Descriptions:
Matthew Berman argues that the most effective AI users have moved beyond prompt engineering and are now focused on token-level optimization — selecting the right model for each sub-task rather than routing everything through a single expensive frontier model. The video opens by explaining why raw per-token pricing is a misleading metric without accounting for how many tokens a given model consumes to complete a task, a concept Berman calls ‘intelligence density.’
Using Artificial Analysis benchmark data, Berman compares GPT-5.6 Soul ($5/$30 per million input/output tokens), Claude Fable (higher still), and Kimmy K3 ($3/$15 per million tokens). Despite Kimmy K3’s lower headline price, the data shows it requires roughly twice as many tokens to complete the same tasks as GPT-5.6, making real-world cost-per-task nearly identical at around $95–$104. Claude Fable comes in at $2.75 per task completed versus GPT-5.6’s $1.04, making it substantially more expensive on an output-adjusted basis.
The actionable recommendation is a three-model workflow: use a frontier model like Claude Fable for high-level planning (where reasoning quality matters most and token volume is low), a cheaper fast model like Grok 4.5 or Cursor’s Composer for code execution (where output volume is highest and quality requirements are lower), and a second frontier model like GPT-5.6 for final review against the original spec. Berman also references a Graphile study on how different AI coding tools cluster their error types, recommending it for teams shipping AI-generated code.
📺 Source: Matthew Berman · Published July 23, 2026
🏷️ Format: Workflow Case Study
