The Picker Is Not the Product
Grok 4.6 gives SpaceX a near-frontier model to go with the coding tool it owns. The neutrality debate that followed is watching the smaller asset — and there's a pricing cliff in the model itself worth more to you this week.
Grok 4.6 gives SpaceX a near-frontier model to go with the coding tool it owns. The neutrality debate that followed is watching the smaller asset — and there's a pricing cliff in the model itself worth more to you this week. What happened Grok 4.6 shipped on 12 August 2026 , scoring 61 on the Artificial Analysis Intelligence Index — tying GPT-5.6 Sol Max, one point behind Fable 5 Max, two behind Claude Opus 5 Max. That index is measured by a third party, not by the vendor. (A note on our own timing: 12 August falls outside this scan's window and our previous scan missed it. The news here is the implication, not the release.) Pricing is $2 / $6 per million tokens, with cached input at $0.50 . Context window: 500,000 tokens. SpaceX acquired Cursor for $60 billion in stock — with a $10 billion walk-away fee if the deal had not proceeded. (Those are two different numbers and they get conflated constantly.) Nine days ago, when we recorded the acquisition, the model-neutrality question was hypothetical: there was no in-house model worth routing to. Grok 4.6 changes that. First, the thing that's worth money this week Before the strategy, the pricing detail — because it is immediately actionable and almost nobody is discussing it. Above 200,000 input tokens, the entire request reprices to $4 / $12. Not the excess. The whole request. Input tokens Input rate Cost of input --- --- --- 199,000 $2.00/M $0.398 201,000 $4.00/M $0.804 A 1% increase in input size doubles the input cost. This is a step function, not a taper. For a chat product this is irrelevant. For an agentic product it is the dominant cost dynamic: Workload Typical input Crosses 200K? --- --- --- Single chat turn 2–10K No RAG query with retrieved docs 20–60K Rarely Codebase-aware completion 50–150K Sometimes Agent, turn 15+, with tool outputs 150–400K Frequently Agents accumulate conversation history, tool outputs and retrieved context by construction . Nobody decides to cross 200K. They just do. An agentic system that manages context to stay under the threshold can cost roughly half as much as an identical system that does not, for identical work. That is an engineering decision worth more than most model selections. We published a routing playbook on 29 July , built on Microsoft's MAI numbers, arguing that where you send a request matters more than which frontier model you standardise on: Cheaper and Better: What Microsoft's MAI Numbers Teach You About Model Routing. The context cliff is the same argument arriving as a pricing structure rather than a capability one. And separately: cached input at $0.50/M is four times cheaper than standard input. If your workload has a large stable prefix — a system prompt, a codebase, a document set — caching that portion cuts 75% off it. The neutrality question, stated fairly Named concerns exist, and they predate Grok 4.6 — they attach to the acquisition. Wired asks whether Cursor can remain "a platform for OpenAI and Anthropic's models inside SpaceX," no