Enterprise AI Budgets Jump 75% as Reasoning Models Hit Production at Scale

Budget Growth Outpaces Previous Enterprise IT Cycles

Enterprise AI spending will grow 75% on average over the next 12 months, according to Andreessen Horowitz's latest survey of enterprise technology leaders. That rate exceeds typical enterprise software expansion and signals a permanent budget reallocation, not a pilot-phase experiment. The same data shows 23% of surveyed enterprises already run OpenAI's o3 reasoning model in production—a model released less than a quarter ago. When nearly one in four buyers moves a reasoning-class model into production workflows within weeks of availability, the enterprise AI market has crossed from evaluation into operational dependency.

The spend increase concentrates in two areas: software development and multi-model deployments. Over 70% of enterprises now use AI for software development in production, up from under 40% a year earlier. At the same time, 37% of enterprises deploy five or more different models simultaneously. Buyers are not standardizing on a single vendor. They are building multi-model stacks where each model handles specific tasks based on cost, latency, and reasoning depth.

Anthropic's Claude 3.7 Sonnet Targets the Mid-Tier Reasoning Gap

Anthropic released Claude 3.7 Sonnet this month, positioning it between the cheaper Haiku and the premium Opus models. The timing matters because enterprises now face a cost problem: reasoning models deliver measurably better results for complex workflows, but running every task on top-tier models drives token costs higher than many budgets can sustain at scale. Sonnet 3.7 is designed as the daily-driver model for software engineering, analytics, and multi-agent orchestration—tasks that need reasoning but do not justify ultra-premium pricing on every API call.

Menlo Ventures' 2025 State of Generative AI report describes Claude 3.5 Sonnet (the predecessor to 3.7) as the point where code generation reached "economically meaningful performance." Enterprises in that study reported step-change productivity in coding workflows, including new code generation, refactoring, and test creation. Claude 3.7 Sonnet extends that capability while maintaining mid-tier pricing, directly competing with OpenAI's GPT-4.1 and Google's Gemini 1.5 Pro in the same cost band.

The competitive landscape is stabilizing around a three-vendor pattern. OpenAI still leads in enterprise mindshare, but Google and Anthropic have made significant share gains over the past year, according to the A16Z survey. Among OpenAI customers, 67% now deploy multiple OpenAI models simultaneously, suggesting that even single-vendor buyers are adopting multi-model strategies within one provider's portfolio. Claude 3.7 Sonnet gives enterprises a credible second source for reasoning workflows, reducing dependence on any single vendor's roadmap or pricing changes.

Model Strategy Shifts from Fine-Tuning to Off-the-Shelf Reasoning

Enterprise model strategy is moving away from heavy fine-tuning and toward using general-purpose reasoning models out of the box. The A16Z data shows enterprises now prioritize models that work well without fine-tuning, a reversal from the pervasive fine-tuning pattern of two years ago. This shift has budget and operational implications. Fine-tuning requires data labeling, retraining pipelines, version management, and ongoing evaluation—all of which add cost and latency to model updates. Off-the-shelf reasoning models eliminate most of that overhead while delivering comparable or better performance on complex tasks.

For buyers, this changes the build-versus-buy calculation. If a general-purpose reasoning model like Claude 3.7 Sonnet or OpenAI's o3 can replace a fine-tuned stack, the operational simplification is immediate. Teams that currently maintain fine-tuned models for code generation, data analysis, or internal documentation should run bake-offs comparing their existing pipeline to current-generation reasoning models. In many cases, the reasoning model will match or exceed the fine-tuned version while cutting operational complexity.

The shift also affects vendor negotiation. When buyers relied on fine-tuning, they faced high switching costs—moving providers meant rebuilding the entire training pipeline. With off-the-shelf reasoning models, switching costs drop. Enterprises can run parallel evaluations across OpenAI, Anthropic, and Google models on the same task and select based on performance, price, and latency. That optionality gives buyers more leverage in pricing discussions and SLA negotiations.

What to Watch: Reasoning Models Increase Risk as They Gain Capabilities

Reasoning models are accelerating adoption, not slowing it—57% of surveyed enterprises say reasoning models are speeding up their AI rollout. But higher reasoning capability creates higher-impact risks. These models can chain actions, call APIs, modify code, and interact with production systems. An error or misalignment in a reasoning model's decision chain can propagate through multiple downstream systems before a human intervenes.

Enterprises deploying reasoning models need stricter guardrails around API access, data scope, and action approval. For software development use cases, that means logging every AI-generated code change, requiring human approval for production commits, and maintaining rollback procedures for AI-driven modifications. For agent workflows, it means defining explicit boundaries on what systems an agent can touch and what actions require escalation.

Buyers running OpenAI, Google, or Anthropic models should treat the next 90 days as a reset point for AI governance policies. If your current policy was written for chat-based assistants or simple summarization, it does not cover reasoning models that write production code or orchestrate multi-step workflows. Update policies now, before reasoning models become embedded in enough workflows that retroactive policy enforcement becomes politically difficult.