SpaceX's Grok Bot will route work to Claude Opus 5.5 whenever Anthropic's model does it better

Musk's Wednesday note sets a policy ^[4]. "Whatever is most likely to give you the best outcome," he wrote ^[5]. That phrase is a forecast made per task, before the task runs. Some component has to make it for every request Grok Bot receives. The post does not say what scores the candidate models, or against which tests ^[5].

The operators Matt Burns interviewed for The New Stack described what that component usually looks like ^[1]. The most common setup was a funnel ^[10]. One security company runs a rules engine over everything first, passes what survives to small models, and lets larger models see only what is left ^[10]. One of its executives said a petabyte of data would cost millions of dollars to process with any model, small ones included ^[11]. A second company sends each agent request through a cheap model to work out what the user wants before the expensive model does anything ^[12].

The classifier stage keeps getting cheaper. Haiku 5.5 now costs $0.10 per million input tokens after a 90% cut. That puts the earlier rate for requests under 100,000 tokens at about $1 ^[18]. Burns notes that small models like Haiku typically handle classification and routing ^[16].

Switching is where routing breaks. One engineering leader told Burns that three days before demoing to an important prospect, his team moved to a newer model that the benchmarks showed was cheaper and no worse ^[13]. It broke the workflow, and the team spent a weekend rolling back ^[13]. His lesson was evals ^[13]. Vendors can cause the same failure from their side. The New Stack's Amanda Caswell reported that when Anthropic made Opus 5.5 20% cheaper than Opus 5, the change broke four things agents depend on ^[14]. Burns's rule is "Always Be Switching" ^[17]. The demo team would want the rule to name a weekday.

Burns wrote in August that "when models converge on price, the money moves to whoever decides which model gets the job" ^[7]. His evidence supports a narrower claim. Cost was the reason most interviewees gave for routing ^[2]. The interviews count teams that route. That is evidence of adoption, while Burns's claim is about where the money goes. The sample is also narrow. The 22 companies are Insight Partners portfolio companies, interviewed for an Insight video series, and Insight owns The New Stack ^[8]. A couple of the interviewees sell small models for a living, and Burns says he weighs their enthusiasm accordingly ^[9].

We think the routing logic is the easy part to copy. A rules engine and a cheap intent classifier are ordinary components ^[10][12]. The harder part is the eval suite that defines "best" for each task and catches a vendor's breaking change before a customer does. If SpaceX gets a lasting advantage from Musk's policy, we'd expect it to come from those tests. A suite like that is what the demo team lacked ^[13].