Most Teams Overpay 90% on Classification

Updated 2026-08-16: this post priced Gemini 3.5 Flash-Lite at $0.25/$1.50. Google never charged that - our registry row was wrong from the day it was written, and the live price is $0.30/$2.50. Correcting it moves the gap from 18.75x to 13.6x and the saving from 95% to 93%, so the headline now reads 90%. The argument is unchanged: classification on a flagship model is an order-of-magnitude mistake.

The cost gap

Classification covers routing, labeling, and categorizing, making it one of the most common API tasks. It is also the task type where the biggest pricing waste happens across teams.

• -Gemini 3.5 Flash-Lite: $0.30/MTok input, $2.50/MTok output, 85/100 confidence
• -Claude Haiku 4.5: $1.00/MTok input, $5.00/MTok output, 90/100 confidence
• -GPT 5.6 Luna: $1.00/MTok input, $6.00/MTok output, 91/100 confidence
• -Claude Opus 4.8: $5.00/MTok input, $25.00/MTok output, 80/100 confidence

Opus scores 80/100 on classification while Flash-Lite scores 85/100 with higher confidence overall. Opus costs 16.7x more on input and 10x more on output for a lower score.

Real numbers

A typical classification call uses roughly 500 input tokens and 50 output tokens per request. Here is the per-call cost breakdown comparing Flash-Lite to Opus side by side:

• -Flash-Lite: (500/1M x $0.30) + (50/1M x $2.50) = $0.000275
• -Opus: (500/1M x $5.00) + (50/1M x $25.00) = $0.003750

That is a 13.6x difference per call, and at 50,000 calls per day it compounds:

• -Flash-Lite: $13.75/day, $412.50/month
• -Opus: $187.50/day, $5,625/month

Annual difference: $63,400 saved by switching models on a single task type alone.

Why teams still use Opus for classification

Three reasons explain this pattern, and each one is solvable with better tooling:

• -Default model syndrome: the SDK was initialized with Opus and the team left it unchanged per task
• -Lack of visibility: teams have no data on what percentage of their calls are classification vs. reasoning
• -Fear of quality loss: teams worry about cheap model errors, but testing on 100 examples resolves this

The fix

1. Identify which of your API calls are classification tasks today 2. Run a quality comparison on 100 representative inputs from your production traffic 3. Switch classification calls to Flash-Lite or Haiku based on results 4. Monitor quality in production for one week before expanding the rollout

Step 1 is the hardest part for teams that lack per-call visibility into task types. The Tokeven dashboard shows which models handle which task types and where the savings are.

Try a free teardown at /teardown or start tracking with a free dashboard.