What is the fastest way to cut AI API costs?

Prompt caching, if your workload repeats a large stable prefix like a system prompt or codebase context. Cached input runs a fraction of standard input price — GLM-5.3 charges $0.26 cached versus $1.40 standard, Claude Opus models charge $0.50 cache-hit versus $5. Restructuring prompts so the stable part comes first is usually a few hours of work for a permanent 40-80% cut on input spend.

Does using a cheaper model actually save money?

Only if success rates hold. A model that costs 5x less but fails 15% more often can cost more once you count retries, human review, and downstream errors. Measure price per completed task, not price per token — run 50 real tasks through each candidate and compare total tokens burned including retries.

How much can batching save on AI API costs?

Roughly 50% on providers that offer a batch tier — Anthropic's batch pricing for Opus-class models is $2.50/$12.50 per MTok versus $5/$25 standard. The trade-off is latency: batch jobs return within hours rather than seconds, so it fits background work like evals, bulk classification, embeddings and content generation, not interactive requests.

Quick Answer

How to Cut AI API Costs in 2026: Seven Real Tactics

Published: August 16, 2026

The Short Answer

Most AI API bills are 40-90% larger than they need to be, and almost none of the savings require a worse product. In rough order of return per hour of engineering effort:

Prompt caching — biggest win for repeated context
Model routing — cheap model for easy steps, frontier for hard ones
Output-token discipline — you pay 3-6× more per output token
Batch tiers — ~50% off for non-interactive work
Off-peak pricing — up to 50% on providers that offer it
Context hygiene — stop resending what the model doesn’t need
Measure price per task, not per token — the one that governs all the others

1. Prompt Caching (Biggest Single Win)

If your requests share a large stable prefix — a system prompt, a schema, a codebase slice, a policy document — you are probably paying full input price for the same tokens thousands of times.

Cached-input rates as of August 2026:

Model	Standard input	Cached input
Claude Opus 5	$5	$0.50 (cache-hit)
GLM-5.3	$1.40	$0.26
DeepSeek V4 Pro	$0.435 (off-peak)	~$0.043
DeepSeek V4 Flash	$0.14 (off-peak)	$0.0028
Muse Spark 1.2	$1.25	$0.15

The work: restructure prompts so everything stable comes first and everything variable comes last. Caching keys on prefixes — a single changed character early in the prompt invalidates everything after it. Common own-goal: injecting a timestamp or request ID at the top of the system prompt.

Anthropic lowered its cache minimum to 512 tokens with Opus 5, which brings caching into range for much smaller prompts than before.

2. Route Instead of Defaulting

Very few workloads need one model for everything. The dominant 2026 pattern is a router: a cheap fast model handles the frequent easy steps, and only genuinely hard steps escalate.

A realistic split for an agent loop:

Step type	Share of calls	Model
Classification, extraction, formatting	~60%	Cheapest tier (GPT-5.6 Luna $0.20/$1.20, Gemini 3.7 Flash $0.75/$3.75)
Ordinary code/content generation	~30%	Mid tier (GLM-5.3 $1.40/$4.40, GPT-5.6 Terra $2/$12)
Long-horizon reasoning, unattended work	~10%	Frontier (Claude Opus 5 $5/$25, GPT-5.6 Sol $5/$30)

Sending 100% of that traffic to a frontier model costs roughly 4-6× more than routing it. Nvidia shipped NeMo Switchyard as an open-source router in August 2026 precisely for this pattern; the decision logic is covered in how to choose an AI model in 2026.

3. Output Tokens Are Where the Money Goes

Every major provider charges 3-6× more for output than input. Claude Opus 5 is $5 in / $25 out. GPT-5.6 Sol is $5 / $30. Gemini 3.7 Flash is $0.75 / $3.75.

Practical levers:

Ask for structured output. JSON with defined fields beats prose that restates the question.
Cap max_tokens deliberately. An unbounded ceiling invites padding.
Suppress preamble. “Explain your reasoning, then answer” doubles output on tasks that don’t need it.
Watch reasoning-token inflation. Models with mandatory thinking (GLM-5.3) or high reasoning effort bill those tokens as output. A “cheaper” model that thinks twice as long is not cheaper.

A verbosity fix in your system prompt is often a 20-30% bill reduction with zero quality loss.

4. Batch Tiers for Anything Non-Interactive

If a result can wait hours, batch pricing is roughly half price. Anthropic’s batch rate for Opus-class models is $2.50/$12.50 versus $5/$25 standard.

Fits: evaluation runs, bulk classification, embeddings backfills, content generation pipelines, dataset labeling, nightly summarization. Doesn’t fit: anything a user is waiting on.

Most teams have more batchable work than they think — the reflex is to build everything synchronously.

5. Off-Peak Pricing

Some providers price by clock. DeepSeek V4 charges roughly half off-peak versus peak (peak windows 1-4 and 6-10 UTC). GLM-5.3 gives 50% points outside 14:00-18:00 UTC+8 on weekdays.

If your batch work is already asynchronous, scheduling it into the cheap window is a config change, not an engineering project.

6. Context Hygiene

Long context is not free, and “just send everything” is the default failure mode of agent frameworks.

Prune conversation history rather than resending 40 turns.
Retrieve, don’t dump. Send the 3 relevant files, not the repo.
Summarize old turns into a compact state object.
Check for silent re-sends. Many frameworks resend full tool definitions and history on every step; measure actual input tokens per call rather than assuming.

7. The Metric That Governs Everything: Price Per Task

Every tactic above can be defeated by optimizing the wrong number. A model at one-fifth the token price that fails 15% more often costs more once you count retries, human review, and downstream cleanup.

Measure this instead: run 50 real tasks through each candidate configuration and record total spend including retries, divided by tasks completed successfully. That number is your actual cost. Full method: price per task vs price per token.

Watch for Dated Prices

Introductory rates expire. Gemini 3.7 Flash’s $0.75/$3.75 ends December 31, 2026 and doubles to $1.50/$7.50. Claude Sonnet 5’s $2/$10 intro ends August 31, 2026, moving to $3/$15. If your unit economics only work at the intro rate, you have a deadline, not a saving.

Last verified: August 16, 2026. Prices from vendor pricing pages.