Conventional wisdom says enterprises choose AI models based on their current and potential capabilities. The market says otherwise. Anthropic now commands 40% of enterprise LLM spend versus OpenAI's 27%, a complete reversal from 2023. The reason isn't that Claude is smarter. It's that Claude is more predictable.

In coding alone, Anthropic's lead is starker still. The company holds 54% market share versus OpenAI's 21%, according to Menlo Ventures' December 2025 report.

Simon Smith, EVP of Generative AI at Klick Health, captured the practitioner experience on X recently. He keeps "turning to Claude for a lot of business output" because its writing "wasn't as negatively affected by its intelligence increase." That's the user-level signal behind the market-level shift.

The consistency gap

Smith's observation cuts to a problem enterprise IT leaders are discovering when it comes to selecting the best models for their organizations. That’s the growing gaps attributed to personality drift. OpenAI's rapid release cadence, with GPT-5.2 launched just one month after 5.1, creates instability that's manageable for consumers but challenging and potentially costly for businesses with established workflows.

"I find it more mechanical than its predecessor," Smith wrote of GPT-5.2. "Yes, I can try and tune its personality, but I increasingly find that unsatisfying."

For individual users, retuning prompts is an annoyance. For enterprises with thousands of employees running standardized AI workflows, it's a procurement risk and an unforeseen time sink that can negate any AI-based productivity gains.

Anthropic's releases tell a different story. Each upgrade maintained behavioral consistency while improving capability. This path reflects the opposite of consumer-oriented personality refreshes.

The greater the safety rigor, the greater the reliability

The connection between Anthropic's safety investments and output reliability isn't coincidental. It's architectural, and it's reflected in their red teaming process.

VentureBeat's analysis of red teaming approaches revealed a fundamental methodological split between the two companies. Anthropic's 153-page system card for Claude Opus 4.5 documents multi-attempt attack success rates from 200-attempt reinforcement learning campaigns. OpenAI's 60-page GPT-5 system card reports single-attempt jailbreak resistance.

Anthropic monitors around 10 million neural features during evaluation using dictionary learning. These features map to human-interpretable concepts including deception, sycophancy, and bias. The same infrastructure that catches safety issues catches behavioral inconsistencies.

The foundation is Constitutional AI, Anthropic's training methodology that gives models explicit principles rather than relying solely on human feedback. That transparency produces predictability. Enterprises can audit what principles guide model behavior rather than spend valuable time reverse-engineering implicit values from inconsistent outputs.

Growing evidence in enterprise accounts

"Anthropic prioritized safety and security a lot more than other LLMs," said Gunjan Patel, Director of Engineering at Palo Alto Networks, which deployed Claude across 2,500 developers. "They discuss security implications in every meeting. As the largest cybersecurity company, that's a big deal for us." The cybersecurity company saw a 20-30% increase in feature development velocity after choosing Claude. Junior developers completed integration tasks 70% faster with Claude's assistance.

Novo Nordisk, creator of Ozempic, says it streamlined pharmaceutical documentation with Claude. Clinical study reports that took 10+ weeks now take 10 minutes — a 90% reduction.

IG Group hit full ROI within three months on its AI initiative. And GitLab's evaluation found that Claude "stood out for its ability to mitigate distracting, unsafe, or deceptive behaviors." That's safety language describing reliability outcomes.

The momentum is accelerating. Anthropic recently announced a partnership with Accenture, with 30,000 professionals trained on Claude, making it one of the largest AI practitioner ecosystems globally. The company has grown from under 1,000 to over 300,000 enterprise customers in two years, with its applied AI team expanding fivefold to support deployments at scale.

Where OpenAI’s strength is most apparent

OpenAI isn't losing the enterprise market. The company retains significant advantages that matter to specific buyer segments.

  • Ecosystem depth: OpenAI's plugin architecture, custom GPTs, and third-party integrations create switching costs that Anthropic hasn't matched. Enterprises already embedded in the OpenAI ecosystem face significant friction if they choose to migrate.

  • Multimodal capabilities: Native image generation, voice interaction, and video understanding give OpenAI an edge for enterprises building consumer-facing AI products where those capabilities matter.

  • Brand gravity: ChatGPT's 2.5 billion daily prompts create familiarity. When enterprise employees already use ChatGPT personally, IT departments face less change management resistance.

  • Reasoning models: Klick Health's Smith acknowledged that "GPT-5.2 Thinking and especially Pro are very smart. For research and reasoning tasks, including strategic planning, I'd bet they're the smartest models available."

Capability still wins some deals. Reliability wins others.

OpenAI's two-front problem

OpenAI's challenge is structural. Smith put it precisely in his X post. "To one side, Anthropic is laser-focused on enterprise. To the other, Google owns consumers with distribution."

OpenAI built its release cadence around capability leaps and consumer engagement. Enterprises need the opposite, starting with predictable behavior, compliance-ready architectures, and operational stability.

"AI is fundamentally rewiring how enterprises buy software," said Joff Redfern, Partner at Menlo Ventures. "Deals close at twice the speed of traditional SaaS, and startups are capturing two dollars for every one incumbents earn."

What this all means for enterprise AI buyers

This isn't a question of which vendor is best. Different development philosophies produce different operational characteristics. Match vendors to use cases.

Buyers evaluating AI in 2026 should ask questions that benchmark scores can't answer.

  • Release stability: How does model behavior change between versions? What's the deprecation policy for capabilities your workflows depend on?

  • Deployment flexibility: Does the vendor support AWS Bedrock and Google Vertex AI? Or only proprietary infrastructure?

  • Compliance documentation: A 153-page system card enables procurement conversations that a 60-page card cannot.

  • Applied AI support: Hands-on deployment teams for enterprise-scale implementations? Or just API documentation?

  • Data sovereignty: For regulated industries, where does inference happen? Who controls the data?

Most enterprise AI initiatives don't stall on model capability. They stall on implementation complexity. Predictable behavior and robust support infrastructure fix that.

What to watch in 2026

Three dynamics will shape the next 12 months.

The stability tax

OpenAI has to decide whether enterprise revenue justifies slowing its release cadence. The April sycophancy rollback showed what happens when consumer optimization backfires.

OpenAI pushed a GPT-4o update intended to make the model "more intuitive and effective," then had to reverse it within days after users reported the model had become excessively flattering and disingenuous. Consumer users found it annoying. Enterprise customers with thousands of standardized workflows found it operationally disruptive. Anthropic built its release process around backward compatibility. OpenAI would have to restructure its development philosophy to match, not just its marketing.

The support scaling problem

Anthropic grew from under 1,000 to over 300,000 enterprise customers while expanding its applied AI team fivefold. That ratio can't hold. Either Anthropic builds partner channels fast, or enterprise deployments start stalling on implementation support. The Accenture partnership signals that the company know this.

The open-source wild card

Llama and DeepSeek aren't competing on reliability yet; they're competing on cost and control. But the gap is closing. When open-source models reach "good enough" reliability for production workloads, the entire pricing structure shifts. Enterprises will run reliability-critical workflows on Claude while commoditizing everything else.

The vendors who win in 2026 won't be the ones with the highest benchmark scores. They'll be the ones who figured out that enterprise AI is an operations problem, not a capability problem.

The bottom line

Anthropic's enterprise dominance wasn't inevitable. Two years ago, OpenAI held 50% market share to Anthropic's 12%. The reversal happened because safety-first development produced the operational characteristics enterprises need. Consistent outputs, predictable behavior, and auditable decision-making are proving to be table stakes.

Smith's post on X captured the user experience. The market data confirms the purchasing behavior. The customer deployments validate the operational reality.

The AI market is learning what enterprise software markets learned decades ago. Capability gets you in the door, but reliability wins the contract and renewals.