IBM's Anthropic Deal Meets a Cheaper AI Reality
IBM's deal with Anthropic landed while companies were already walking away from premium AI models. The shift was underway before the announcement.
IBM’s deal with Anthropic landed while companies were already walking away from premium AI models. The shift was underway before the announcement. It has only accelerated since.
The spending reversal
Companies are increasingly using lower-priced AI models, including those from China. The move marks a reversal in mindset from the early AI boom, when access to the most powerful systems was the priority. Now the priority is cost per outcome.
Cursor, a startup, advises companies on optimizing AI spending through tokenomics. The term refers to treating AI usage as an economic problem measured in tokens. Mike Saeks, who leads token ROI strategy at Cursor, put it plainly. “The most powerful and expensive AI models aren’t necessary for relatively mundane tasks.”
Some U.S. companies have received significant free token usage from AI vendors during the early scramble for adoption. That era is fading. Finance leaders are now asking which teams drive real value from their AI budgets, not which teams consume the most tokens.
The model-agnostic turn
Cursor’s software is model-agnostic, allowing easy switching between various AI models. This flexibility is becoming a requirement, not a luxury. Engineers rarely volunteer to downgrade to a cheaper model on their own. Tooling that routes tasks to the cheapest model that can solve them is changing that behavior.
Marty Kausas of Pylon has watched companies test multiple providers in quick succession. “There’s zero loyalty that I’m seeing.” The statement reflects a broader pattern. When intelligence becomes a commodity measured in cost per task, switching costs drop and vendor relationships thin.
The China factor
U.S. AI models are closed, while Chinese models are often cheaper and open-weight. DeepSeek, MiniMax, and Moonshot AI have become common names in enterprise evaluations. Their pricing sits well below U.S. frontier tiers. DeepSeek’s V4-Pro model costs a fraction of Anthropic’s state-of-the-art Fable 5 service on a per-token basis.
Anthropic and OpenAI have accused Chinese model-makers of technology theft. Anthropic has described what it calls industrial-scale distillation campaigns by DeepSeek, Moonshot AI, and MiniMax. The companies deny the claims. Beijing has dismissed the accusations as groundless. The dispute remains unresolved and now sits inside procurement decisions.
Security teams weigh the tradeoff. Cheaper models can mean lower bills and faster iteration. They can also mean less control over where inference runs and how data is handled. For some workloads, the risk is acceptable. For others, it is not.
Anthropic’s lower-cost option
Anthropic released a lower-cost model that offers a choice between higher intelligence or lower cost. The move mirrors OpenAI’s recent pricing adjustments. OpenAI’s GPT-5.6 Sol is designed to be more token-efficient, and the company has cut prices on its Luna and Terra tiers by as much as 80 percent.
The pattern is clear. Frontier providers are racing down the cost curve to keep enterprise workloads from drifting to open-weight alternatives. The question is whether the price cuts arrive fast enough to change behavior that has already hardened.
IBM’s position
IBM has struck a deal with Anthropic. The partnership integrates Claude models into IBM’s software portfolio and consulting practice. It was announced before IBM’s August 2026 alliance with OpenAI, which brings GPT-5.6, Codex, and ChatGPT Work into IBM Consulting Advantage.
The dual partnerships signal a model-agnostic strategy at the system-integrator level. IBM is not betting on a single AI lab. It is betting on distribution. Enterprises want governed, multi-vendor portfolios. They want the ability to route work by cost, security, and performance without rewriting their stack.
What the pattern suggests
The rapid shift to cheaper AI models threatens the high valuations of leading AI companies like Anthropic and OpenAI. If enterprises treat intelligence as a utility priced per task, margin compression follows. Price cuts can defend volume. They do not restore pricing power.
The precedent here is not subtle. Cloud compute, advertising inventory, and mobile app distribution all followed a similar arc. Early scarcity gives way to oversupply. Buyers optimize. Incumbents cut prices. New entrants compete on cost and openness.
This time the geopolitical layer adds friction. U.S. restrictions on advanced chips constrain Chinese training runs. Yet Chinese inference costs remain dramatically lower. Open-weight distribution allows multiple providers to compete on the same model. That competition drives prices down faster than closed ecosystems can match.
The unresolved question is security. Cheaper models may be less secure by design, or they may be secure enough for the tasks that matter most to the bottom line. Procurement teams are making that call every week. Their choices will determine whether premium U.S. models retain enterprise mindshare or become niche tools for high-stakes workloads.
The pattern does not predict the outcome. It only frames the pressure. Valuations built on scarcity assumptions face a market that now measures intelligence in cost per accepted outcome. The companies that adapt fastest to that metric will set the next standard. The others will learn what happens when loyalty disappears and price becomes the only language buyers speak.