Enterprise AI competing on control
Vendors are shifting the frontier: enterprise buyers now choose models for governance, cost and monitoring rather than headline performance, and Anthropic has made its model economics public so customers can compare. (siliconangle.com) Anthropic’s published pricing lets procurement teams evaluate trade-offs in a way that rewards auditability, permissions and cost predictability. (platform.claude.com)
For big companies buying artificial intelligence in 2026, the first question is no longer “Which model wins the benchmark?” It is “Who can keep 5,000 employees, 20 internal tools, and one finance team from losing control of the bill?” (siliconangle.com) That is why Anthropic’s April 9 update focused on “organization-wide controls” for Claude Cowork instead of just raw model performance. Claude Cowork, launched in January, is an autonomous agent that can organize files, create reports, and run browser tasks on a worker’s computer. (siliconangle.com) Once software can click around a browser and edit files, enterprise buyers start asking for the same things they ask from payroll software or cloud servers. Anthropic says Enterprise admins can now use role-based access controls to decide which Cowork capabilities each employee can use. (siliconangle.com) The next problem is not intelligence but visibility. Anthropic said it is expanding OpenTelemetry support so companies can send events like tool calls and file modifications into their security information and event management systems. (siliconangle.com) The bill is the other half of the argument. Anthropic says admins can set group spend limits for teams, and its earlier enterprise controls let admins cap extra per-user usage at standard application programming interface rates so finance teams can predict monthly costs before a rollout spreads. (siliconangle.com) (anthropic.com) Anthropic is also making that math unusually public. Its official pricing pages list Claude Haiku 4.5 at $1 per million input tokens and $5 per million output tokens, Claude Sonnet 4.6 at $3 and $15, and Claude Opus 4.6 at $5 and $25. (anthropic.com 1) (anthropic.com 2) (anthropic.com 3) That matters because procurement teams can compare models the way they compare cloud storage or database tiers. Anthropic also advertises up to 90% savings from prompt caching and 50% savings from batch processing, which turns “How smart is it?” into “How much will this workload cost at scale?” (anthropic.com 1) (anthropic.com 2) (anthropic.com 3) OpenAI is moving in the same direction, but with a different sales shape. OpenAI’s help center says ChatGPT Enterprise workspaces disable apps by default, let admins assign app access through role-based access control, and now support flexible seat types including a Codex-only seat added on April 2, 2026. (help.openai.com 1) (help.openai.com 2) OpenAI is also cutting prices where heavy users feel them fastest. SiliconANGLE reported that OpenAI introduced a $100-per-month Pro tier for Codex users, down from the prior $200 level for its top subscription, while Anthropic’s top Claude Code subscription still starts at $200 per month. (siliconangle.com) Underneath the product launches, both companies are selling the same promise: not just a smart model, but a governable one. OpenAI’s platform page highlights role-based access controls, billing and usage alerts, single sign-on, and zero data retention by request, while Anthropic’s trust materials highlight single sign-on, System for Cross-domain Identity Management, audit logs, and role-based permissions. (openai.com) (anthropic.com) So the frontier for enterprise artificial intelligence is shifting from demos to controls. When a vendor publishes prices, logs actions, limits permissions, and lets a security team watch every tool call, it stops looking like a magic trick and starts looking like software a Fortune 500 company can actually buy. (anthropic.com) (openai.com)
