GPT-5 for business: evaluate the workflow, not the launch claim

OpenAI introduced GPT-5 on 7 August 2025. Its announcement described a unified ChatGPT system combining a fast model, a deeper reasoning model and a router that selected between them. The accompanying business announcement connected the release to workplace adoption and API applications. This analysis is dated 25 October 2026 and looks back at that launch. It does not describe GPT-5 as today’s newest model, reproduce historical plan access as a current entitlement or establish present-day prices. 1

The central business question is not whether the launch was impressive. It is whether the particular system an organisation deploys produces more accepted work under its own constraints.

The launch description matters because ChatGPT’s router could choose different processing paths according to the conversation, task complexity and tool needs. A person testing the chat application was therefore evaluating an experience, not necessarily a fixed model invocation. An API experiment needs its own record of model identifier, reasoning configuration, available tools and prompt. Treating both results as one undifferentiated GPT-5 score makes later comparisons difficult.

Separate user-experience trials from controlled model evaluations. Both are valuable, but they answer different questions: whether employees can work effectively in an application, and whether a configured component meets a repeatable technical requirement.

OpenAI reported improvements in factual reliability, instruction following and reduced sycophancy. Those claims identify useful test categories; they do not establish an error rate for a company’s own documents. A relative reduction in factual mistakes is especially easy to misread if the underlying task distribution is different. A customer-service answer, a financial reconciliation and an engineering explanation have very different consequences when wrong.

Construct a small reference set with source-backed answers and explicit failure categories. Include requests based on false premises. The model should challenge an incorrect assumption rather than produce a more polished justification. Grounding depends on company knowledge, not model branding alone.

A hypothetical support pilot should measure approved responses per staff hour, not just generation speed. Record how often a reviewer must fix facts, remove unjustified promises or find the right policy. Add a separate measure for harmful errors that survive review. Fast drafting can be valuable even without full automation, but only if the saved writing time exceeds the additional checking burden.

Use comparable cases and alternate the order of assisted and unassisted work to reduce learning effects. Do not present such a pilot design as a measured saving. AI advisory can help select a decision threshold before enthusiasm changes what counts as success.

GPT-5’s launch positioned it for coding and agent workflows as well as everyday writing. Connecting any reasoning system to tools changes the evaluation: an incorrect answer can become an incorrect action. Start by separating drafting, recommendation and execution. A model that prepares a refund explanation does not automatically need permission to issue a refund.

Require approval at consequential boundaries and test refusal, escalation and recovery alongside ordinary completion. The dedicated agent safety analysis provides the broader permissions framework. Here the practical implication is that a model upgrade should trigger a controlled re-evaluation, not an automatic expansion of what the application is allowed to do.

Define an owner, a limited use case and an exit condition. Keep a baseline system available and preserve the test set when changing prompts or models. Record costs from actual usage, including retrieval, retries and review, rather than assuming a historical token price describes the whole workflow. A launch announcement cannot establish the terms of a deployment more than a year later.

For leadership discussion, an AI keynote can explain why a unified consumer experience and an auditable business component require different evidence. The lasting lesson from GPT-5’s launch is to evaluate the whole path from request to approved outcome, with enough instrumentation to know which part actually improved.