Story 1 — Grok 4.5: What the Private Beta Actually Signals
On June 28, Elon Musk confirmed that Grok 4.5, built on xAI’s 1.5-trillion-parameter V9 foundation model with Cursor coding data added post-pre-training, entered private beta at SpaceX and Tesla. The performance claim — “close to, perhaps exceeding Claude Opus” — is vendor-asserted and unverified. No independent benchmark has been run. No system card has been published.
Deploying inside SpaceX and Tesla — which generate aerospace engineering problems, rocket code, vehicle software, and manufacturing data at scale — is a real-world validation environment that no benchmark suite replicates. Reinforcement learning is still running. The Grok Build coding harness is showing daily improvements.
xAI stated it will release models trained completely from scratch every month through end of 2026. This raises a direct question for any team building production applications on Grok: will version-pinned API access be maintained across monthly releases? Anthropic and OpenAI both support pinned model strings. xAI has not confirmed equivalent versioning for its monthly cadence.
Story 3 — Tesla Robotaxi in Miami: The Cameras-Only Bet Gets Bigger
Tesla launched its Robotaxi service in Miami on July 5 — its fifth city and first outside Texas and California — running without an in-car safety monitor and using remote monitoring only. The Cybercab platform runs on FSD v13 with camera-only sensing and no lidar. Waymo operates with lidar, radar, and cameras plus in-car safety availability.
Grok 4.5 is in private beta at Tesla for next-generation FSD reasoning, but is not yet in production. Current Miami deployments run on the existing FSD v13 stack trained on Dojo 2. The twelve-state expansion target by end of 2026 means Tesla needs twelve separate state regulatory approvals under twelve different permit frameworks — because no federal autonomous vehicle framework exists. Florida is among the most permissive states for AV operation.
Story 4 — Grok Video 1.5: xAI’s Surprise Win in the Video War
Grok Imagine Video 1.5 beat Sora 2, Veo 3.1, and Kling in blind user preference benchmarks. The wins are concentrated in abstract and stylised content and short social-format clips. The audio-in-one-pass capability — native synchronised audio generated simultaneously with video rather than as a separate step — is a genuine workflow advantage. The 86% cost reduction versus Sora translates to approximately $0.02 per clip at SuperGrok’s 50/day limit.
SuperGrok at $30/month now has three distinct advantages over Claude Pro and ChatGPT Plus — live X firehose data, best-in-test voice naturalness, and now the highest-volume lowest-cost video generation with synchronised audio. The caveat: Veo 3.1 still leads on photorealism for premium cinematic content.
xAI is moving fast on multiple fronts with no independent benchmark validation on its flagship model, a 530M legal reserve SpaceX set aside for Grok-related litigation is a real liability heading into a public company context.