xAI · Grok 4.5
The internal beta
finally shows its cards.
Since 6/30, Grok 4.5 had been distributed only inside SpaceX and Tesla — no architecture, no rollout plan for anyone else. On 7/9, xAI published the specs and the release calendar. It's the day model-selection spreadsheets can finally add a real Grok column.
Ten Days In The Dark
"Runs at Tesla" — and
that was all we knew
On 6/30, xAI dropped Grok 4.5 as a beta inside SpaceX and Tesla. Rumors leaked out, but externally there was only a single line — "it's running inside Musk's companies." Parameter count, context window, price: none of the numbers a buyer needs were public.
For those nine or ten days, anyone keeping Grok on their shortlist was picking blind. On 7/9, xAI released the architecture, pricing, and external rollout plan in a single disclosure. The story of the day is simple: the pieces needed to put Grok "next to the others in the spreadsheet" are finally on the table.
Specs Unveiled
Three lines for the comparison sheet
Three things from today's reveal that matter for real model selection work.
Architecture
The MoE (Mixture of Experts) structure is now confirmed. Only a fraction of the total parameters activate per step, so inference costs will come in lower than a same-tier dense competitor, all else equal.
Distribution
Web and mobile (via X) are live already. The API begins a staged public rollout from today. Enterprise contracts are handled case-by-case; direct integration with other clouds is expected within weeks.
Where it leans
Working backward from the SpaceX / Tesla usage, tuning skews toward long-context and heavy numeric or engineering tasks. The motivation to replace GPT / Claude in general chat is thin — and worth saying plainly.
Position
The fourth name in the deck
Bottom Line
Should it be on your shortlist?
Honestly, the case for replacing an existing GPT / Claude / Gemini deployment is weak today. But moving from "can't put its name in the comparison at all" to "can" is a small but real step for anyone who owns model selection. Watching implementation reports over the next week or two makes sense — especially for cost-optimization and second-string routing for specific task types.
For everyday users, nothing changes. That's still true.