LLMs

Decision models are becoming commodities—customization emerges as the differentiator

As decision-model adoption accelerates, startups and hyperscalers are competing on data-tuning capabilities rather than base-model performance.

Last verified:

Decision models capture over one-quarter of Vercel router traffic

According to TechCrunch AI, more than one in four requests on Vercel’s AI model router were directed to decision models in October 2026—a striking concentration for a category that emerged only weeks prior. Decision models are LLMs tuned to quickly select between preset options and register confidence in the choice, rather than generate unbounded reasoning text. This design trades reasoning depth for speed and cost, making them ideal for production environments where fast, deterministic decisions outweigh elaborate explanations. TypeSafe CEO Diogo Almeida underscored that the company’s Jev family of models rose to dominance through meticulous calibration using statistically-adjusted synthetic data, establishing a template other builders would follow.

Hyperscalers and startups flood the market

The barrier to entry for decision models proved remarkably low. According to TechCrunch AI, one Amazon engineer built a decision model in a weekend, and within weeks, every major cloud provider released its own variant. On November 16, 2026, Microsoft CEO Satya Nadella announced the rollout of Microsoft-Decision-1.1 across Azure, following the company’s careful tuning of cost, latency, intelligence, and calibration metrics under engineering VP Achint Srivastava’s direction. Simultaneously, U.K. startup Sqwish, founded by Cambridge grad students Ushnish Sengupta and Federica Freddi, who had been developing small LLMs when Jev arrived, announced a $3 million pre-seed round. TechCrunch AI reports that Sqwish’s kettei fat model has performed competitively on some benchmarks, but Sengupta emphasized the company’s focus on post-training customization rather than raw benchmark scores.

Post-training customization emerges as the competitive moat

Commodity pressure is mounting across the decision-model category, forcing differentiation beyond base-model performance. According to TechCrunch AI, Sqwish’s core pitch centers on built-in tooling that allows customers to upload their own production data for continuous model improvement—a capability absent in most competitors. Sengupta told TechCrunch AI: “What makes us not a commodity, I think, is the fact that we are offering our customers the same power…they can keep uploading their production data, and just the same way that we trained our models, their models can also keep getting better at the tasks that they care about.” TechCrunch AI notes that most decision-model builders agree the real validation comes from production context switching rather than published benchmarks. Sengupta framed benchmarks as “an invitation to try the model out, and then the continual tuning is the reason why they should stick to us.”

Why This Matters

For ML operations leaders and platform teams running high-volume classification workloads, the decision-model market is transitioning from a capability race (baseline accuracy) to a retention game (continuous customization). TypeSafe’s dominance (1.5B+ weekly queries on OpenRouter, according to TechCrunch AI) is now vulnerable to competitors offering superior post-training data infrastructure—meaning vendor selection will increasingly hinge on API-native fine-tuning capabilities rather than published benchmark leads. Startups like Sqwish are betting that the ability to privately improve models on customer data will outweigh the distribution advantages of Microsoft or cloud incumbents. By Q2 2027, expect decision-model contracts to be won and lost on customization efficiency (cost per fine-tuning iteration, convergence speed) rather than base-model quality.

Frequently Asked Questions

What is a decision model?

A decision model is an LLM optimized to select between preset options and report confidence in its choice, without generating unbounded output tokens. They are faster and cheaper than reasoning models, making them ideal for high-volume classification workloads in production.

Why did decision-model adoption explode so quickly?

TypeSafe's Jev family, released in October 2026, demonstrated the category's utility and cost advantages. Within weeks, hyperscalers (Amazon, Microsoft, Google) and startups released competing models, lowering barriers to entry and accelerating adoption across platforms like Vercel.

How do companies differentiate in a crowded decision-model market?

Benchmark performance is table stakes. Competitive advantage is shifting to post-training customization—the ability to fine-tune models on a customer's own production data to improve task-specific accuracy and calibration.

Who is winning the decision-model market so far?

TypeSafe's Jev family remains dominant by usage metrics (1.5B+ queries per week on OpenRouter). Microsoft has entered with Decision-1.1 on Azure; Sqwish is differentiating on built-in customer data tuning after securing a $3M pre-seed round.

#decision-models #customization #TypeSafe #Microsoft #Sqwish #classification #cost-optimization