TRAIBE ← Library
Notes / IE4 min read

Different tasks,
different costs of intelligence.

A call-center case from a talk at IE: 1,200 calls a day, a different AI model for each step, and every call now reviewed.

Raoni LotarOctober 2026
Three blocks on a desk, linked by arrows: a sound wave, a brain and a report
Three steps, three models, three price points.

At a talk at IE, a team shared a familiar problem. A sales team makes 1,200 calls a day. The quality team listened to a handful of recordings, filled in a form, and nobody looked at the rest: what was promised, what was left out, what went wrong. Only 2% of calls were ever reviewed, and feedback took three weeks.

Their answer was not one big agent. It was an ecosystem of agents, each running on the model that fit its task, balancing the cost of tokens against the quality of the output. The real workflow has more steps; to keep it simple, think of three.

How the agents split the work1,200 calls a day
01 · Transcribe Gemini Flash Turns every call into text. A simple job at high volume, so the cheapest capable model wins. From $0.75 in / $3.75 out
02 · Analyse GPT-6 Astra Reads each transcript and spots possible mistakes. This is where reasoning matters. $10 in / $50 out
03 · Report Claude Opus 5 Writes the summary managers actually read. Clarity and tone matter most here. $5 in / $25 out
What the main models cost · US$ per 1M tokens
ProviderModelTierInputOutput
GoogleGemini 4 ArgonFlagship$2.00*$10.00*
GoogleGemini 3.8 FlashFast$0.75*$3.75*
GoogleGemini 3.5 Flash-LiteLightest$0.30$2.50
OpenAIGPT-6 AstraFlagship$10.00$50.00
OpenAIGPT-6.1 SolMid$2.00$10.00
OpenAIGPT-6 LunaLightest$0.10$0.50
AnthropicClaude Opus 5.5Flagship$4.00$20.00
AnthropicClaude Opus 5Previous flagship$5.00$25.00
AnthropicClaude Sonnet 5.5Mid$2.00$10.00
AnthropicClaude Haiku 5.5Lightest$0.10$0.50
Standard API prices for text, October 2026. Highlighted: the models used in the case. *Launch prices: Gemini 4 Argon doubles on 1 December 2026, Gemini 3.8 Flash on 1 January 2027.
The resultsBefore and after
Before 2% of calls reviewed
Time to review a call25 min
Time to feedbackThree weeks
After 100% of calls reviewed
Time to review a call6 min
Time to feedbackNext day
How to choose a model
  1. 01
    Match the model to the task.Transcribing, sorting and extracting rarely need a frontier model. Judgement and nuance do.
  2. 02
    Price the volume.The step that runs 1,200 times a day decides the bill. Make that one cheap first.
  3. 03
    Spend on what people read.The output a manager acts on is where quality pays back.
The bigger picture

This call center is not an exception. As companies learn to match each task to its price tag, they stop sending everything to the most powerful model. Frontier models are not losing their place. They are being saved for the steps that need them: judgement, nuance, the output people act on. Everything else moves to cheaper models that are good enough.

The numbers already show it.

14% of the AI usage routed through OpenRouter in Q3 2026 went to the ten highest-rated models, down from 55% in early 2025.
The best model is not the smartest one. It is the right one for the step.
Sources Google: Gemini Developer API pricing OpenAI: API pricing Anthropic: Claude API pricing TechFlow, via The Information: share of OpenRouter usage on top models drops from 55% to 14% (Oct 2026) The call-center case was presented at a talk at IE.