AI model prices fell again: how a small business should choose without chasing every launch

The right AI model for a small business is usually the one that completes one useful job reliably at a cost you can explain—not the newest one with the loudest launch. Prices changed again in July and August 2026, so compare a real workflow's total cost: model usage, tool calls, retries, and the time a person spends checking the output. Start with a small, supervised test before moving customer-facing work.

What changed

• On 30 July 2026, OpenAI said it reduced GPT-5.6 Luna pricing by 80% and Terra pricing by 20%. See OpenAI's dated update.
• Anthropic says Claude Sonnet 5's introductory API price of $2 input / $10 output per million tokens runs through 31 August 2026, then becomes $3 / $15. Its updated tokenizer can also change how many tokens the same content uses. See Anthropic's announcement.
• Google announced Gemini 3.6 Flash at $1.50 input / $7.50 output and Gemini 3.5 Flash-Lite at $0.30 / $2.50 per million tokens on 21 July 2026. See Google's announcement.

Those are published vendor rates, not a promise of what your business will pay. Availability, plan terms, regions, output length, search or connector fees, and usage patterns can change.

What it means for your business

Lower headline prices are good news, but a cheap model is not automatically a cheap workflow. A customer-reply automation that creates incorrect answers, triggers repeated retries, or needs ten minutes of editing per message can cost more than a slightly pricier model that produces a useful first draft.

For most owners, the better question is: "Can this workflow save more time than it creates?"

Use this simple comparison before changing tools:

Choose by the job, not the brand

Start with one low-risk, repetitive job:

• Turn a sales call into follow-up notes and a draft email.
• Sort inbound enquiries into urgent, routine, and needs-human-review.
• Turn a finished job into a first draft for a social post or case study.
• Summarize weekly numbers into a short owner update.

Run the same sample through the options you are considering. Keep the prompt, source material, and success criteria the same. Then score the result for accuracy, clarity, speed, and how much editing it needs.

That small test is more useful than a benchmark chart because it reflects your business language, documents, and customers.

Where AI bills quietly grow

Before approving a new AI workflow, include these line items in your estimate:

1. Model usage — input and output volume, not just the advertised rate.
2. Connections and tools — automation platforms, search grounding, CRM access, or other add-ons.
3. Retries — failed or vague outputs often mean another call.
4. Human review — especially for customer messages, pricing, finance, or public posts.
5. Setup and maintenance — prompts, permissions, testing, and a person who owns the workflow.

A helpful rule: set a monthly spend cap and review the first two weeks of usage before you expand. Never move a process because a promotional price is ending; move it because the measured outcome is better.

A sensible next step

If you are not sure which AI tool belongs in your workflow, begin with the job—not the model list. Use the AI plan finder to identify a practical first use case, then compare the relevant tools in the AI price index. For a workflow that touches customers or business systems, plan the approvals and handoff before you connect anything through AI integration services.

The model market will keep changing. A well-owned workflow, a clear review step, and a bill you understand will age much better than any launch-day recommendation.