How to Cut AI Costs With a Multi-Model Workflow

TL;DR

Use a frontier model for planning, a cheaper fast model for execution, and a different frontier model for final review. This allocation places expensive intelligence where reasoning matters most, shifts output-heavy code generation to a lower-cost model, and uses cross-model review to catch mistakes, reducing the example workflow cost from $46.50 with GPT 5.6 alone to $25.55.

Transcript

The best people in the world at using AI don't care as much about prompt engineering anymore. What they do care about is which tokens they're using for which task. Not all tokens are made equal. And I'm going to explain exactly how to increase the quality of your AI usage while also decreasing the cost. This is what the absolute frontier of AI user... Read More

Key Insights

• A token is the basic unit processed by a large language model, and it may represent a complete word or only part of one. Models generate responses by predicting the next token from the preceding sequence.
• Token price is not the same as task price because models can consume different numbers of tokens to solve identical problems. A model sold at half the per-token price offers no task-level savings if it needs twice as many tokens.
• Intelligence density is the amount of useful reasoning produced per token. The speaker considers OpenAI and ChatGPT strongest on this measure because their models reportedly require fewer tokens than some competitors to solve the same problem.
• Task completion cost is a more useful comparison than benchmark scores or headline API rates. The cited Artificial Analysis results place Kimmy K3 at $0.95 and GPT 5.6 Soul at $1.04 per completed task, despite their different token prices.
• Different models make different categories of coding mistakes, so cross-model review can expose defects that self-review misses. The cited Greptile findings show GPT 5.5 finding more bugs when reviewing code authored by Claude than Claude found reviewing its own work.
• The recommended workflow assigns planning to a frontier model, implementation to a cheaper fast model, and final review to a different frontier model. This arrangement reserves costly reasoning for the stages where high-level judgment and independent verification matter most.
• Execution is the most expensive stage because writing code generates far more output tokens, and output tokens carry higher listed prices than input tokens. Planning and review primarily consume input while producing comparatively short specifications or findings.
• Output speed has economic value because a lower token price may still waste time if the model takes longer to solve the task. The speaker argues that a fast execution model can reduce both the direct expense and total completion time.
• More videos from this channel:
  • How does Deepseek V4.1 Flash improve efficiency and open weights perf…
  • How to test VO3 AI video generation capabilities
  • How does Grok 4 perform in industry reactions and tests
• More from this channel: Matthew Berman

Questions & Answers

Q: How can I reduce AI coding costs with multiple models?

Divide the work into planning, execution, and review. Give the planning stage to a frontier model that can inspect the codebase and create a detailed specification. Send that specification to a cheaper, faster model for implementation, where most output tokens are generated. Finally, use a different frontier model to compare the code with the specification and identify bugs or omissions.

Q: Why is token price alone a poor way to compare AI models?

Token price ignores how many tokens a model needs to finish a task. The transcript compares GPT 5.6 Soul at $30 per million output tokens with Kimmy K3 at $15 per million. If GPT 5.6 needs one million tokens while Kimmy requires two million for the same result, both cost $30 despite the apparent twofold price difference.

Q: What is intelligence density in an AI model?

Intelligence density describes how much useful reasoning or task completion a model delivers through each token. A model with higher intelligence density can solve a problem with fewer tokens, making its apparently expensive tokens more valuable. The speaker argues that OpenAI and ChatGPT lead on this measure because competing models may use substantially more tokens to reach the same outcome.

Q: Why should a frontier model handle the planning stage?

Planning requires high-level reasoning about how to solve the problem, including reviewing the codebase, designing the feature, anticipating complications, and producing a complete specification. The speaker treats this as the hardest stage and therefore assigns it to a frontier model. Its output is relatively short, so premium reasoning can be purchased without generating the large output volume required during implementation.

Q: Why can a cheaper AI model handle code execution?

A cheaper model can implement code effectively after a strong model has already determined the architecture and written a detailed plan. Execution involves producing the largest volume of output tokens, which makes it the most expensive phase when frontier pricing is used. Moving this output-heavy work to a capable lower-cost model therefore creates the workflow's largest opportunity for savings.

Q: Why should a different model review AI-generated code?

Different models have different weaknesses and tend to cluster around different bug categories. The Greptile findings cited in the transcript show that GPT 5.5 found more bugs when reviewing a pull request authored by Claude than Claude found through self-review. An independent model can therefore notice missing behavior, semantic mistakes, or other defects that the authoring model is inclined to overlook.

Q: How do input and output tokens affect coding workflow costs?

Planning uses many input tokens because the model must read the codebase, but it produces relatively few output tokens because the result is only a plan. Execution reads that shorter plan and generates much more output as code. Review returns to heavy reading and limited output. Since the cited GPT 5.6 prices are $5 for input and $30 for output per million tokens, execution dominates cost.

Q: How much can the mixed-model workflow save in the example?

The example calculates a cost of $81 when Fable handles the entire task and $46.50 when GPT 5.6 handles every stage. Assigning planning, execution, and review to different models reduces the stated cost to $25.55. The speaker also expects the mixed approach to complete faster because the output-heavy implementation stage uses a faster, more efficient model.

Summary & Key Takeaways

• Tokens are the basic units processed by large language models, but their listed prices do not reveal their practical value. A cheaper model may require more tokens or time to complete the same task, so users should compare total task cost, output speed, and achieved quality rather than price per million tokens alone.

• The recommended coding workflow separates planning, execution, and review. A frontier model studies the codebase and writes a specification, a cheaper and faster model implements that plan, and another frontier model checks the completed work against the specification. This division uses each model where its characteristics provide the greatest practical value.

• Execution is the best stage for cost reduction because it produces many expensive output tokens while requiring less difficult reasoning once a strong plan exists. In the presented example, one-model workflows cost $81 with Fable or $46.50 with GPT 5.6, while the mixed-model workflow costs $25.55 and can finish faster.