Key Takeaways
- In roughly thirty days, Amp shifted its backend traffic from 80 percent Anthropic (Claude and Opus) to 80 percent OpenAI, while customer use of GLM-5.2 grew alongside them.
- Amp removed user-facing model selection dropdowns entirely, replacing them with flat agent intensity controls because model differences no longer dictate user outcomes.
- Founder Quinn Slack argues that benchmark launches no longer drive developer loyalty; buyers no longer trust published leaderboards over raw task completion.
- Product defensibility has moved away from foundational model access to the surrounding software: the context extraction system and the runtime that manages parallel agents.
The Eighty Percent Backend Swing
For two years, AI engineering teams lived in terror of missing a model drop. If Anthropic released a new Claude or OpenAI dropped a new GPT version, dev tools scrambled to update their integrations within hours. Founders assumed users would churn if they could not immediately run the single most expensive model on the market.
That dynamic has broken. Slack watched it happen in Amp's production telemetry. Within about four weeks, Amp swapped the vast majority of its underlying traffic without triggering user revolts.
When three competing providers produce roughly equivalent results across complex software engineering tasks, the foundational layer becomes infrastructure rather than product identity. Developers do not care which provider answers their query as long as their pull request builds cleanly.
Dropdowns Confuse Users Who Just Want Work Done
Most software interfaces still expose the underlying plumbing. They hand the user a picker with eight models, distinct context windows, and different per-token prices. That design choice treats the buyer like an AI researcher instead of a builder trying to ship code.
Amp scrapped the dropdowns. Instead of asking a developer to guess whether Claude 3.5 Sonnet or GPT-4o will handle their refactor better, Amp offers flat intensity controls. The user dictates how hard the system should push and how many concurrent threads to fire; the platform manages routing under the hood.
This shift matches a broader collapse in developer faith regarding public benchmarks. As Slack observed: “It's no longer about using the best model. The models are going to keep getting better and better, but in the past, remember, everyone used to switch to the new model on launch day, and everyone was so excited about it. And now, it's still exciting to see, but you don't really trust the benchmarks.” (play from 45:36)
The Real Moat Is Scaffolding and Human Intent
If the top models perform similarly, buying API credits cannot differentiate your startup. A competitor can purchase identical tokens five minutes after reading your landing page.
The actual moat is the mechanism that translates human thought into an actionable execution loop. Slack puts the priority plainly: “Because now what is most important is the thing that gets that information out of your brain to the agent, and then the thing that lets you run a bunch of agents at once.” (play from 44:14)
When a model acts as a fungible engine, your value comes from two layers: context retrieval and orchestration. How well does your tool capture what the engineer actually wants? How does it structure repo context, verify test outputs, and handle background execution without human supervision? Slack's experience proves that an agent tool can thrive, stay profitable, and swap 80 percent of its backend without users caring, provided the runtime surrounding the model does the heavy lifting.
What to Do With This
Audit your application interface this week and delete your model selector dropdown. Replace it with an intent or speed selector (such as standard, thorough, or background) and build an internal router that dispatches tasks to the cheapest provider meeting your evaluation bar.