Eight ways to cut a token bill, in the order that actually works
Most teams start by switching models. That is usually the fourth-best option available.
Cost work on an LLM product is a sequence, not a single trick. The order matters, because the first four levers act on the volume of tokens you buy and the later ones only act on the rate you pay for them. Teams that start at the bottom of this list usually spend a week and save a few percent.
1. Stop paying twice for the same prefix
Caching is the largest lever because it attacks the biggest line item. The strongest published discount in the index today is 99% (Mimo V2.5 Pro, Novita AI) — and a cache read applies to work you were already doing, with no change in output quality.
2. Trim what travels with every request
A 900-token system prompt resent 10,000 times a day is 270 million input tokens a month for text that never changes. Audit the prompt: examples, tone rules and tool schemas are frequently longer than the task requires.
3. Fix the history policy
Full-transcript history grows every turn, so input grows superlinearly across a conversation. A rolling window, a summary of earlier turns, or re-sending only the last two exchanges keeps input flat.
4. Buy the right shape of model for the workload
The same index ranks completely differently depending on the workload. Input-heavy work should be ranked by input price (long documents, retrieval); output-heavy work by output price (reports, code). The current floors for a general profile:
| Model | Provider | Input /1M | Cached /1M | Context |
|---|---|---|---|---|
| Qwen2.5 Coder 7B | Nebius | $0.01 | — | 33K |
| Qwen2.5 Coder 3B Instruct | Nscale | $0.01 | — | — |
| Qwen2.5 Coder 7B Instruct | Nscale | $0.01 | — | — |
| Nemotron 3.5 Lightning 30b A3b | Perplexity | $0.011 | $0.0011 | — |
| Granite 4.0 H Micro | Cloudflare Workers AI | $0.017 | — | 131K |
5. Route, don't standardise
Not every request needs the model that writes your hardest answer. Classification, extraction and formatting can go to a cheaper model while the reasoning-heavy minority stays on the expensive one. The saving comes from the share of traffic you move, not from a global downgrade.
6. Control output length
Output is billed above input on virtually every model here (median $1.5 vs $0.6 per 1M). Where a shorter answer is acceptable, an explicit length instruction is a permanent discount — and it compounds with every cached prefix you keep.
7. Check the cross-platform spread
The widest spread in the index is on GPT Oss 120b: 11 platforms, from $0.037 to $0.35 per 1M input tokens (9.5×). If you are running an open-weight model, buying it from a cheaper host is a pure rate cut — verify the operational terms first.
8. Watch the changelog instead of the invoice
Prices move weekly. A model you ruled out on price three months ago may be competitive now, and a model you standardised on may have been repositioned. The changelog and the weekly reports exist so that re-evaluation is a two-minute job rather than a project.
Rank every model for your workload shape, recomputed on every refresh.
See the current rankingsFrequently asked
› What is the fastest way to reduce LLM costs?
Measure what you resend. On most chat products the system prompt and retrieved context are re-billed on every call, so caching and prompt trimming move the bill before any model change does.
› Does switching to a cheaper model always help?
It lowers the rate, not necessarily the cost. A cheaper model that needs more retries, longer outputs or extra guardrail calls can cost more per completed task than a more expensive one that gets it right first time.
› How often should I re-check my model choice?
Monthly is enough for planning, weekly if your spend is material. Prices move constantly — the changelog logs every detected move, and the weekly reports summarise them.
Where the numbers come from
The data pages this article draws on — check any figure against them, they are recomputed on every refresh.