What’s the fastest way to cut 50% of cost?

Move default traffic to a cheaper SKU that still passes your eval. Routing beats prompt nits.

Should I self-host to save money?

Only past a volume/policy threshold. See open vs closed and local vs cloud.

Do reasoning models wreck cost?

They can. Use them on a small percentage of traffic.

Where do I see live prices?

Leaderboard, compare, and each model’s monthly cost table.

LLM cost optimization — cut the bill without wrecking quality

LLM cost optimization is usually routing and prompt shape, not a new vendor logo. Most bills are 90% default-path tokens. If the default path uses a flagship, you will hate finance meetings.

Measure tokens by route. Put a cheap model on the default. Cache static prefixes. Retrieve fewer chunks. Reserve reasoning SKUs for hard jobs. Use this site’s live $ / 1M and monthly tables to sanity-check each move.

Cheapest models in the live catalog

Live OpenRouter pricing and Arena quality, cached about an hour. Not a static blog table.

Model	Provider	Blend $ / 1M	Quality	Context
Mistral Nemo	Mistral	0.02	—	131,072
Ling 3.0 Flash	inclusionAI	0.04	—	262,144
Llama 3 8B Lunaris	Sao10k	0.04	—	8,192
MythoMax 13B	Gryphe	0.06	—	4,096
Nex-N2-Mini	Nex Agi	0.06	—	262,144

The default path is the bill

If 95% of requests are “reset password” macros, those should not hit Opus-class models. Classify intent cheaply; escalate. Compare a flash row to a flagship on a pair page and compute monthly cost at your token volume.

Prompt caching

Put the huge static system prompt first and keep it stable so caches hit. Look at cached input prices in the compare table. A model with cheap cached input can beat a slightly cheaper uncached rival.

Stop stuffing context

Every extra retrieved chunk is input money and slower TTFT. Tune retrieval. See the RAG and context-window guides.

Output budgets

Ask for JSON with small keys. Ban essays in support bots. Output tokens are often pricier than input. Max-output settings are both a safety rail and a cost rail.

Revisit the catalog

New cheap rows appear constantly. An hour of leaderboard time per month beats a heroic rewrite. Sort by cost with a quality floor, open two detail pages, done.

FAQ

What’s the fastest way to cut 50% of cost?: Move default traffic to a cheaper SKU that still passes your eval. Routing beats prompt nits.
Should I self-host to save money?: Only past a volume/policy threshold. See open vs closed and local vs cloud.
Do reasoning models wreck cost?: They can. Use them on a small percentage of traffic.
Where do I see live prices?: Leaderboard, compare, and each model’s monthly cost table.

LLM cost optimizationreduce LLM API costcheaper GPT alternativecut AI token costs

Find cheaper high-enough models