AI Agents Just Got Up To 100x Cheaper: What The Collapse In Running Costs Means For UK Businesses And Smaller Firms

Up to 100x - Reported reduction in the cost of running enterprise AI agents using new open-weight models  ·  Cost, not capability - The thing that has quietly been killing many AI projects - escalating running costs, not lack of ability  ·  Open-weight - Models you can run on your own terms are driving the fall, versus defaulting to expensive frontier models  ·  New use cases - Automations that never justified frontier-model costs suddenly make financial sense at a fraction of the price

Among the flashier AI announcements of September 2026, one of the more consequential developments got comparatively little attention: the cost of actually running AI agents is falling fast. New open-weight models built specifically for enterprise AI agents - the likes of Abacus.AI's Smaug family - are reportedly cutting the cost of running agents by up to 100 times compared with the frontier models many businesses reached for by default. It is easy to file that under technical trivia, but doing so would be a mistake, because it changes the economics of business AI in a way that matters more to ordinary firms than most model launches. The reason is simple and under-appreciated: cost, not capability, has quietly been one of the biggest killers of AI projects. Plenty of AI pilots work technically and still get shelved because running them at scale is too expensive to justify. When that cost falls by an order of magnitude or more, the maths behind a huge range of everyday automations flips from 'not worth it' to 'obviously worth it'. As an AI Agency London, we think this trend deserves far more attention than it is getting, and this analysis is why.

What Is Driving The Cost Collapse

The fall in running costs is being driven mainly by a shift in what businesses run, not by frontier models simply getting cheaper. For a long time the default was to point the biggest, most capable frontier model at every problem, which is a bit like hiring a top barrister to answer routine emails - enormously capable, and wildly overpriced for the job. What is changing is the arrival of capable open-weight models built specifically for enterprise agent workloads, which can be run far more cheaply and on your own terms. Two things make them cheap: they are sized and optimised for the practical tasks agents actually do, rather than for maximum general intelligence, and being open-weight they avoid the premium pricing of always calling someone else's frontier model. The result is that for the large majority of everyday business tasks - classifying, extracting, routing, drafting, answering from known information - you can now use a model that is more than good enough at a tiny fraction of the cost, and reserve the expensive frontier models for the genuinely hard problems that need them. That right-sizing is the heart of the cost collapse.

Why This Disproportionately Helps Smaller UK Firms

Falling running costs are good news for everyone, but they are especially good news for smaller and mid-sized businesses, and that matters for Britain where such firms are the backbone of the economy. Large enterprises could always absorb high AI running costs on high-value use cases; a small firm often could not, which is a big part of why the AI divide between big and small businesses has been widening. When the cost of running an agent drops by up to 100 times, automations that were previously the preserve of well-resourced companies come within reach of a small business - automating customer responses, processing documents, handling routine admin, answering questions from a knowledge base. The economics that used to say 'only worth it if you're big' start to say 'worth it for almost anyone'. This is genuinely levelling: it lets a small UK firm apply AI to the same volume of routine work as a larger competitor without a prohibitive bill. It will not close the gap on its own - smaller firms still need the know-how to deploy AI well - but it removes one of the biggest barriers that has been holding them back, and that is something to welcome.

• Right-size the model - use a cheap, capable model for routine, high-volume tasks and reserve expensive frontier models for genuinely hard problems.
• Do the maths per task - work out what an automation costs to run against what it saves; the cost collapse flips many previously-borderline cases to clearly worth it.
• Volume is where it pays - the biggest wins are high-volume routine work, exactly the work that used to be too expensive to automate at scale.
• Watch for hidden costs - cheaper per-call does not mean free; agents that loop or over-call can still add up, so cost monitoring still matters.
• Don't chase cost at the expense of quality - the goal is the cheapest model that is genuinely good enough for the task, not the cheapest model full stop.

The Bottom Line

The up-to-100x fall in the cost of running AI agents is one of the most practically important developments of 2026, even if it made fewer headlines than the flashier launches. It matters because cost, not capability, has quietly been killing AI projects: plenty of pilots work and still get shelved because running them at scale is too expensive. The collapse is driven not by frontier models getting cheaper but by capable open-weight models, built for real agent workloads, letting businesses right-size the model to the task instead of overpaying for maximum intelligence on routine work. That flips the economics of a whole range of everyday automations from 'not worth it' to 'clearly worth it', and it disproportionately helps the smaller and mid-sized UK firms that could never absorb high running costs - a genuinely levelling shift. The way to benefit is disciplined, not reckless: match the model to the task, do the maths per automation, reserve the expensive models for the hard problems, and keep an eye on cost even as the per-call price falls. Cheaper agents do not make strategy optional; they make far more of it affordable - which is exactly the ground where good deployment pays off.

References & Further Reading

• Blog.mean.ceo - Latest AI announcements, September 2026 (open-weight enterprise agent models and cost reduction): https://blog.mean.ceo/latest-ai-announcements-news-september-2026/
• AI Agent Store - AI Agents News, week of September 13 2026 (cost and model innovations): https://aiagentstore.ai/ai-agent-news/this-week
• Gartner - AI agent spending and the risk of escalating, unmanaged cost in agentic projects: https://www.gartner.com/en/newsroom
• MIT / NANDA - the state of AI in business: why pilots stall before scale (cost among the reasons): https://mlq.ai/media/quarterly_decks/v0.1_State_of_AI_in_Business_2025_Report.pdf
• CMArix - powerful agentic AI trends to watch in 2026 for enterprises: https://www.cmarix.com/blog/agentic-ai-trends/