xAI tests Grok 4.5 private beta on a 1.5T V9 model with Cursor data

• According to testingcatalog's beta report and ai_for_success's repost of Musk's claim, Grok 4.5 is in private beta at SpaceX and Tesla.
• testingcatalog's screenshot says the model is based on a 1.5T V9 foundation model with Cursor data added during supplemental training, while rohanpaul_ai's summary frames that as roughly 3 times the 0.5T size previously associated with Grok 4.3.
• The performance claim is still mushy: Musk's quoted post only says Grok 4.5 is close to, and perhaps exceeds, "Opus," while koltregaskes notes nobody has clarified which Opus version that means.
• haider1's recap and scaling01's reaction surfaced the more interesting roadmap claim, that SpaceX plans to release new scratch-trained models every month this year.

You can read the quoted Musk post for the full wording, check koltregaskes's thread for the missing version detail on "Opus," and testingcatalog's iOS screenshots show Grok's "Tasks" feature quietly renamed to "Automations" at the same time.

The core update is simple: Grok 4.5 has moved into internal testing, not a public rollout.

Both posts point back to the same Musk statement, which places the beta inside SpaceX and Tesla rather than xAI's public product surfaces.

The most concrete technical detail is the jump to a 1.5T V9 base model plus supplemental Cursor data.

Musk also said reinforcement learning is still improving the model and that the "Grok Build harness" is getting better every day, which suggests xAI is framing the gain as both a larger base model and a better post-training stack, per the quoted Musk post.

The benchmark brag landed before any benchmark.

The only public claim so far is "close to, perhaps exceeding Opus" in Musk's quoted post. koltregaskes is right to call out the gap: without a named model slug, eval suite, or numbers, the comparison is more teaser than datapoint.

The other headline is release frequency.

Musk's post says completely new models trained from scratch will ship from SpaceX every month this year. That is a bigger claim than the Opus comparison, because it implies xAI thinks its training and harness loop is now fast enough to treat frontier models more like a recurring cadence than a once-a-season event.

One adjacent product detail slipped out through the iOS app.

The screenshots show Grok's "Tasks" feature renamed to "Automations," with the same basic scheduled-agent flow, a "Project Instructions" box, and push or email notifications. testingcatalog said it looks like a rename plus UI refresh for now, not a new capability.