
+  - Complex bug root cause analysis
+  - Performance optimization reasoning
+  - Mathematical code verification
+- **Cost at 1M tokens/day:** ~$10.00/day
+
+#### 5. Budget-Conscious Batch Processing
+**Recommended:** GPT-4o-mini or GPT-5-mini via Batch API
+- **Why:** 50% discount + already low prices
+- **Use for:**
+  - Overnight documentation generation
+  - Weekly codebase summaries
+  - Non-urgent analysis tasks
+- **Cost at 10M tokens/day:**
+  - GPT-4o-mini Batch: ~$3.75/day
+  - GPT-5-mini Batch: ~$11.25/day
+
+### Cost Comparison Table
+
+| Model | Input ($/1M) | Output ($/1M) | Batch Input | Batch Output | Best For |
+|-------|-------------|---------------|-------------|--------------|----------|
+| GPT-4o-mini | $0.15 | $0.60 | $0.075 | $0.30 | High-volume basic tasks |
+| GPT-5-nano | $0.05 | $0.40 | $0.025 | $0.20 | Ultra-low-cost simple summaries |
+| GPT-5-mini | $0.25 | $2.00 | $0.125 | $1.00 | Balanced cost/quality |
+| GPT-5 | $1.25 | $10.00 | $0.625 | $5.00 | Standard complex tasks |
+| GPT-5.2 | $1.75 | $14.00 | $0.875 | $7.00 | Latest flagship |
+| GPT-5.2-Codex | ~$1.25 | ~$10.00 | ~$0.625 | ~$5.00 | Agentic coding |
+| GPT-4.1 | $2.00 | $8.00 | $1.00 | $4.00 | Large context (1M) |
+| o3 | $2.00 | $8.00 | $1.00 | $4.00 | Deep reasoning |
+
+---
+
+## Special Considerations for Code Analysis
+
+### Context Window Strategy
+
+1. **Small Files (<10K tokens):** Any model works
+2. **Medium Projects (10K-128K tokens):** GPT-4o, GPT-5.2, GPT-5
+3. **Large Codebases (128K-400K tokens):** GPT-5.2, GPT-5, GPT-4.1
+4. **Massive Projects (>400K tokens):** GPT-4.1 (1M context) or chunking strategy
+
+### Token Optimization Tips
+
+- **Minify Code:** Remove comments/whitespace before sending (if not needed for analysis)
+- **Selective Inclusion:** Only send relevant files
+- **Hierarchical Summarization:** Summarize small chunks, then summarize summaries
+- **Caching:** OpenAI offers prompt caching - reuse common context
+
+### Security Considerations
+
+- **Sensitive Code:** Ensure compliance with data handling policies
+- **GPT-5.2-Codex:** Specifically enhanced for cybersecurity use cases
+- **Private Deployment:** Consider Azure OpenAI for enterprise security requirements
+
+---
+
+## API Capabilities Relevant to Code Analysis
+
+### Supported Features
+
+1. **Function Calling:**
+   - All GPT-4 and GPT-5 models support structured function calls
+   - Useful for integrating code analysis into tools
+
+2. **JSON Mode:**
+   - Guaranteed JSON responses
+   - Perfect for structured code summaries
+
+3. **Vision (Multimodal):**
+   - GPT-4o and GPT-5 series support image inputs
+   - Analyze architecture diagrams, UI screenshots, flowcharts
+
+4. **Fine-tuning:**
+   - Available for GPT-4o and GPT-4o-mini
+   - Can specialize for specific code summarization formats
+
+5. **Embeddings (Separate Pricing):**
+   - text-embedding-3-large: $0.13/1M tokens
+   - text-embedding-3-small: $0.02/1M tokens
+   - Useful for semantic code search
+
+### Developer Experience
+
+- **Streaming:** All models support streaming responses
+- **Retry Logic:** Built-in exponential backoff recommended
+- **SDKs:** Official Python, Node.js, and community libraries
+- **Monitoring:** Usage dashboard at platform.openai.com
+
+---
+
+## Pricing Trends & Future Outlook
+
+### Recent Changes
+- **80% price reduction** for o3 reasoning model (2026)
+- **Significant rate limit increases** (GPT-5 TPM: 30K → 500K)
+- Introduction of ultra-low-cost nano variants
+
+### Competitive Landscape
+- OpenAI faces competition from:
+  - Anthropic Claude (Opus 4.5, Sonnet 4.5)
+  - Google Gemini 3 Pro
+  - DeepSeek R1 (reasoning model)
+- Price reductions driven by market competition
+
+### What to Expect
+- Continued price decreases as competition intensifies
+- More specialized models for specific tasks
+- Larger context windows becoming standard
+- Enhanced agentic capabilities across model lineup
+
+---
+
+## Quick Decision Matrix
+
+### Choose GPT-4o-mini if:
+- Budget is primary constraint
+- Tasks are straightforward (basic summarization)
+- Processing millions of tokens daily
+- Quality requirements are moderate
+
+### Choose GPT-5-mini if:
+- Need better quality than GPT-4o-mini
+- Still cost-sensitive
+- Want newer model capabilities
+- Willing to pay ~3-4x more for better results
+
+### Choose GPT-5.2 if:
+- Need best general-purpose performance
+- Code summarization requires deep understanding
+- Working with complex codebases
+- Can justify ~10x cost vs mini models
+
+### Choose GPT-5.2-Codex if:
+- Performing agentic coding tasks
+- Large refactors or migrations
+- Security-critical code work
+- Need multi-step reasoning with code
+
+### Choose o3 if:
+- Need deep logical reasoning about code
+- Algorithm verification required
+- Mathematical correctness matters
+- Budget allows reasoning model costs
+
+### Choose GPT-4.1 if:
+- Need 1M token context window
+- Processing entire large repositories
+- Context window more important than latest features
+
+---
+
+## Sample Cost Scenarios
+
+### Scenario 1: Daily Codebase Documentation
+- **Task:** Summarize 100 files/day, avg 500 tokens each = 50K tokens
+- **Model:** GPT-5-mini
+- **Cost:** (50K × $0.25/1M) + (50K × $2/1M) = $0.0125 + $0.10 = **$0.11/day**
+- **Monthly:** ~$3.30/month
+
+### Scenario 2: Weekly Architecture Analysis
+- **Task:** Deep analysis of 1M tokens once weekly
+- **Model:** GPT-5.2
+- **Cost:** ($1.75 + $14) = **$15.75/week**
+- **Monthly:** ~$63/month
+
+### Scenario 3: Continuous Code Review Pipeline
+- **Task:** 10M tokens/day, basic summaries
+- **Model:** GPT-4o-mini (Batch API)
+- **Cost:** (10M × $0.075/1M) + (10M × $0.30/1M) = $0.75 + $3.00 = **$3.75/day**
+- **Monthly:** ~$112.50/month
+
+### Scenario 4: Enterprise Refactoring Project
+- **Task:** 500K tokens/day for 30 days (major migration)
+- **Model:** GPT-5.2-Codex
+- **Cost:** (500K × $1.25/1M) + (500K × $10/1M) = $0.625 + $5.00 = **$5.625/day**
+- **Monthly:** ~$168.75/month
+
+---
+
+## Resources & Links
+
+### Official Documentation
+- [OpenAI API Pricing](https://openai.com/api/pricing/)
+- [Platform Pricing Docs](https://platform.openai.com/docs/pricing)
+- [Rate Limits Guide](https://platform.openai.com/docs/guides/rate-limits)
+- [Models Documentation](https://platform.openai.com/docs/models/)
+
+### Model Announcements
+- [Introducing GPT-5](https://openai.com/index/introducing-gpt-5/)
+- [Introducing GPT-5.2](https://openai.com/index/introducing-gpt-5-2/)
+- [Introducing GPT-5.2-Codex](https://openai.com/index/introducing-gpt-5-2-codex/)
+- [Introducing o3 and o4-mini](https://openai.com/index/introducing-o3-and-o4-mini/)
+
+### Developer Resources
+- [GPT-5.2 Prompting Guide](https://cookbook.openai.com/examples/gpt-5/gpt-5-2_prompting_guide)
+- [Rate Limit Handling](https://cookbook.openai.com/examples/how_to_handle_rate_limits)
+- [OpenAI Academy](https://academy.openai.com/)
+
+### Third-Party Resources
+- [Price Per Token Calculator](https://pricepertoken.com/pricing-page/provider/openai)
+- [Finout OpenAI Pricing Guide](https://www.finout.io/blog/openai-pricing-in-2026)
+- [Cost Calculator](https://costgoat.com/pricing/openai-api)
+
+---
+
+## Conclusion & Recommendations
+
+For **mdcontext** code summarization use cases:
+
+1. **Start with GPT-5-mini** for basic file summarization
+   - Excellent cost/quality balance
+   - ~$0.25-$2.00 per million tokens
+   - Suitable for 80%+ of summarization tasks
+
+2. **Upgrade to GPT-5.2** for complex analysis
+   - When understanding architecture matters
+   - When context relationships are critical
+   - When quality justifies ~7x cost increase
+
+3. **Use Batch API** whenever possible
+   - 50% cost savings for non-urgent tasks
+   - Perfect for overnight documentation generation
+
+4. **Reserve GPT-5.2-Codex** for agentic tasks
+   - Large-scale refactoring projects
+   - Security-critical analysis
+   - Multi-step code transformations
+
+5. **Monitor costs closely**
+   - Set up billing alerts
+   - Track token usage per operation
+   - Optimize prompt engineering to minimize tokens
+
+6. **Consider hybrid approach**
+   - Use mini models for initial pass
+   - Use flagship models for complex files flagged by mini models
+   - Can reduce costs 60-80% vs. using flagship for everything
+
+---
+
+**Last Updated:** January 26, 2026
+**Next Review:** April 2026 (quarterly review recommended)