Comparative LLM Usage Across Sectors

Narrative

Developer usage of LLMs, as captured by the Stack Overflow 2025 Developer Survey (49,000+ respondents across 177 countries), is heavily concentrated at the top. OpenAI GPT models lead with 81–82% developer adoption, followed by Claude Sonnet models at 43–45% and Gemini Flash at 35%. The survey, the first edition to ask about specific LLMs rather than AI tools in general, also found that 84% of developers use or plan to use AI tools, yet 46% distrust the accuracy of output - a trust deficit that grew 15 percentage points year-on-year. The 2025 DORA State of AI-assisted Software Development report, drawn from nearly 5,000 survey responses and 100+ hours of qualitative interviews, confirmed near-universal adoption: 90% of respondents reported using AI in daily software development work, spending a median of two hours per day on it. Critically, DORA found that AI functions as a multiplier of existing team conditions - strengthening high performers while amplifying dysfunction in weaker organisations - not as a uniform productivity lever.

The Thoughtworks Technology Radar shifted markedly across its two 2025 editions. Volume 32 (April 2025) highlighted retrieval-augmented generation and prompt engineering as emergent practices. By Volume 33 (November 2025), the signal had moved to context engineering, the Model Context Protocol, and growth of agentic systems - a transition Thoughtworks CTO Rachel Laycock described as a step change in industry thinking. Separately, InfoQ's practitioner coverage tracked the same arc: agentic AI moved from architectural curiosity to a distinct engineering discipline requiring specification-driven development, atomic decomposition, and observable workflows. A 2025 practitioner study surveying 306 engineers across 26 domains found that production agents are typically constrained: 68% execute at most 10 steps before requiring human intervention, and 70% rely on prompting off-the-shelf models rather than weight tuning, with reliability cited as the primary development challenge. The ZenML analysis of 1,200 production LLM deployments reinforced this: the organisations extracting real value were those investing in evaluation pipelines and infrastructure-based guardrails, not those with the most impressive demos.

Deployment architecture is fragmenting into a well-defined set of patterns. Cloud-managed services (Azure OpenAI, Bedrock, Vertex AI) remain dominant for enterprise teams lacking inference infrastructure. Self-hosted inference via vLLM is consolidating as the production-grade open alternative: Stripe reportedly achieved a 73% inference cost reduction by migrating 50 million daily API calls to vLLM on one-third of a prior GPU fleet. Ollama serves prototyping and local development, with practitioners citing a clear migration path from Ollama to vLLM when real users appear. The Thoughtworks Radar noted a growing set of organisations moving toward self-hosted, governed deployments to address multi-tenancy, access control, and data residency requirements. Open-weight adoption is expanding but remains below closed-model share in large enterprises: a16z's 2026 enterprise CIO survey found that 23% of large enterprises were running OpenAI's o3 in production compared to 3% for DeepSeek, though open-weight models - particularly Llama 4, DeepSeek V3/V4, and Qwen 3.x - are gaining faster in regulated sectors where third-party API calls face procurement or compliance obstacles.

LLM cost measurement is undergoing rapid professionalisation under the banner of "TokenOps", a FinOps-adjacent discipline that applies visibility, allocation, and optimisation principles to token consumption. Per-token prices fell approximately 80% across the major provider families between early 2025 and early 2026, yet enterprise AI spend still rose: CloudZero's State of AI Costs report measured average monthly AI spend growing from $63,000 in 2024 to $85,500 in 2025, a 36% increase. The a16z 2025 enterprise survey found average LLM spend growing from $4.5 million to $7 million over two years, with CIOs forecasting $11.6 million by end of 2026. The primary driver of spend growth is not per-token price but volume expansion driven by agentic loops: agentic workflows trigger 10–20 LLM calls per user task versus single-query chatbots, with one analysis citing a Gartner figure that agentic models require 5–30x more tokens per task than standard chatbots. The evidence base for published enterprise spend figures remains thin and largely vendor-driven; independently verified budget data at organisation level is almost absent, and ROI claims from enterprises themselves remain, as a16z noted, "less dramatic than one might expect."