Home Process Plans About
From the blog

AI Just Got Cheaper

October 28, 2026 · 5 min read

OpenAI cut the cost of capable AI agents in half this week. If you have been waiting for the economics to make sense, your wait is over.

Subscribe now

OpenAI has four model tiers. Terra is the middle one - positioned explicitly for routine computer work: reading what is on a screen, clicking through applications, filling out forms, drafting messages, entering invoice data into accounting systems, editing videos, interpreting technical drawings. The stuff your team spends hours on every single day.

The benchmark numbers are worth noting. On OpenAI’s internal test designed to expose model weaknesses, GPT-6.1 Terra’s error rate dropped from 53.6% on the previous generation to 16.7% at low reasoning effort. At maximum effort, it fell to 4.3%. Coding performance more than doubled. It completes tasks more than three times faster than the premium GPT-6 Astra model while costing a fraction as much.

But the benchmark that matters most to a business owner is the one in your own operation. What does it actually cost to have an AI agent process an invoice, route a customer inquiry, or update a CRM record? And is that cost low enough now that it beats doing it manually?

For a growing number of tasks, the answer has just become yes.


The Economics Changed This Week

Before yesterday, the per-task cost of running an AI agent for routine work was viable but not obvious. You could make the numbers work at scale - processing thousands of invoices or handling high-volume customer triage. But for a business doing, say, 50 invoices a week or handling 100 support inquiries a day, the math was marginal.

A 50% price cut flips that calculation.

Here is a concrete example. Say your business processes 200 invoices a month. Each invoice requires extracting vendor name, date, amount, account code, and purchase order number - roughly 2,000 tokens of input, 200 tokens of output. At the old pricing, that cost about $0.52 per invoice, or $104 per month. At new pricing, it is $0.26 per invoice - $52 per month. Add in retries, edge cases, and human review time, and you are still under $100 a month for a process that probably eats 8-10 hours of someone’s time.

The math gets better the more volume you have. A business processing 1,000 customer inquiries a month through an AI triage system goes from roughly $260 to $130 in direct model costs.

The savings are real. But they come with a catch that most businesses miss.


The Trap: Buying a Cheaper Model Won’t Fix a Broken Process

This is where most AI adoption efforts go wrong. A business owner hears about a new model, gets excited, signs up for a tool, and expects results to appear. When they do not, the conclusion is that the AI does not work.

The data on this is sobering. MIT’s Project NANDA found that 95% of generative AI pilots delivered no measurable return to the bottom line. Gartner estimates roughly 31% of AI tools go unused within 90 days of purchase. The average small business spent $2,340 on AI subscriptions in 2025, and roughly a third of that went to tools nobody actually used.

These failures are not model failures. They are process failures.

GPT-6.1 Terra is genuinely cheaper and more capable than what was available last week. But if you point it at a messy workflow - a CRM with duplicate records, an inbox with no routing rules, a spreadsheet with three different date formats - it will amplify the mess, not fix it. A cheaper model running on bad data is just cheaper chaos.

The businesses that get real ROI from AI start with a different question. Not “what can this tool do?” but “what is costing me the most time, and what is the cheapest way to take an hour off the top of it?”


Where to Start (and Where Not To)

The use cases that make sense at the new pricing share three characteristics: high volume, clear rules, and low stakes for mistakes. Think invoice processing, expense categorization, email triage, meeting transcription and summary, inventory reorder triggers, lead qualification.

The use cases that do not make sense are strategic decisions, relationship-sensitive communication, anything requiring legal or regulatory judgment, and processes that are already broken. Automating a bad process does not fix it. It just makes the bad process run faster.

A well-designed AI automation replaces a repetitive step, not a human role. The businesses reporting the best outcomes - McKinsey found 55% of AI high-performers fundamentally redesigned their workflows when deploying AI - treat the model as a tool inside a redesigned process, not as a replacement for thinking about how work actually gets done.


The Bottom Line

GPT-6.1 Terra changes the math on AI automation for SMBs. At half the cost and significantly lower error rates, the economic case for automating routine, high-volume tasks just became straightforward. You can build a working automation for under $100 a month in model costs, and most implementations pay for themselves inside 60 days.

But the model is not the strategy. The strategy is identifying the right bottleneck, scoping the smallest solution that addresses it, and building the workflow so the tool fits into how your team already works - not the other way around.

If you are ready to find the automation that makes sense for your business, we can help. We run a Discovery engagement that maps your actual workflows, identifies the highest-ROI automation opportunities, and builds a working prototype in two weeks. It is the same process we use for our own clients, and it is designed to get you to a working result, not a slide deck.

Reply to this post or reach out at blackandtanlabs.com. The economics just got better. The work of getting it right did not change - but it is work worth doing.

Enjoyed this? New essays on AI, operations, and building in contrast - delivered to your inbox.

Subscribe on Substack →