OpenAI o3-Flex + Cline & Roo: BYE Gemini! This is THE MOST COST-EFFECTIVE AI Coding SETUP YET!

Key Concepts:

• GPT-3.5 Turbo (03) price reduction
• Flex processing for 03 and O4 Mini
• Requesty as a tool for managing and utilizing Flex processing
• MCP (Managed Context Provider) servers for enhanced AI coding
• Firecrawl MCP server for web searching and scraping
• Klein Rode and Kilo Code as platforms for using 03 and MCP servers

1. GPT-3.5 Turbo (03) Price Reduction and Comparison

• OpenAI has reduced the price of 03 significantly, making it 80% cheaper.
• New pricing: $2 per million input tokens and $8 per million output tokens.
• Previous pricing: $10 per million input tokens and $40 per million output tokens.
• This price drop makes 03 competitive with models like Gemini 2.5 Pro and Claude Sonnet.
• The speaker considers 03 superior to Gemini 2.5 Pro due to its reasoning capabilities and tool calling abilities.
• Claude Opus 4 is still more expensive than 03.

2. Flex Processing for 03 and O4 Mini

• OpenAI introduced Flex processing for 03 and O4 Mini, offering even cheaper rates for slower responses and occasional unavailability.
• Flex pricing: $2 per million input tokens and $4 per million output tokens.
• Suitable for non-production, lower priority, or asynchronous tasks like model evaluation and data enrichment.
• Not recommended for real-time responses or guaranteed uptime.

3. Comparison of Pricing with Flex and Standard Tiers

• Gemini 2.5 Pro: Input rates range from $1.25 to $2.50 per million tokens, and output is $10 to $15.
• With Flex, 03 is cheaper on output and competitive on input compared to Gemini 2.5 Pro.
• Even without Flex, 03 is cheaper than Gemini 2.5 Pro while offering better capabilities.

4. Requesty: A Tool for Managing Flex Processing

• Requesty is presented as a tool similar to OpenRouter but with more features.
• It allows easy use of 03 Flex by setting the service tier to "flex" in the model name (e.g., "03;flex").
• Requesty handles routing, load balancing, caching, monitoring, and cost control.
• The speaker uses Requesty for all of his projects.

5. Integration with Klein Rode and Kilo Code

• Klein Rode: Users can set up a new profile, select Requesty as the provider, and choose 03 or 03 Flex models. Users can also adjust the reasoning effort levels between high and low.
• Kilo Code: Offers free $20 credits and no OpenRouter markup fees for using 03 and O4 Mini.

6. MCP (Managed Context Provider) Servers

• MCP servers enhance AI coding by providing additional context and information.
• The speaker uses Context 7 (for documentation) and Firecrawl MCP server.
• Firecrawl MCP server: Allows AI coder to gather knowledge from the internet by crawling and scraping URLs and web pages.
• Firecrawl's new search endpoint allows performing web searches and scraping search results in one API call.
• Firecrawl has a free plan with 500 credits and paid plans starting at $16.

7. Firecrawl's Web Scraping Capabilities

• Firecrawl can gather clean data from all accessible subpages, even without a site map.
• It can parse and output content from web-hosted PDFs.
• The search endpoint allows giving a search query and performing web searches, optionally scraping search results in one operation.
• It can search the web and get LLM-ready page content for each result, making it one API call to discover pages and scrape their full content.

8. Example Application: Building a Minecraft Replica

• The speaker demonstrates using 03 with Firecrawl MCP server to build a replica of Minecraft using HTML, CSS, and JS.
• The AI coder successfully creates a functional replica, although with some issues.
• The speaker notes that Gemini also has similar issues, but 03 is cheaper.

9. Conclusion

• The price drop in 03 and the introduction of Flex processing make advanced reasoning more accessible and affordable.
• Requesty simplifies the management and utilization of Flex processing.
• MCP servers, especially Firecrawl, enhance AI coding capabilities by providing access to web-based knowledge.
• The speaker finds 03 to be a compelling alternative to Gemini, especially for smaller codebases, due to its lower cost and tool calling abilities. The only thing that the speaker misses is the 1 million token context window.