Anthropic Takes 2:1 Enterprise Lead as Decision Models Hit 25%

(SAN FRANCISCO —) New enterprise engineering teams onboarding onto Vercel's AI gateway selected Anthropic's models twice as often as OpenAI's over the past month, according to production routing telemetry disclosed by Vercel AI product lead Harpreet Arora on the MTS podcast and reported by BigGo Finance. Sitting directly in the inference path between foundation model laboratories and enterprise applications, Vercel's gateway records actual API procurement rather than self-reported developer sentiment, revealing a decisive October-to-November swing back toward Anthropic following the release of Claude Opus 5.5 and Fable 5.2.

Yet the headline 2-to-1 ratio masks a deeper structural transformation in how production AI systems are architected. Across the same gateway traffic, specialized "decision models"—a category virtually absent from enterprise architectures in January—now process roughly 25% of all requests, while open-weight models command approximately 25% of total spend. Rather than locking into a single frontier vendor, engineering organizations are decomposing monolithic agent loops into tiered, multi-provider pipelines where frontier models explore novel capabilities in sandboxes and cheaper, lower-latency runtimes execute the resulting workflows at scale.

Technical Architecture & Engineering Breakdown: The Scout-and-Workhorse Routing Paradigm

To understand why gateway telemetry diverges so sharply from public model-war narratives, systems architects have to examine how an enterprise AI gateway functions as an abstraction layer. Positioned between application backends and upstream model endpoints, the gateway normalizes authentication, failover, rate-limiting, and cost attribution across closed APIs and hosted open-weight clusters. Because every production token traverses this control plane, the resulting telemetry reflects hard architectural commitments rather than marketing mindshare.

Arora's primary architectural insight is that frontier model releases and low-cost model adoption operate as a coupled expansion loop rather than a zero-sum battle. When a lab ships a major capability leap—such as Opus 5.5 or Fable 5.2—engineers do not immediately wire that expensive frontier endpoint into high-volume production cron jobs. Instead, they treat the frontier model as an exploratory scout inside an isolated sandbox, probing complex multi-step problems they previously considered intractable. As organizations formalize these experimental environments, mitigating agent sandbox execution risks becomes a prerequisite before any discovered workflow can graduate to production.

Decomposing Monolithic Prompts into Decision Models

Once an engineer validates a repeatable execution plan over a handful of sandbox runs, the architecture undergoes immediate cost-down refactoring. Deterministic or bounded subtasks—such as customer-support triage, go-to-market pipelines that scrape LinkedIn profiles, or automated social-signal scanners—are stripped away from the frontier model and offloaded to smaller family members like Claude Haiku or open-weight alternatives. Because the frontier release expanded the total surface area of tasks the enterprise attempts, the downstream execution tier grows continuously even during months when frontier progress plateaus.

This exact refactoring pressure explains the explosive rise of "decision models," which grew from zero visibility at the start of the year to handling roughly one-quarter of all Vercel gateway requests. Historically, early agentic applications routed every node in a directed acyclic graph (DAG) through a general-purpose flagship LLM, paying steep latency and token penalties just to classify intent, validate schema compliance, or trigger a binary branch. In modern enterprise stacks—including structured orchestration layers like Microsoft Copilot Studio—teams are replacing general-purpose reasoning calls with narrow decision models purpose-built for fast routing, state classification, and gating.

Benchmark Evaluation & Rigorous Comparison: Market Swings vs. Production Feel

While Anthropic currently holds a 2-to-1 advantage among newly onboarded gateway teams, Arora explicitly cautions against treating that lead as permanent. The 2026 telemetry record reads like a pendulum: Anthropic's Opus 4.5 release sparked a dominant run through April and May, OpenAI clawed back share in the subsequent cycle, Anthropic experienced a noticeable lull across August and September, and then Opus 5.5 alongside Fable 5.2 pulled engineering teams back into Anthropic's orbit in October and November.

Critically, Vercel's data exposes a sharp split between "share of headlines" and "share of requests." On the OpenAI side, flagship conversations often center on visual and spatial breakthroughs like Astra, which Arora credits with unlocking entirely new workflows in tools such as Blender. Yet in raw request volume, one of OpenAI's quietest performers is the compact GPT-6 Luna series (now at version 6.1). Luna rarely dominates synthetic benchmark tables, but mobile application developers gravitate toward it for its speed, low cost, and responsiveness.

Why Model "Personality" Overrides Static Leaderboards

Arora argues that developers select production models much the way teams hire colleagues—evaluating temperament, communication style, and predictability alongside raw capability. If autonomous agents are designed to function as digital knowledge workers, engineering leads do not want every worker exhibiting identical phrasing, verbosity, or failure modes. Subtle differences in how a model interprets terse prompts, formats structured outputs, or recovers from tool-call errors determine which endpoint survives code review and ships to users.

Because Vercel's gateway captures live routing volume and spend rather than synthetic academic evaluations, public SWE-bench, HumanEval, and MMLU-Pro figures for these newly shipped commercial tiers are not reported in the gateway dataset. Instead, the matrix below synthesizes the verified gateway telemetry, architectural roles, and weight availability across the primary model classes competing for enterprise traffic.

Deployment, VRAM Footprint & Inference Optimization: Sovereign Runtimes and Modality Shifts

One of the most counterintuitive findings in Vercel's telemetry involves the geography of open-weight inference. Open-weight models now capture roughly a quarter of all gateway spend, and the most capable frontier releases in that category—notably Kimi K3 and GLM 5.3—have largely originated from Chinese AI labs. Yet approximately 70% (described interchangeably by Arora as two-thirds) of enterprise open-weight spend lands with US-based inference providers.

For regulated industries—including healthcare, financial services, and enterprise SaaS—model provenance is secondary to runtime jurisdiction. Compliance teams routinely approve open-weight architectures only when weights are pulled into sovereign, US-resident GPU clusters with strict zero-retention guarantees, mirroring the broader industry push toward in-VPC agent secrets and self-hosted execution. By decoupling weight origin from inference hosting, enterprises capture open-weight unit economics without exposing customer payloads to foreign API endpoints.

The Open-Weight Supply Bottleneck and NVIDIA's Retreat

Despite durable enterprise demand, Arora flags a troubling supply-side slowdown in the open-weight ecosystem. While Kimi K3 and GLM 5.3 delivered genuine step-function improvements, the months since have lacked an equivalent leap, suggesting that leading closed labs have unlocked internal post-training or synthetic data techniques that open-weight teams have yet to replicate. Compounding that gap is NVIDIA's apparent pullback from frontier open-weight development—a notable shift for a $5 trillion company whose CEO Jensen Huang had championed open-weight models globally while investing in Nemotron and recruiting talent from Poolside.

If the open-weight quality curve remains flat while closed labs iterate rapidly and optimize specialized pricing—such as high-margin specialized API tiers—open weights risk hitting a 25% spend ceiling rather than expanding further. Still, architectural diversification is accelerating across the stack: xAI is actively rewriting its Grok stack off LangChain onto a custom inference engine, Reflection and Thinking Machines are pushing alternative open-weight paradigms, and Meta's Muse models are emerging as challengers to the Anthropic–OpenAI duopoly. Simultaneously, Anthropic's anticipated Nasdaq IPO under the ticker ANT—reportedly targeting a valuation above $1.2 trillion—will soon force public markets to price frontier lab economics directly.

Voice Hits 1-in-9 Teams as Video Pipelines Loom

Beyond text and code routing, Vercel's gateway data highlights a rapid modality shift. Voice interfaces surged from 0% at the beginning of 2026 to roughly one in nine enterprise teams on the gateway. Inside Vercel itself, engineers increasingly hold down a command key to dictate unstructured thoughts, relying on audio-capable models to synthesize rambling speech into clean technical specifications. Arora projects that voice adoption will continue climbing while video generation and understanding models—currently too unpredictable for repeatable production loops—will cross the reliability threshold into enterprise workflows within the next 12 months.