---
title: "How to reduce AI costs: 9 ways that work in 2026"
description: "Nine ways a business can cut what it pays for AI: free models, cheaper workers under a strong model, lower effort, caching and batch, with vendor numbers."
url: "https://madsalt.tech/blog/how-to-reduce-ai-costs/"
author: "Denys Zaichenko"
author_url: "https://madsalt.tech/about/"
language: "en"
published: 2026-11-11
---

# How to reduce AI costs: 9 ways that work in 2026

TL;DR

The fastest way to reduce AI costs is to stop paying top-model prices for every step. Turn the effort level down first: in [Anthropic's own test](https://platform.claude.com/docs/en/about-claude/models/optimizing-for-cost-and-intelligence), one model solved 87.4% of coding tasks at low effort for USD 0.11 each against 95.3% at high for USD 0.29, and running low then retrying only the failures at high reached about 97% for about USD 0.17. Then cache repeated prompts, batch work that can wait ([50% off](https://platform.claude.com/docs/en/about-claude/pricing)), give simple steps to free or cheap models, and keep the strong model as the one that decides.

I'm Denys, a full-stack and AI engineer. I build AI agents and automations for small firms. This list is the one I use. It is for the owner who sees an AI line on the invoice growing and wants to know which levers move it. Numbers are vendor figures from November 2026, linked where they appear; benchmark numbers from Anthropic are its own internal tests.

## Where does an AI bill come from?

From tokens, the pieces of text a model reads and writes. Writing costs more than reading: on Anthropic's price list a strong model charges USD 4 per million tokens read and USD 20 per million written, and its smallest model USD 0.10 and 0.50 ([Anthropic pricing](https://platform.claude.com/docs/en/about-claude/pricing)). So a bill grows with three things: which model you call, how much it thinks and writes, and how often you send it the same long text again. Every way below cuts one of those.

| Way                                      | What it cuts             | Typical saving (vendor figures)     |
| ---------------------------------------- | ------------------------ | ----------------------------------- |
| 1. Lower the effort level                | thinking and writing     | medium 30–52% less than high        |
| 2. Low first, retry failures high        | thinking on easy tasks   | about 40% less than all-high        |
| 3. Free models for simple steps          | price per token          | USD 0 within daily caps             |
| 4. Strong model decides, cheap ones work | price per token          | about a third, on big jobs only     |
| 5. Prompt caching                        | re-reading the same text | cache reads at 5–10% of input price |
| 6. Batch what can wait                   | everything               | 50% off                             |
| 7. Trim what you send                    | reading                  | 21–26% on one small agent           |
| 8. Ask for shorter answers               | writing                  | about 13%                           |
| 9. Seats vs API                          | unused seats             | depends on use                      |

## 1. Turn the effort level down

Most current models have an effort or reasoning setting, and it is the biggest single lever after the model itself. Anthropic's levels are low, medium, high, xhigh and max, and effort changes all tokens in the answer, including tool calls and thinking ([Anthropic effort docs](https://platform.claude.com/docs/en/build-with-claude/effort)).

Anthropic's published test on Claude Opus 5.5 (a SWE-bench Pro subset, its own benchmark) ([cost and intelligence guide](https://platform.claude.com/docs/en/about-claude/models/optimizing-for-cost-and-intelligence)):

| Effort           | Tasks solved | Cost per task |
| ---------------- | ------------ | ------------- |
| low              | 87.4%        | USD 0.11      |
| medium (default) | 92.8%        | USD 0.21      |
| high             | 95.3%        | USD 0.29      |

The same guide says medium cost about 30% to 52% less than high on long coding tasks, and low about two thirds less. OpenAI's models have the same switch (`reasoning.effort`, low to max), and OpenAI notes reasoning tokens are billed as output, but it publishes no percentage, so I don't quote one ([OpenAI reasoning guide](https://developers.openai.com/api/docs/guides/reasoning)).

Set it too low and the model stops before it has enough evidence. Start at medium, measure, then go lower where results hold.

## 2. Run low first, retry only the failures at high

In the same Anthropic test, running everything at low and re-running the 13% of tasks that failed at high solved about 97% for about USD 0.17 a task, against 95.3% for USD 0.29 with everything at high. Anthropic calls it about 40% less at the same pass rate ([guide](https://platform.claude.com/docs/en/about-claude/models/optimizing-for-cost-and-intelligence)).

![Cost per task by effort level. USD per coding task, one model, vendor benchmark. Low effort: 87.4% solved: 0.11; Medium effort: 92.8% solved: 0.21; High effort: 95.3% solved: 0.29; Low, failures retried at high: about 97%: 0.17.](https://madsalt.tech/charts/en/how-to-reduce-ai-costs.svg)

*Source: Anthropic cost and intelligence guide, November 2026. Vendor's internal benchmark.*

This works when you can tell automatically that a task failed: a test, a form check, a missing field. Without that check, you don't know what to retry.

## 3. Use free models for simple steps

OpenRouter, one API in front of many model providers, lists free versions of some models, marked `:free`. Its limits ([OpenRouter limits](https://openrouter.ai/docs/api-reference/limits)):

- 20 requests a minute;
- 50 requests a day if you have bought less than USD 10 of credits in total;
- 1,000 requests a day once you have bought USD 10 or more;
- failed attempts count toward the daily cap.

OpenRouter's free collection in November 2026 included NVIDIA Nemotron 3 Ultra and Nemotron 3.5 Super VL, Cohere North Mini Code and Google Gemma 4 ([free models](https://openrouter.ai/collections/free-models)). The list changes weekly, and some free listings expire.

Read the data policy before you send anything private. OpenRouter itself does not store prompts unless you opt in, but each provider has its own policy, and some are marked "may train" on your data ([provider logging](https://openrouter.ai/docs/guides/privacy/provider-logging)). Free models are for public or anonymised text: summaries of web pages, sorting, first drafts, not client records.

## 4. Let a strong model decide and cheap models do the work

Keep the expensive model as the one that plans and checks, and hand the bulk steps to cheaper ones. Both main coding agents support this: Claude's subagents take their own `model` setting, and Codex has `default_subagent_model` and `default_subagent_reasoning_effort` ([Anthropic subagents](https://code.claude.com/docs/en/sub-agents), [Codex subagents](https://learn.chatgpt.com/docs/agent-configuration/subagents)).

What that split looks like on one screen, with the cost of each agent as a bar, is in my [multi-agent orchestration dashboard](https://madsalt.tech/work/multi-agent-orchestration/) prototype (sample data).

Be honest about where it pays. Anthropic measured a strong orchestrator with mid-size subagents at about a third less cost for about the same score, but on work larger than one context window. When the work was one dependent chain or fitted in a single context, one model at lower effort came out cheaper in every case it measured ([guide](https://platform.claude.com/docs/en/about-claude/models/optimizing-for-cost-and-intelligence)). Its advice: tune effort on one model first, because most workloads end there.

## 5. Cache the text you send every time

If every request starts with the same instructions, documents or tool list, cache them. On Anthropic a cache read costs 10% of the normal input price, and 5% on its newest Opus and Sonnet models; the cache write costs a bit more once ([Anthropic pricing](https://platform.claude.com/docs/en/about-claude/pricing)). OpenAI prices cached input at 5% on GPT-6.1 Sol ([model page](https://developers.openai.com/api/docs/models/gpt-6.1-sol)). Anthropic measured caching cutting a research agent's cost by 4.8 to 7.2 times and calls it usually the largest lever.

One trap from the same guide: a 25-token status line at the very start of the prompt changed every request, so every request rewrote the whole cache. Put anything that changes at the end.

## 6. Batch what can wait

Reports, overnight classification, bulk extraction: anything that can wait up to 24 hours can go through the batch API at 50% off on both Anthropic and OpenAI, and on Anthropic the discount stacks with caching ([Anthropic pricing](https://platform.claude.com/docs/en/about-claude/pricing), [OpenAI pricing](https://developers.openai.com/api/docs/pricing)).

## 7. Send less

Every document, tool definition and old message you attach is read again on every call. Anthropic measured input trimming taking a further 21% to 26% off a small triage agent's bill, and loading tool definitions only when needed saving 45% with 500 tools attached ([guide](https://platform.claude.com/docs/en/about-claude/models/optimizing-for-cost-and-intelligence)). Long prompts can also cross a price step: Haiku 5.5 charges 5 times as much for prompts over 100,000 tokens.

## 8. Ask for shorter answers

Writing costs about five times reading on most price lists. Anthropic measured "ask for shorter answers" at about 13% less from 39% fewer output tokens ([guide](https://platform.claude.com/docs/en/about-claude/models/optimizing-for-cost-and-intelligence)). "Answer in three sentences" or "return only the JSON" is a free change.

## 9. Check seats against API use

Team plans are per person: ChatGPT Business and Claude Team standard seats are USD 25 a month, or USD 20 on a yearly plan ([ChatGPT Business](https://help.openai.com/en/articles/8542115-chatgpt-business-general-faq), [Claude Team](https://support.claude.com/en/articles/9266767-what-is-the-team-plan)). Ten seats on yearly billing are USD 2,400 a year. If half the team opens it twice a month, an internal tool on the API with a spend limit can cost less. If everyone uses it daily, seats are usually the better deal.

## My similar case

My own work runs on ways 1, 3 and 4 together. Claude or Codex stays the orchestrator: it plans the task and checks the result. The bulk of the work goes to free models on OpenRouter. And I change the effort level from task to task instead of leaving it high. Effort alone is a big lever: on Anthropic's numbers above it moves the cost of one task from USD 0.29 at high to USD 0.11 at low.

What a whole agent costs per run is in [how much an AI agent costs](https://madsalt.tech/blog/how-much-does-an-ai-agent-cost/), and what went wrong when I ran many agents in parallel is in [multi-agent orchestration: what broke](https://madsalt.tech/blog/multi-agent-orchestration-what-broke/).

Send me what you pay for AI now and what it does, and I'll show you where it can cost less. This work falls under [AI integration](https://madsalt.tech/services/ai-integration/).

Is your AI bill growing every month?

Send me what you pay and what it does. I'll show you which of these levers fit, and set them up.

[Write to me](https://madsalt.tech/#contact)

## Limits of this list

- **Vendor benchmarks.** The effort, caching and orchestration percentages are Anthropic's own internal tests on coding and agent work. Your tasks may move differently; measure on your own.
- **No OpenAI percentages.** OpenAI publishes no token or cost difference between effort levels, so none is given.
- **Prices change.** Read on 9 November 2026. OpenRouter's free list changes weekly, and two promotions on vendor pages end on 30 November and 31 December 2026.
- **Free models** were not tested here for quality or uptime; check them on your own task before relying on one.
