Urgent.News

What's breaking now, across hundreds of outlets.

AI

Output tokens are 17% of the volume and 50% of the bill — a year of Gemini 4 Neon vs Claude Opus 5.5

Most API pricing pages quote three numbers per million tokens, and it's easy to read them as "this one is $1 more expensive". The input prices of Gemini 4 Neon and Claude Opus 5.5 differ by exactly $1 per million tokens. Over a year of realistic traffic, that $1 becomes $240. Here is the arithmetic, so you can redo it with your own token mix. The only formula you need Both prices are per 1M…

A year of using Gemini 4 Neon and Claude Opus 5.5 revealed stark differences in cost. The input charges for both models differed by only $1 per million tokens, but over 12 months of typical usage, that $1 escalated to $240. The pricing structure is straightforward: one per million tokens, with cached input billed separately. No tiers or minimums exist in the published tables.

For consistent comparison, the analysis used a fixed workload: 10 million input tokens and 2 million output tokens per month, with no cached hits. After deducting the 12 million tokens used, the monthly and yearly costs for each model were calculated. The output tokens constitute 16.7% of the total tokens, contributing half of the total bill. This isn't unique to Neon; every row in the table shows output priced at exactly five times its input rate, with output occupying 50% of the total cost.

A significant factor in cost is cached input. Google's Gemini 4 Neon offers a cached-input rate of $0.15 per million tokens, which is just 5% of its input rate. If a significant portion of the input is cached, Neon becomes much cheaper. However, Anthropic's Claude Opus 5.5 does not disclose a cached-input rate. Comparisons can only be made when both models have published cached rates.

The model with cache hits priced at zero is cheaper than the one with the published rate. The "fast" tier of Claude Opus 5.5 doubles the costs compared to the standard tier, making it $100 more expensive per month than Neon and 167% higher.

To perform similar calculations for your own usage, gather actual input and output token counts from your provider's usage dashboard. Multiply these by the per 1 million token rates from the vendor's pricing page and record the date of reading. If caching is used, split the input into cached and uncached portions and price them separately. It's crucial to re-run these calculations whenever a vendor adjusts their rates, as they can change rapidly.

Written by urgent.news from Dev.to's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.

Also reported by 1 other outlet

Every outlet that reported this, and when →

Read the original at dev.to →

More in AI

Stop Sitting With Your AI in a Chat

I do not sit with AI in a chat anymore. I give it tasks in a tracker, the same way I give tasks to a team. It calls me only when it cannot go on without me. It has been about a month.

  • Delegates tasks to ten AI assistants, each with specific role
  • Uses task tracker like Todoist for assignment and communication
  • Spartan assistant oversees new requests and delegates tasks

More from Saturday 21 November →