AI summary by TwinMind
OpenAI is back on top
Theo - t3․gg
30m 17s
OpenAI o3 Pro Launch and Pricing Changes
- OpenAI released o3 Pro model today
- o3 standard model price cut by 80%, now cheaper than GPT-4.0, Claude models, and Gemini 2.5 Pro
- o3 Pro is 87% cheaper than previous o1 Pro model
- o3 Pro takes ~4 minutes reasoning time on simple prompts
- Available for Pro users ($200/month) in ChatGPT and API
- Plus users ($20/month) get access to o3 as part of core plan
- T3 Chat offers o3 access for $8/month (promo code "o3-PLS" for $1/month for new subscribers)
Model Tier Structure and Performance
- Three-tier system: Mini, Main (standard), Pro
- Mini: Smaller, faster, cheaper
- Main: Balanced performance and cost
- Pro: Smartest but slowest and most expensive
- o3 Pro has 64% win rate vs o3 on human evaluations
- o4 Mini performs surprisingly well, often matching higher-tier models
- Benchmarks show steady improvement: o1 mini < o3 mini < o4 mini < o3 < o3 Pro
- Gemini 2.5 Pro still beats o3 Pro on some math benchmarks
User Experience and Model Characteristics
- Pro models create poor UX due to slow response times
- ChatGPT site bugs with o1 Pro due to slow generation
- Reasoning models feel more clinical vs GPT-4.0's conversational tone
- GPT-4.0 remains default for general users due to immediate response without reasoning
- o4 Mini becoming preferred default for many power users
Prompting Best Practices for Reasoning Models
- Minimize system prompts to focus context on actual problem
- Anatomy of good prompts:
- Start with clear goal
- Specify return format
- Include warnings/important details
- Dump all necessary context
- Provide capabilities (MCPs, tool calls)
Tool Integration and Capabilities
- ChatGPT now has "Choose Tool" button for selecting available tools
- o3 Pro better at discerning environment and available tools
- Better at asking questions about outside world vs hallucinating
- Deep Research feature: 250 reports available for $200 subscription (~$1 per report)
- OpenAI embracing MCP (Model Context Protocol)
Cost Analysis and Context Management
- o3 Pro pricing: $20 per million input tokens, $80 per million output tokens
- 150K token context example costs $3 just for input
- Human-in-the-loop interactions expensive due to context re-ingestion
- Caching helps but limited to ~30 minutes
- Tool calls preferred over human interaction to avoid context rebuilding
- Input tokens major expense for applications like Cursor
Competitive Landscape Analysis
- Claude 4 Opus now looks expensive and less capable
- Anthropic Sonnet models now most expensive on comparison charts
- o3's new pricing directly competes with Gemini 2.5 Pro
- Artificial Analysis benchmark costs:
- Gemini 2.5 Pro: $852 (most expensive due to reasoning tokens)
- Claude 4 Opus: ~$400
- 2.5 Flash Reasoning: $319 vs 2.0 Flash: $3
- o3 Mini: Cheaper than 2.5 Flash Reasoning despite higher per-token cost
Technical Specifications and Limitations
- o3 context window: 200K input, 100K output (smaller than competitors)
- o4.1 and Google models: 1M context window
- Knowledge cutoff: May 31, 2024 (over a year old)
- Temporary chats disabled for o3 Pro due to streaming issues
Use Case Recommendations
- GPT-4.0: General chat, immediate responses, voice interactions
- o4 Mini: Fast reasoning tasks, good balance of speed/intelligence
- o3: Deep analysis with adequate context
- o3 Pro: Complex research, large data analysis, report generation
- Not recommended: o3 Pro for simple coding tasks or chat
- Better for: Analyzing large codebases, complex problem-solving with tools
Market Position Assessment
- OpenAI maintaining lead despite Google's brief competitiveness
- Consistent APIs and reliable tool protocols advantage over competitors
- Price wars driving costs down rapidly
- Power users benefit most from model selection optimization
- Regular users should stick with GPT-4.0 default
