I Can't Avoid the Price Gap Anymore

For two years nobody at my company asked what any of this cost. Then it got a $1,500 cap for Claude, separate from Cursor’s $1,000 and separate from ChatGPT Codex’s own $1,500 — three unrelated numbers nobody ever reconciled against each other.

That ended this month. The three caps merged into one pool: $2,000 total per engineer, covering every model and every tool, exceptions granted case by case. Spend more on Claude and there’s less left for everything else.

Losing the separate buckets stings a little — I used to burn $1,500 on Claude without thinking about what that meant for anything else. But unmetered spending was never going to survive contact with a finance department once the novelty wore off, and pooling it was the right call.

The July price comparison I used made GLM substantially cheaper per token than Opus. I did not retain an archived rate card, so I have removed the precise table rather than present those rates as verified current prices. Check Anthropic’s pricing and Fireworks’ pricing for the model, cache tier, and billing mode you actually use.

We do not self-host GLM; Fireworks serves it. Hosted rates, cache hits, and the number of calls per task all affect the bill. A lower token price is only one part of the comparison.

Separate budgets let every tool pretend the others didn’t exist. One pool has to cover all of them, and covering all of them means every model choice now competes with every other one. The price gap between Opus and GLM isn’t new — separate invoices just hid it. What’s new is that I can’t avoid looking at it anymore.