A year ago, there was clear daylight between OpenAI’s GPT-4 and everything else. Today, that gap has nearly closed. Anthropic’s Claude 3.5, Google’s Gemini, Meta’s Llama 3, and a dozen other models perform comparably on most tasks.

This commoditization is happening faster than almost anyone predicted—and it has significant implications for how enterprises should think about AI strategy.

The Evidence of Convergence

Look at benchmark performance across major models:

• On standard coding tasks, multiple models score within 5% of each other
• On complex reasoning, the gap between top performers is narrowing quarterly
• For typical business applications, users often can’t distinguish outputs

Open-source models have closed much of the gap with proprietary offerings. Llama 3 performs comparably to GPT-4 on many tasks, at a fraction of the inference cost.

Chinese models from Alibaba, Baidu, and startups like Moonshot and DeepSeek are reaching near-frontier performance, creating additional competitive pressure.

What Drove This Speed

Several factors accelerated commoditization:

Research Publication

Despite concerns about AI safety, most significant advances still get published. Techniques developed at leading labs become available to competitors within months.

Open-Source Investment

Meta’s decision to open-source Llama created a foundation that hundreds of organisations are building on. Mistral, Cohere, and others release competitive models under permissive licenses.

Inference Optimisation

Smaller, more efficient models can match larger models’ output quality for specific tasks. Specialised fine-tuning means you don’t always need the largest possible model.

Data Accumulation

As more organisations deploy AI, more feedback data becomes available for training. The incumbents’ data advantage is eroding.

What This Means for Enterprises

Don’t Bet on One Provider

Building deeply into any single model provider’s ecosystem creates unnecessary lock-in. The provider that’s best today might not be best next year—and the switching costs are lower than you might think.

Design applications with model abstraction. The prompt and orchestration logic should be separable from the specific model serving the request.

Application Value Matters More Than Model Choice

As models converge, the differentiation shifts to:

• How well you understand your specific use case
• How effectively you integrate AI into existing workflows
• The quality of your data and feedback loops
• Your ability to evaluate and improve outputs

Companies building custom AI solutions increasingly compete on application design, not model selection.

Evaluate on Your Specific Tasks

Benchmark comparisons don’t tell you which model works best for your use case. A model that performs 3% worse on general benchmarks might perform 20% better on your specific domain.

Test candidates on actual tasks with actual users. The results often surprise people.

Cost Becomes a Real Factor

When quality is comparable, cost matters. Inference pricing varies significantly:

• Major providers (OpenAI, Anthropic, Google): Premium pricing, highest reliability
• Mid-tier providers: Lower cost, generally reliable
• Open-source deployment: Lowest ongoing cost, requires infrastructure

For high-volume applications, the savings from moving to alternative providers or self-hosted models can be substantial.

What This Means for AI Companies

Frontier Capability Is Necessary But Not Sufficient

Building the best model doesn’t guarantee market success. Distribution, integration, enterprise relationships, and trust all matter.

OpenAI has first-mover advantage and brand recognition. Anthropic has a safety reputation and strong enterprise positioning. Google has distribution through Cloud. These advantages don’t evaporate just because models converge.

But being slightly behind on benchmarks is increasingly survivable. The “winner take all” scenario looks less likely.

Differentiation Shifts to Vertical Solutions

Horizontal model providers increasingly compete with vertical-specific solutions:

• Legal AI companies training on case law and legal workflows
• Healthcare AI companies with clinical data and regulatory compliance
• Financial AI companies with market data and risk frameworks

These specialised players don’t need frontier capabilities on general tasks. They need to be exceptional at their specific domain.

Pricing Pressure Is Real

Open-source alternatives cap what commercial providers can charge. As capabilities converge, pricing power decreases.

This creates tension: continued research investment requires significant revenue, but competitive pricing limits that revenue. Expect consolidation among model providers over the next 2-3 years.

The Strategic Response

For enterprises investing in AI:

1. Build portable applications - Don’t assume your current model provider is permanent
2. Invest in evaluation capability - You need to assess which models actually work for your tasks
3. Focus on data and workflow - These create durable advantage; model choice doesn’t
4. Watch open-source closely - The cost-quality trade-off is changing rapidly
5. Consider multi-model architectures - Different models for different tasks based on cost and capability

For AI companies:

1. Specialise or scale - The middle ground between OpenAI scale and vertical focus is uncomfortable
2. Build ecosystem lock-in carefully - Some lock-in is inevitable; extractive lock-in alienates customers
3. Prepare for pricing pressure - Current margins probably aren’t sustainable
4. Find defensible differentiation - Pure model capability is not a moat

Looking Forward

Commoditization doesn’t mean AI becomes unimportant—the opposite. As basic AI capabilities become widely available, the question shifts from “can we use AI?” to “how well can we use AI?”

Winners will be organisations that apply AI effectively to real problems, with strong data foundations, clear use cases, and continuous improvement processes.

The model behind the application matters less than the application itself.

This is normal technology evolution. Database technology commoditized; value shifted to applications built on databases. Cloud infrastructure commoditized; value shifted to services running on cloud. AI models are following the same pattern.

The exciting work isn’t building a slightly better model—it’s building things that genuinely help with real problems.

That’s where the opportunity lies now.