GPT-6 Astra and Claude Fable cost exactly the same: $10 per million input, $50 per million output. That single fact changes the decision more than any benchmark did.
When two frontier models are priced identically, procurement stops being the question and routing starts. You are not choosing a vendor. You are choosing, per task, which profile you want.
The routing rule follows straight from the profile. Send Astra the work where the task is operating something over time. Send reasoning-dense work elsewhere, because at the same price you can currently get better judgment.
The third band on that table is the one companies skip, and it is the one that costs money when skipped. Long autonomous computer use is exactly the capability that makes an unsupervised mistake expensive, which is the subject of Part 6.
One caution against over-routing, from Ethan Mollick, and it cuts both ways: firms consistently underestimate the value of higher intelligence on exactly the tasks where a weaker model merely seems good enough. Thorsten Ball's version is the arithmetic one. A model that is not smart enough retries, and retries burn tokens, so per-token price is not per-task cost. Route the task, not the token.
Vendor loyalty is already dissolving without you
The behavioural version of this arrived before the pricing one. Peter Yang published a ten-use-case bake-off across ChatGPT, Claude, Grok and Gemini and reported a real shift in his own habits: "A few months ago, I used Claude for almost everything. Today, my AI usage looks very different."
That is not a benchmark result and it is more interesting than one. It says the default habit is now unstable, and that per-task routing is replacing single-vendor loyalty even where nobody has built a router. Suhail Doshi supplied the mood underneath it, from inside a frontier lab: "Cannot. Keep. Up. With. All. These. Models."
If the people building models cannot track the releases, expecting your team to maintain a considered per-task model policy by hand is unreasonable. That is the demand case for routing as infrastructure rather than as discipline. Stripe paid a reported $7.5B for OpenRouter on that thesis.
Routing is a margin lever, not a cost-control measure
The framing that makes this worth an engineering week comes from Olivia Moore at a16z: "It feels like we've hit diminishing returns on intelligence for many tasks. We may no longer see every product auto-switch to the next frontier model upon release. This is great for app-layer builders."
Two things follow. The reflex of shipping the newest frontier model on release day was never a product decision, it was a fear of being outclassed, and if marginal intelligence stops mattering for most tasks then that fear is optional. And routing becomes something you pull deliberately for margin rather than something a CFO imposes after a bad month.
Garry Tan supplied the practitioner proof in the same window, from the person who coined tokenmaxxing: he cut GStack's token load by 50% with no capability reduction. Waste, not usage, was the variable.
Three rules that survive the next model release
Route on task shape, not on release date. The question is never "what is newest," it is "is this task operating something or judging something." That question does not expire when the next model ships.
Make the routing inspectable. A probabilistic classifier chooses the path and deterministic routes execute it, so you can audit afterwards why a given request went where it did. A router nobody can explain is a second black box stacked on the first.
Treat topology as the cost model. Cheap models for extraction and classification, strong models for decomposition and synthesis. Which model runs at which node is the COGS decision, and writing it as a diagram makes it arguable by people who do not write prompts.
Which of your workflows is on the wrong model?
Book a free Diagnostic: 30 to 45 minutes, no deck, no pitch. We map your workloads onto the two axes and find the ones paying frontier prices for classification.
Book the Diagnostic →Part 3 of Working with GPT-6 Astra: The Operator Guide.
