# How to Add AI to Your SaaS Without Wrecking Your Margins

> Add AI features to your SaaS by pricing per task, routing volume to a cheap tier, caching context and planning on list prices, not launch discounts.

- Canonical: https://zeroraised.com/blog/how-to-add-ai-features-to-a-saas
- Published: 2026-10-29
- Author: ZeroRaised Team (https://zeroraised.com/blog/authors/zeroraised)
- Publisher: ZeroRaised (https://zeroraised.com/)
- Section: Growth
- Tags: artificial-intelligence, saas, revenue, bootstrapping

Adding AI to a SaaS wrecks margins when you price the feature the way you buy the model. Tokens are your cost, not the customer's unit of value, and introductory rates expire. The fix is unglamorous: choose one job, measure cost per completed task, route volume to a cheap tier, cache what does not change, and price on list rates.

**TL;DR**

- Pick one job with a measurable outcome before you choose a model.
- Price the feature per task, never per token.
- Send high volume work to a cheap tier such as Claude Haiku 5.5 at $0.10 and $0.50 per million tokens.
- Cache stable context, because cache reads cost far less than fresh input.
- Model margin on list prices and put usage caps in every plan.

**Disclosure:** ZeroRaised publishes this guide and is a launch and discovery site for bootstrapped products. Facts come from the linked sources as checked on 29 Oct 2026; ZeroRaised has no commercial relationship with the companies named. The steps are our own recommendations, not vendor claims.

**Cost per task** is the total model spend required to complete one unit of user value, such as one summary, one classification or one resolved support reply. It is the number that belongs in your pricing, not the price per million tokens.

## Start with one job, not a feature list

A feature list produces a demo; a job produces a margin. Before you add anything, write one sentence naming the job, the user and the result: "turn this support ticket into a draft reply" or "classify this invoice by vendor".

That sentence gives you a test set, a success definition and a pricing unit. It also narrows the model choice. Anthropic says Haiku 5.5 was designed for high volume, cost sensitive tasks including summaries, classification and live customer support ([Anthropic](https://www.anthropic.com/claude-haiku-5-5)). If your first job is one of those, you likely do not need the most expensive model. Ship one job well, then add the second.

## How should you price an AI feature?

Price the outcome, not the plumbing. A buyer will not accept "4,200 input tokens and 600 output tokens". They will accept "$0.30 per drafted reply" or "50 drafts included, then $0.25 each".

Two shapes work: an included monthly allowance with per task overage, which caps the damage from a heavy user; or pure usage from the first task, which is fairer at low volume but unpredictable. Either way, set the price from measured cost per task. The [guide to pricing an AI product](https://zeroraised.com/blog/how-to-price-an-ai-product) covers margin targets, and the [SaaS pricing guide](https://zeroraised.com/blog/how-to-price-a-saas-product) covers packaging.

## Why should a cheap tier carry the volume?

Because most traffic is boring. Claude Haiku 5.5 costs $0.10 per million input tokens and $0.50 per million output for prompts up to 100K tokens, rising to $0.50 and $2.50 above that, with cache reads at $0.01 ([Anthropic](https://www.anthropic.com/claude-haiku-5-5)). Haiku 4.5 was $1.00 and $5.00 ([Anthropic](https://www.anthropic.com/claude-haiku-5-5)). Anthropic says prompts up to 100K were around 90 percent of requests to the previous Haiku model, and that Haiku 5.5 costs around 75 percent less to run on average ([Anthropic](https://www.anthropic.com/claude-haiku-5-5)).

That is a ten times cut on input for the tier carrying most requests. If every ticket goes to a frontier model because one endpoint was easier to wire, you pay a reasoning premium for work that does not reason.

Keep the expensive tier for ambiguous inputs, long documents and anything published without review. A length or confidence check is often enough to split the traffic.

## Where do introductory prices bite?

Gemini 4 Argon launched at $2 input and $10 output per million tokens through 30 November, then moves to $4 and $20 ([llmwise](https://llmwise.ai/data/ai-price-and-limit-changelog)). Gemini 4 Carbon has the same introductory terms, cached input at 95 percent off, and the same $4 and $20 list rate; it is also not yet available to developers ([Google](https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-carbon/)).

Gemini 3.8 Flash launched at $0.75 and $3.75 and moves to $1.50 and $7.50 from 1 January 2027 ([llmwise](https://llmwise.ai/data/ai-price-and-limit-changelog)). For comparison, Sonnet 5.5 and GPT-6.1 Sol list at $2 and $10, Opus 5.5 at $4 and $20, and GPT-6.1 Terra at $1 and $5 ([llmwise](https://llmwise.ai/data/ai-price-and-limit-changelog)).

Build two spreadsheets: today's promotional rate, and the list rate with its start date. If the feature only works in the first, redesign it. The [price increase analysis](https://zeroraised.com/blog/gemini-mistral-ai-price-increases) shows how often list prices move.

## Does effort level really change the bill?

Yes, more than most founders expect. Claude Opus 5.5 scores 35 at low effort and 58 at max, at ten times the cost per task, $0.98 against $10.04 ([Artificial Analysis](https://artificialanalysis.ai/articles/artificial-analysis-intelligence-index-v5)). Median output tokens per task rose by two thirds, from 35,000 to 59,000 ([Artificial Analysis](https://artificialanalysis.ai/articles/artificial-analysis-intelligence-index-v5)).

Effort is a product decision. Classification rarely needs maximum reasoning; contract review might. Set it per job and revisit on quality complaints, not because a new model defaults to high.

## Why does cost per task beat model rank?

Cost per task varies by more than 13 times among the ten highest scoring models, from $0.96 for GPT-6.1 Terra at max effort to $13.04 for Claude Fable 5.1 at max ([Artificial Analysis](https://artificialanalysis.ai/articles/artificial-analysis-intelligence-index-v5)). GPT-6.1 Sol scores within 4 points of Opus 5.5 at about an eighth of the cost per task ([Artificial Analysis](https://artificialanalysis.ai/articles/artificial-analysis-intelligence-index-v5)). No single model leads all nine evaluations, and 60 percent of the index now uses private test sets or held-out answers ([Artificial Analysis](https://artificialanalysis.ai/articles/artificial-analysis-intelligence-index-v5)).

Rank tells you what to shortlist; your test set tells you what to ship. The [LLM API selection guide](https://zeroraised.com/blog/how-to-choose-an-llm-api) and the [LLM cost reduction guide](https://zeroraised.com/blog/how-to-reduce-llm-costs) cover the comparison and caching work.

## A practical sequence to ship without wrecking margins

1. Pick one job and one success measure. Collect 50 real inputs with expected outputs.
2. Benchmark two or three tiers on that set. Record quality, latency and cost per completed task, not vendor demo prompts.
3. Default the volume path to the cheap tier and route only hard cases upward.
4. Cache what is stable. Anthropic says the Sonnet 5.5 cache read price halved, making it around 20 percent cheaper on most agentic work ([Anthropic](https://www.anthropic.com/claude-haiku-5-5)).
5. Price on list rates, with the promotional rate on a separate dated line.
6. Add usage caps and overage.
7. Review margin per account every month; aggregate margin hides the customer who eats the plan.

## Where ZeroRaised fits

ZeroRaised is a launch and discovery site for products built without institutional venture capital or accelerator equity. If you have shipped an AI feature and want a page that explains the job and the price, [submit your product](https://zeroraised.com/submit). The [pricing page](https://zeroraised.com/pricing) lists the current options. We do not promise traffic or results, and a listing will not fix negative margin. If you want more breakdowns, [join the ZeroRaised newsletter](https://zeroraised.com/newsletter).

## Sources

- [llmwise: AI price and limit changelog](https://llmwise.ai/data/ai-price-and-limit-changelog)
- [Anthropic: Introducing Claude Haiku 5.5](https://www.anthropic.com/claude-haiku-5-5)
- [Google: Introducing Gemini 4 Carbon](https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-carbon/)
- [Artificial Analysis: Intelligence Index v5](https://artificialanalysis.ai/articles/artificial-analysis-intelligence-index-v5)

## Frequently asked questions

### How do I price an AI feature inside a SaaS plan?

Start from cost per completed task, not cost per token. Measure what one successful outcome costs on your real inputs, add a margin you can defend, then either include a monthly allowance or charge per task. Publish usage caps so a heavy user cannot silently turn your best customer into your worst margin.

### Which model should I use for high volume AI features?

Use a cheap, fast tier for the high volume path. Claude Haiku 5.5 costs $0.10 per million input tokens and $0.50 per million output tokens for prompts up to 100K tokens, and Anthropic says prompts up to 100K made up around 90 percent of requests to the previous Haiku model. It was designed for summaries, classification and live support.

### Why should I avoid introductory model prices in my forecasts?

Because they are temporary. Gemini 4 Argon launched at $2 input and $10 output per million tokens and moves to $4 and $20, and Gemini 4 Carbon has the same introductory terms. Gemini 3.8 Flash goes from $0.75 and $3.75 to $1.50 and $7.50 on 1 January 2027. A margin that only works during the discount is not a margin.

### Does model effort level really change my costs?

Yes. Artificial Analysis reports Claude Opus 5.5 scoring 35 at low effort and 58 at max effort, at ten times the cost per task, $0.98 against $10.04. Median output tokens per task also rose from 35,000 to 59,000. Match effort to the job rather than running everything at maximum.

### Should I build on the top ranked model?

Not automatically. Artificial Analysis found cost per task varies by more than 13 times among the ten highest scoring models, and no single model leads all nine evaluations. A model a few points lower can cost an eighth as much per task. Rank is a starting point for shortlisting, not a purchasing decision.