How to Reduce LLM API Costs in 2026 (5 Levers, Priced)
Short answer: the biggest cuts come from prompt caching, where cached input costs one twentieth of normal input on OpenAI's gpt-6.1-sol and Claude Sonnet 5.5, and the Batch API, which halves the price on OpenAI, Anthropic and Google for work that can wait. Then route easy requests to a small model, cap output length and trim context. Prices verified on each vendor's page, October 2026.
Take a support assistant making 100,000 API calls a month on gpt-6.1-sol. Each call sends 4,000 input tokens, 3,500 of them the same system prompt and help-center excerpt, and gets back 500 output tokens. At standard prices that is a $1,300 month: $800 of input and $500 of output (our arithmetic). Here is what each lever does to that bill.
The 5 levers at a glance
| Lever | How it works | Discount (vendor, Oct 2026) | Scenario bill after lever | Catch |
|---|---|---|---|---|
| 1. Prompt caching | Reuse the repeated prefix instead of paying full input price | Cached input $0.10 vs $2.00 per 1M on gpt-6.1-sol | $635 | Cache writes cost more than normal input |
| 2. Batch API | Send work that can wait as an asynchronous batch | 50% off input and output | $650 (or $317.50 with caching) | No real-time answers |
| 3. Model routing | Send easy calls to a small model | gpt-6-luna at $0.10 in / $0.50 out per 1M | $435.50 (70% routed) | Needs a way to judge which calls are easy |
| 4. Shorter outputs | Cap answer length | Output is 5x the input price on sol | $1,100 (300-token answers) | Too tight a cap cuts answers off |
| 5. Leaner context | Send fewer tokens per call | Every input token removed saves $2 per 1M on sol | $1,000 (2,500-token calls) | Removing context can hurt accuracy |
Prices read on OpenAI's API pricing page on October 28, 2026, standard short-context rates. Scenario bills are our arithmetic, each lever applied alone to the baseline; cache-write costs are covered under lever 1.
Dupple data: Toolradar, our software directory, tracks 50 AI observability tools and 51 AI model deployment tools as of October 28, 2026 (AI observability, AI model deployment). You need one of the first kind to see which requests drive the bill before you pull any lever below.
Use prompt caching for the repeated part of every call
Prompt caching is the first lever to pull whenever many calls start with the same block of text. On OpenAI's page, gpt-6.1-sol input costs $2.00 per million tokens, but cached input costs $0.10. Claude Sonnet 5.5 is priced the same way on Anthropic's pricing page, for normal input and cache reads alike.
In our scenario, 3,500 of every 4,000 input tokens are the shared prefix. Cached, those 350 million tokens a month cost $35, the other 50 million cost $100, and output stays at $500, for $635 a month, less than half the baseline (our arithmetic).
The catch is the write. Putting the prefix into the cache costs more than sending it normally: on sol, a 30-minute cache write is $2.50 per million tokens and a 12-hour write is $5.00. Writing our 3,500-token prefix once costs under a cent, so even one fresh write every 30 minutes around the clock, about 1,440 a month, adds roughly $13 (our arithmetic, assuming steady traffic). Anthropic's listed cache prices reflect a 5-minute lifetime, so a quiet API that lets the cache expire pays the write more often.
Caching only helps when the repeated part comes first. Put the system prompt, instructions and reference documents at the top of every call, and the user's question last.
How much does prompt caching save on each vendor?
| Vendor and model | Input per 1M | Cached input per 1M | Cache write per 1M |
|---|---|---|---|
| OpenAI gpt-6.1-sol | $2.00 | $0.10 | $2.50 (30 min), $5.00 (12 h) |
| OpenAI gpt-6-luna | $0.10 | $0.01 | $0.125 (30 min), $0.25 (12 h) |
| Claude Sonnet 5.5 | $2 | $0.10 | $2.50 (5-min lifetime) |
| Claude Haiku 5.5 (prompts up to 100K tokens) | $0.10 | $0.01 | $0.125 |
| Gemini 4 Argon, through November 30, 2026 | $2.00 | $0.10 | Implicit caching only |
| Gemini 3.5 Flash-Lite | $0.30 | $0.03 | Storage $1.00 per 1M tokens per hour |
Read on OpenAI's, Anthropic's and Google's Gemini API pricing page on October 28, 2026.
Two Google details change the math. Gemini 4 Argon supports implicit caching only, so you cannot pin a prefix yourself, and Gemini 3.5 Flash-Lite charges for storage by the hour: keeping our 3,500-token prefix cached all month costs about $2.52 (our arithmetic).
Move anything that can wait to the Batch API
Batch processing halves the price at all three vendors. OpenAI's batch rate for gpt-6.1-sol is $1.00 input and $5.00 output, Anthropic's page says "Save 50% with batch processing", and Google's paid tier lists the Batch API at a 50% cost reduction.
If our whole support workload could wait, the bill would fall to $650. But a live chat cannot wait, so batch fits the jobs around it: tagging yesterday's tickets, summarizing conversations, drafting help articles, running evaluations.
The levers stack. OpenAI prices cached input in batch at $0.05 per million on sol, so the same workload with both caching and batch would cost $317.50 a month: $17.50 for the cached prefix, $50 for the rest of the input and $250 for output (our arithmetic).
Route easy requests to a smaller model
Most traffic does not need the flagship model. OpenAI's gpt-6-luna costs a twentieth of gpt-6.1-sol for both input and output, and Claude Haiku 5.5 carries the same rates as luna for prompts up to 100,000 tokens.
Say 70% of our calls are simple: order status, password resets, opening hours. Sending those 70,000 calls to luna costs $45.50 a month, and the remaining 30,000 hard calls on sol cost $390, so the bill drops to $435.50 (our arithmetic). Combine routing with caching and a 300-token output cap, and the same month comes to about $147.
Routing needs a decision rule: a keyword filter, a cheap classifier call, or a gateway that sends each request to a model by policy. Our best LLM gateways guide covers the tools, and OpenRouter pricing explains what a router charges on top of model prices.
Is OpenAI, Claude or Gemini cheapest for the same job?
At the flagship-model level, the three are priced alike today. Our 100,000-call month costs the same $1,300 on gpt-6.1-sol, Claude Sonnet 5.5 and Gemini 4 Argon at its current rate (our arithmetic). Small models change the picture: Haiku 5.5 or gpt-6-luna brings it to $65, and Gemini 3.5 Flash-Lite to $245, or $122.50 in batch.
Watch the calendar on Google. Gemini 4 Argon is $2.00 input and $10.00 output per million through November 30, 2026, then $4.00 and $20.00 from December 1, 2026. The same workload doubles to $2,600 a month on that date unless you move it.
Cap output length
Output tokens cost five times input on gpt-6.1-sol, so long answers are expensive. If our assistant's average answer drops from 500 to 300 tokens, output falls from $500 to $300 a month and the bill to $1,100 (our arithmetic).
Do it in the prompt ("answer in at most three sentences") and with a maximum output setting as a hard ceiling. Thinking tokens count as output too: Google's page states that Gemini output prices include thinking tokens, so a reasoning model that thinks at length costs more even when the visible answer is short.
Send less context per call
Every token you do not send is a token you do not pay for. Trimming our 4,000-token calls to 2,500, for example by retrieving three help-center passages instead of eight, would cut input from $800 to $500 and the bill to $1,000 (our arithmetic, an example size rather than a measured one).
Context trimming and caching pull in different directions. Once a prefix is cached, shortening it saves little, because cached tokens are already cheap. Trim the part that changes per call, such as retrieved passages and chat history, and cache the part that does not.
If you need fewer API calls at all, running a small open model on your own machine is another route; our guide on how to run an LLM locally prices the hardware.
Hidden costs that push the bill up
- Speed tiers. OpenAI's Fast mode doubles standard prices (gpt-6.1-sol at $4 input, $20 output) and Ultrafast is six times standard. Anthropic's fast mode for Opus 5.5 is 2x, and Gemini 4 Argon's Priority tier is $3.60 input and $18.00 output.
- Region rules. OpenAI charges a 10% uplift on data residency and FedRAMP endpoints for models released on or after March 5, 2026, and Anthropic's US-only inference costs 1.1x.
- Search grounding. On Gemini 3 and newer models, 5,000 grounded search requests a month are free, then $14 per 1,000.
- Long prompts on Haiku. Above 100,000 tokens, Haiku 5.5 input jumps fivefold, to $0.50 per million.
- Cache writes. OpenAI and Anthropic charge more to write a cache than to send normal input, so caching a prefix that is rarely reused can cost more than not caching.
Do LLM gateways reduce cost?
Indirectly. A gateway does not lower the model's list price; it makes routing, caching rules, fallbacks and per-team budgets easier to run, which is how the savings above get applied across a whole company. Check whether your gateway adds its own fee, and compare it with what routing saves. Pair a gateway with one of the LLM observability tools to see cost per feature before and after each change.
How we priced these levers
Every price on this page was read on developers.openai.com, claude.com and ai.google.dev on October 28, 2026, at standard paid-tier rates for short prompts. The scenario (100,000 calls a month, 4,000 input and 500 output tokens per call, a 3,500-token shared prefix) is an example workload, and every bill is our arithmetic on those printed prices. If tokens are new to you, start with what a token is in AI. We re-check these prices every 60 days.
FAQ
What is the fastest way to reduce LLM API costs?
Prompt caching, if your calls share a long fixed prefix. On gpt-6.1-sol and Claude Sonnet 5.5, cached input costs $0.10 per million tokens against $2 for normal input, which halves our example $1,300 month to $635 before cache writes.
How much does the Batch API save?
Half. OpenAI, Anthropic and Google all price batch work at 50% of standard rates, and on OpenAI batch stacks with caching: cached batch input on gpt-6.1-sol is $0.05 per million tokens.
Does prompt caching work on Gemini?
Yes, with limits. Gemini 4 Argon supports implicit caching only, at $0.10 per million cached tokens through November 30, 2026. Gemini 3.5 Flash-Lite supports cached input at $0.03 per million plus a storage charge of $1.00 per million tokens per hour.
Does limiting output tokens lower cost?
Yes, and by more than input trimming, because output costs five times input on gpt-6.1-sol and Claude Sonnet 5.5. Cutting answers from 500 to 300 tokens saves $200 on our 100,000-call month.
Which cheap model can I route simple requests to?
OpenAI's gpt-6-luna and Claude Haiku 5.5 both cost $0.10 per million input tokens and $0.50 per million output (Haiku for prompts up to 100,000 tokens). Gemini 3.5 Flash-Lite costs $0.30 input and $2.50 output, with a free tier.
Will Gemini 4 Argon get more expensive?
Yes. Google's page lists $2.00 input and $10.00 output per million tokens through November 30, 2026, and $4.00 and $20.00 from December 1, 2026, with context caching rising from $0.10 to $0.20.
Cite this: Dupple, "How to Reduce LLM API Costs in 2026 (5 Levers, Priced)", November 2026.
Related
- How much does it cost to build an AI agent?: where API spend sits in a full agent budget.
- Best AI agent frameworks: the code that makes these calls.
- How to build an AI agent: the step-by-step version.
- Best free local LLM tools: models with no per-token bill.
- Dupple learning hub: more guides on AI tools for work.
Cache first, batch what can wait, then route; most teams can cut a flagship-model bill by half or more without changing vendor. Want weekly picks like this for AI at work? Try Dupple X →
Some offers on this page may be paid placements or contain affiliate links.
