IBM Says the Model Is No Longer the Differentiator — What the Shift to AI System Competition Means for Every Business

Introduction

At some point in the last twelve months, without a specific press release marking the transition, artificial intelligence models stopped being the competitive frontier. The race to announce a more capable model — more parameters, better benchmarks, higher scores on reasoning tests — has not ended, but it has become secondary to a different kind of competition that is harder to see and far more consequential for the businesses actually trying to build lasting value with AI.

Gabe Goodhart, Chief Architect for AI Open Innovation at IBM, named the shift directly in an interview with IBM Think. "We're going to hit a bit of a commodity point," he said. "It's a buyer's market. You can pick the model that fits your use case just right and be off to the races. The model itself is not going to be the main differentiator." That statement is either obvious or alarming depending on where your organisation currently is in its AI journey. For every business that has spent the past two years debating whether to standardise on GPT, Claude, or Gemini, it is a reframing that changes the entire question.

Why AI Models Have Become Commodities

A commodity, in economic terms, is something that can be substituted for a similar version from a different supplier without meaningful loss of value. Oil from one well is functionally equivalent to oil from another. That condition does not require perfect interchangeability — it only requires that the cost of switching from one supplier to another is low enough that no single supplier can sustain a durable pricing premium or capability advantage.

AI models have arrived at that condition, and the journey there was faster than most people expected. At IBM Think 2026, IBM CEO Arvind Krishna described frontier AI models as exactly that: commodities where "you can substitute one for the other — takes a little bit of work, but you can substitute one for the other." The mechanism behind that convergence is well understood. DeepSeek's January 2025 release demonstrated that competitive model quality could be achieved at a fraction of the training compute that Western labs were spending. The economics that once justified treating frontier model access as a scarce resource collapsed. Within months, multiple open-source models matched the performance of frontier commercial models on the majority of enterprise tasks. The gap between the best closed model and a capable open-source alternative narrowed faster than the market had priced in.

The result, as of 2026, is a market in which GPT-5, Claude Opus, and Gemini Ultra all produce outputs that a typical enterprise cannot meaningfully distinguish on most practical tasks. For the minority of tasks where the distinction genuinely matters — highly complex reasoning chains, specialised scientific analysis, tasks requiring contextual judgment across very long documents — model selection remains relevant. For the majority of enterprise workloads, the model is increasingly a commodity input, not a source of competitive differentiation.

You Are Not Talking to a Model — You Are Talking to a System

Goodhart made a point in his IBM Think interview that reframes the question more starkly than any benchmark comparison can. "If you go to ChatGPT," he explained, "you are not talking to an AI model. You are talking to a software system that includes tools for searching the web, doing all sorts of different individual scripted programmatic tasks, and most likely an agentic loop."

That description is accurate and important. When a user asks ChatGPT about a current news event, the model does not produce the answer from its training data alone. It calls a web search tool, receives results, synthesises them, and returns a response. When it writes code and runs it, it calls a code interpreter. When it generates an image alongside a written response, it calls a separate image model. The GPT-4o or GPT-5 model at the centre of that experience is a component, not the product. The product is the orchestrated system that knows when to call which tool, in what order, with what context, and how to assemble the results into a coherent output.

This distinction matters enormously for how businesses think about their own AI strategy. If what a product like ChatGPT is really selling is a well-orchestrated system, then competing with it — or building something equally valuable for a specific enterprise use case — is not primarily a question of model quality. It is a question of orchestration design, tool selection, context management, data integration, and workflow construction. A business that has access to the same underlying models as its competitors but has built a better system around them is not at parity with those competitors. It has a structural advantage that does not expire when the next model version is released.

Cooperative Model Routing: The Architecture IBM Is Betting On

Goodhart went further in his analysis, describing the architectural pattern he expects to define competitive AI systems in 2026. "In 2026, I think we'll see more sort of cooperative model routing," he said. "You'll have smaller models that can do lots of things and delegate to the bigger model when needed."

Cooperative model routing is a specific and practical idea. Not all tasks require the most capable, most expensive model available. A smaller, faster, cheaper model can handle the majority of routine queries — answering a factual question, formatting a document, generating a standard email response — at a fraction of the cost per token of a frontier model. The frontier model is reserved for tasks that actually require its capabilities: complex reasoning across many pieces of information, judgment calls with significant consequences, tasks where the quality ceiling of a smaller model would produce unacceptable results.

The economic logic is compelling. If 80% of an organisation's AI queries can be routed to a model that costs one-tenth as much per token as the frontier option, the average cost per query drops dramatically without meaningful quality loss on the tasks where quality matters most. The orchestration layer that makes those routing decisions — analysing each incoming query, selecting the appropriate model tier, monitoring the output quality, and escalating to a higher-capability model when the smaller one's response is insufficient — is the source of the cost efficiency. And building that routing layer well is a system engineering challenge, not a model quality challenge.

IBM has been developing this architecture for its own consulting platform, which Goodhart and his colleagues have described as a multi-year internal transformation. The same principles that IBM applied to its own operations are now being offered to enterprise clients through its Enterprise Advantage platform, which Goodhart's IBM colleague Francesco Brenna described on the Mixture of Experts podcast as centred on "context engineering, orchestration, and control" rather than on model selection.

What This Means for Enterprise Software Vendors

For enterprise software companies — the vendors of CRM systems, ERP platforms, productivity suites, and specialised industry applications — the shift Goodhart describes is creating an identity crisis that is playing out in real time across the sector.

An IDC analysis published in May 2026 described the predicament directly: if AI agents become the primary orchestration layer in enterprise computing, competitive advantage shifts from interface quality to data quality and API comprehensiveness. A cross-application agent coordinating work across a company's CRM, ERP, and HR systems is not choosing platforms based on how they look. It is choosing them based on how reliably they expose their capabilities, how trustworthy their data is, and how cleanly they can be integrated into a multi-agent workflow. This is existential for vendors whose moats are built on user experience and interface stickiness. If an AI agent becomes the primary user of enterprise software rather than a human being, the metrics that have historically defined product success — daily active users, session duration, feature adoption rates — become secondary to API depth, data model quality, and MCP compatibility.

Salesforce has already signalled its response to this dynamic. Having built Agentforce as a proprietary agent orchestration harness inside its own ecosystem in 2024, the company pivoted in 2025 toward supporting the Model Context Protocol and making its systems accessible to external agents rather than exclusively to agents built on Agentforce. The logic is the same one that Goodhart described: a vendor that tries to own the entire stack — model, orchestration, data, and workflow — in a world where models are commodities and orchestration protocols are open standards is betting against the architecture that the industry is building toward. The vendors best positioned for the next phase are those investing in API depth, clean data models, and open protocol compatibility, while making a clear strategic choice about whether they want to own the orchestration layer or be the most capable system that orchestration layers call.

What a Business Trying to Build Durable AI Advantage Should Do

The practical implication of everything Goodhart described is that the most important AI investment a business can make in 2026 is not model selection. It is system architecture. And system architecture has several components that are more durable than any model version cycle.

The first is data. The insight that the AI research community at IBM, Anthropic, and across the enterprise AI market has converged on is that the competitive advantage in AI is not the model — it is the data the model runs on and the business logic that governs its behaviour. An AI system running on a company's proprietary customer interaction data, internal process knowledge, and institutional expertise will outperform a generic AI system running on the same underlying model. That advantage does not disappear when a new model version is released, because it is grounded in data and context that the company uniquely owns. As one Bridgenext senior architect put it in the context of enterprise AI strategy, "the true competitive advantage will belong to the enterprises that have meticulously documented, secured and exposed their proprietary business logic and systems as high-quality, agent-callable APIs."

The second component is workflow integration. The gap between a company that has deployed AI tools in isolated use cases and a company that has integrated AI into end-to-end workflows is not a technology gap. It is an orchestration gap. Goodhart's point about cooperative model routing applies equally at the workflow level: the organisation that has designed its AI systems to hand tasks between models, between human reviewers and automated agents, and between different specialised tools in a coordinated sequence has built something that competitors with better model subscriptions but weaker orchestration cannot match.

The third component is governance. A McKinsey 2025 State of AI report found that while individual AI adoption was widespread, fewer than a third of organisations had successfully scaled AI into genuine enterprise-wide transformation. The gap was not capability. It was governance — the absence of standards for how AI outputs are reviewed, how errors are caught, how models are updated, and how responsibility is assigned when an AI-assisted decision produces a bad outcome. Enterprises that establish those standards early build the institutional infrastructure that makes AI safe to scale. Enterprises that skip that step accumulate operational risk that compounds with every additional AI deployment.

Conclusion

Gabe Goodhart's statement that the model is no longer the main differentiator is not a prediction about the distant future. It is a description of the present. The AI model market in 2026 is a buyer's market, as he put it, and the businesses that are still treating model selection as their primary AI strategy are optimising for a competitive advantage that no longer exists at the model layer and never will again.

The competition has moved up the stack. What IBM is calling system competition — the combination of models, tools, workflows, data, and orchestration architecture — is the actual battleground for durable AI advantage. Cooperative model routing, where smaller models handle routine tasks and escalate to more capable models when genuinely necessary, is the pattern that makes that competition economically sustainable. And the system engineering skills required to build a well-governed, well-integrated, cost-efficient AI architecture are far rarer than the ability to write a good API call to a frontier model.

The businesses that will win the AI era are not the ones with the best model subscription. They are the ones that understood, early enough, that the model was always a component — and that the system around it was always the product.