GPT-5 Adds Modes But Model Picker Lives On

OpenAI revisits GPT-5 strategy after bumpy rollout

When OpenAI announced GPT-5 it pitched a simpler ChatGPT: one model with a router that would automatically pick the best internal model for each prompt. The aim was familiar — reduce menu fatigue and make model selection invisible to users. But reality diverged from the promise.

Instead of a single seamless experience, OpenAI exposed three modes in the model picker: Auto, Fast, and Thinking. Auto behaves like the advertised router, but Fast and Thinking let users override the router and pick models with different speed and response characteristics.

The company also re-enabled legacy models for paid customers — notably GPT-4o and others that were deprecated last week — after a user backlash. Many users had developed attachments to specific models' tones, verbosity, and quirks, and removing them triggered criticism and even public displays of mourning in the broader AI community.

OpenAI admits the rollout was bumpy: the router underperformed on launch day, prompting public explanations from Sam Altman and rapid iteration from the ChatGPT team. Routing is harder than it looks — it must infer user intent and preferences in milliseconds, then choose a model that balances speed, accuracy, and tone.

Beyond speed and correctness, model choice affects subjective qualities: one model may be more concise, another more playful, another more contrarian. Users attach to those personalities, and the industry is just beginning to understand how real that attachment is and how it shapes adoption.

The practical implications for product teams, enterprises, and platform operators are immediate:

• Design deprecation paths with user-facing controls and clear timelines to avoid surprises.
• Capture per-user preferences (tone, verbosity, risk tolerance) so routing decisions reflect real humans, not heuristics alone.
• Run gradual experiments and A/B tests to validate a router’s choices rather than switching models globally overnight.

For regulators and risk teams, the episode highlights another issue: opaque routing decisions can complicate audits and safety reviews if different prompts are handled by different model variants. Traceability and explainability must be built into routing logic.

What should organizations do now? Start by instrumenting user feedback: track when people switch from Auto to Fast or Thinking and why. Combine quantitative telemetry with qualitative interviews to map which model attributes matter most to different user groups.

QuarkyByte's approach is pragmatic: treat the router as a product feature that needs design, measurement, and rollback plans. That means short experiment cycles, clear KPIs tied to satisfaction and accuracy, and migration playbooks that preserve familiar personalities for users who depend on them.

OpenAI’s quick reintroduction of older models shows companies can course-correct. The bigger lesson is that one-size-fits-all routing is an attractive goal but must be married to per-user customization and careful change management. Expect more iteration from OpenAI — and more nuance for teams building on these APIs.

In short: GPT-5 didn’t erase the model picker. It turned model routing into a higher-stakes product problem — one that will reward teams that combine solid telemetry, user research, and staged rollouts over a simple flip-the-switch approach.