Grok 4.5 Private Beta: 7 Essential Facts About xAI’s 1.5T Model in 2026

Grok 4.5 entered private beta at SpaceX and Tesla on June 28, 2026. Here are 7 confirmed facts about xAI’s 1.5T V9-Medium model and what they mean for AI in 2026.

xAI’s Grok 4.5 entered private beta at SpaceX and Tesla on June 28, 2026. Elon Musk confirmed it on X. Early internal evaluations reportedly place performance at or above Anthropic’s Claude Opus. Public release expected shortly after the internal deployment stabilizes.

That’s the headline. Here’s what it actually means, what’s confirmed versus what’s claimed, and what most coverage gets wrong about where this model fits in xAI’s lineup.

Grok 4.5 Private Beta: What xAI Actually Shipped

First, the naming confusion. xAI runs two parallel naming systems. Consumer release names appear in the app: Grok 4, Grok 4.1, Grok 4.3. Foundation model versions run internally: V8, V9. The model being called Grok 4.5 is the consumer and enterprise deployment name for the V9-Medium foundation, a 1.5 trillion parameter model that completed base training on May 26, 2026.

Some coverage treats V9-Medium and Grok 4.5 as separate products. Not the case. Same model. Two names.

The current Grok production model, v8-small, runs on roughly 500 billion parameters. 500 billion. The new model at 1.5 trillion is a threefold jump. Grok 4.4 shipped with 1 trillion parameters in late May 2026, so the step from 4.4 to 4.5 is a 50% increase in a single release. That scaling jump is real. Whether it translates to proportional capability gains is what the SpaceX and Tesla beta is designed to determine.

7 Confirmed Facts About Grok 4.5

1. 1.5 Trillion Parameters on V9

xAI’s Grok 4.5 runs on the V9 foundation at 1.5 trillion parameters. Most frontier models sit in the 200B to 500B effective parameter range. Threefold jump. Not incremental.

That said, raw parameter count is a capacity signal, not a capability guarantee. Mixture-of-Experts architectures like DeepSeek’s match dense models at a fraction of active parameters. What makes the 1.5T number meaningful here isn’t the scale. It’s what the model was trained on.

2. Cursor Training Data Is the Real Story

The most consequential detail about Grok 4.5’s architecture isn’t parameters. It’s the training corpus.

V9-Medium was trained on Cursor workflow data. Cursor is one of the most widely used AI code editors, with engineers at companies including OpenAI, Stripe, and Perplexity using it daily for production work, and its training dataset captures not just what code looks like but how engineers actually build, debug, and iterate on code in real production environments over hours-long sessions. Most coding models train on public GitHub repositories, which capture finished code but miss the messy, iterative, multi-file debugging process that constitutes most of an engineer’s actual day. This one didn’t. V9-Medium trained on how real engineers actually work: multi-file refactors, debugging sessions, production codebase extensions, and agentic coding loops that never appear on GitHub at all.

IDE-level data. Not public repo data. That difference is more likely to close benchmark gaps than the parameter increase alone, because SWE-Bench Verified tests the kind of reasoning required to understand incomplete codebases and fix bugs that span multiple files, which is exactly the situation Cursor users face every day when working through real production codebases with the tool.

The connection to Cursor goes deeper than a data deal. In April 2026, SpaceXAI signed an agreement with Anysphere, Cursor’s developer, that includes an option to acquire the company outright for $60 billion, and on June 16, the same day V9-Medium appeared in the consumer Grok product on X and SuperGrok, SpaceX filed the binding merger agreement for that acquisition. If the deal closes, xAI gains not just historical training data but the ongoing Cursor usage firehose going forward, creating a compounding data advantage for every model that follows Grok 4.5. That’s the structural story underneath the parameter count.

3. SpaceX and Tesla Are Not Just Brand Names

The choice to run the private beta at SpaceX and Tesla is a deliberate technical decision. Not positioning. Not a press beat.

Most AI companies test models on public benchmarks and synthetic tasks designed to isolate specific capabilities under controlled conditions that rarely resemble actual production environments. SpaceX and Tesla give xAI something fundamentally different. High-stakes environments. Real scale. SpaceX workflows span orbital mechanics calculations, Starlink constellation management, and real-time manufacturing decisions under pressure that no benchmark suite simulates, while Tesla’s engineering pipelines involve full-stack software development across an active fleet of millions of connected vehicles whose software must ship reliably and be updated over-the-air to drivers around the world. The breadth and depth of real-world coding and reasoning tasks those environments generate is exactly the kind of coverage a synthetic test suite misses.

Grok 4.5 is being tested against actual rocket trajectory analysis and vehicle software architecture. Not curated prompts. Not controlled conditions. The feedback from those deployments feeds directly into reinforcement learning that continues to refine the model, and no benchmark score can replicate the signal you get from a SpaceX engineer catching a model error in a real trajectory calculation. Training signal competitors cannot buy. No other lab has access to this operational data, and the February 2026 SpaceX-xAI merger made that access permanent rather than ad hoc.

4. Reported Performance at or Near Opus

Musk stated that early internal evaluations place performance “close to, and potentially above, Opus.” Anthropic’s Claude Opus is the reference point, currently sitting at approximately 87.6% on SWE-Bench Verified for coding tasks.

75%. That’s where the current Grok production model sits on SWE-Bench Verified. Twelve points behind Opus. Closing that in a single release would be significant.

The word “potentially” matters. Internal evaluations from the company that built the model. Pre-release claims routinely outperform independently verified results. That’s not unique to xAI. Every lab does it. Independent benchmarks haven’t run because the model hasn’t left the building yet, and while the SpaceX and Tesla operational stress testing is more rigorous than any benchmark suite, the third-party numbers are what the industry actually uses to compare. Once Grok 4.5 exits private beta, those numbers arrive quickly. Wait for them.

5. Monthly Release Cadence Through Year-End 2026

xAI’s roadmap anticipates new models trained from scratch at SpaceX every month through the end of 2026. The model entering private beta today is the current entry in that cadence.

Monthly. Frontier-scale AI. Unusual doesn’t cover it.

Most labs operate quarterly or longer because training runs take months, and post-training alignment adds more time on top of that. xAI runs multiple training jobs simultaneously on Colossus 2, and that infrastructure is what makes the monthly pace feasible at all. Colossus 2 is a gigawatt-scale facility in Memphis, Tennessee, equipped with approximately 550,000 NVIDIA GB200 and GB300 GPUs, making it the first confirmed gigawatt-scale AI training facility in the world since it crossed that threshold in January 2026. Seven concurrent training jobs running right now, spanning multiple parameter-count variants and specialized models including Imagine V2. Monthly cadence: production schedule, not aspiration.

6. Reinforcement Learning Is Still Running

The model in private beta today is not a finished product. xAI confirmed reinforcement learning continues to improve it after base training completion. Still running. Musk’s post also referenced updates to the Grok Build framework, the agentic coding environment V9 feeds into.

By the time a public API endpoint ships, the model will have accumulated additional refinement rounds from SpaceX and Tesla deployment data, ongoing RL training, and human feedback from the internal beta. Moving target. What enters public release won’t match today’s private beta snapshot.

7. V9-Medium Was Already in Consumer Grok Before Today

Before the formal announcement, V9-Medium began appearing in the consumer Grok product on X and SuperGrok on June 16, 2026. X users who’ve interacted with Grok in the past two weeks have already used the foundation behind Grok 4.5. The private beta at SpaceX and Tesla refers specifically to an enterprise track with access to proprietary operational environments. Different access tier. Same underlying model.

Consumer availability and API availability are separate gates at xAI. The API endpoint for V9-Medium hasn’t appeared in xAI’s documentation as of late June. For developers who need an xAI coding API today, grok-build-0.1 is in public beta. The full model endpoint will follow.

Grok 4.5 vs Grok 5: Two Different Models

The most consistent error in coverage of xAI’s lineup is treating these as the same release or a simple successor path.

They’re not the same product. Different use cases. Different scales.

Grok 4.5: 1.5 trillion parameters, coding-specialized, already deployed at SpaceX and Tesla. Grok 5: xAI’s next-generation general-purpose flagship targeting 6 trillion and 10 trillion parameter variants in a Mixture-of-Experts architecture, still in training on Colossus 2 as of late June 2026.

Prediction markets assigned roughly 33% probability of a Grok 5 public release by June 30. That window has closed.

Grok 5’s parameter counts start at 6 trillion. The model entering beta today is a specialized technical product built at 1.5T. Separate roadmap. Separate purpose. Neither substitutes for the other.

What Grok 4.5 Means for Tesla Owners

Tesla’s in-car AI has consistently lagged behind xAI’s development pace. The current in-car Grok assistant runs several model generations behind what xAI ships to API users. If Grok 4.5 moves from the internal beta to production deployment inside vehicles, it would be a meaningful step forward in how the car handles natural language, complex queries, and multi-part voice commands.

Tesla distributes software via over-the-air updates across the entire connected fleet simultaneously. A certified production deployment could reach millions of vehicles in a single OTA push. No dealer visit. No manual download. No driver action required.

Timeline for that production move: unconfirmed. The current beta is the step before it. Tesla’s OTA release notes are the signal to watch. They typically list software version details and new features. When Grok 4.5 appears in those notes, the transition is live.

The Structural Position Behind the Release

Grok 4.5’s deployment illustrates something that extends beyond a single model release.

Most AI labs spend heavily to acquire both users and compute, negotiating distribution deals, building developer relations from scratch, and competing for the same pool of engineering talent that every other frontier lab is also recruiting aggressively. xAI’s empire supplies both internally. X provides hundreds of millions of users and a real-time data stream. Tesla: a global hardware footprint of connected vehicles that update simultaneously. SpaceX: engineering talent and operational environments no other lab can access. Colossus 2: compute to train at a pace most labs can’t sustain. And when the Cursor acquisition closes, xAI adds the most detailed coding workflow data pipeline available anywhere.

The concern for competitors watching Grok 4.5 isn’t specifically whether it tops their best model on a single benchmark. It’s structural. Deploy to an enormous installed base instantly. Improve against captive real-world usage. No distribution negotiation. That loop compounds. A better model score on one leaderboard doesn’t easily counter it.

What the February 2026 SpaceX-xAI merger changed is worth noting. Before it: informal talent and infrastructure sharing between two companies that happened to share a founder. After it: SpaceX engineering workforce, GPU clusters, and operational data became formal xAI resources, and the line between “Musk’s AI company” and “Musk’s rocket company” stopped existing as a meaningful boundary. Grok 4.5 is the first model entering private beta in the post-merger company. Its testing environment, SpaceX and Tesla operations, is fundamentally different from anything a pre-merger Grok model tested against.

For more on how xAI’s Colossus 2 compares to other frontier training infrastructure, see the official xAI blog

Grok 4.5 by the Numbers

1.5 trillion parameters. V9-Medium, the foundation.

500 billion. v8-small, the current production model.

Threefold jump. One release.

1 trillion. Grok 4.4, shipped late May 2026.

50 percent. Parameter increase from 4.4 to 4.5.

87.6 percent. Claude Opus on SWE-Bench Verified.

75 percent. Current Grok production model on SWE-Bench Verified.

12 points. The gap between them.

550,000 GPUs. Colossus 2, Memphis, Tennessee.

One gigawatt. Power threshold Colossus 2 crossed in January 2026.

Seven. Concurrent training jobs running right now on Colossus 2.

$60 billion. SpaceXAI’s option to acquire Cursor.

June 16. V9-Medium arrived in consumer Grok on X and SuperGrok.

June 28. Private beta confirmed at SpaceX and Tesla.

Monthly. xAI’s stated model release cadence through year-end 2026.

33 percent. Prediction market odds that Grok 5 would ship by June 30. It didn’t.

February 2026. SpaceX formally absorbed xAI, renaming the AI division SpaceXAI and making the companies one entity for the first time.

The monthly cadence points to another new model in July 2026, with the series continuing through year-end. Grok 5 remains in training in its 6T and 10T variants. Two parallel timelines.

The near-term signal to watch: independent benchmark results once Grok 4.5 exits private beta. SWE-Bench Verified, HumanEval, and MBPP will confirm whether the Opus-level claim holds. Current 12-point gap behind Opus. Close it, and the Cursor data strategy is validated. Fall short, and the parameter scaling story needs more time.

The SpaceX and Tesla beta is the real test regardless. Rocket trajectory. Vehicle software architecture. Those environments don’t grade on a curve.

FAQs About Grok 4.5

When will Grok 4.5 be publicly available? No specific date confirmed. Musk’s post referenced a release “in the near future,” but xAI hasn’t committed to a timeline. The private beta at SpaceX and Tesla is explicitly the step that precedes public release. Until the company announces otherwise, treat any leaked date as speculation.

How is Grok 4.5 different from Grok 4.4? Two differences. Parameters: Grok 4.4 shipped with 1 trillion parameters in late May 2026; the new model adds another 500 billion, a 50% increase. Training data: Grok 4.5 introduces supplemental training on real Cursor developer workflows, not just public GitHub repositories. The training data difference is more significant than the parameter gap.

Is Grok 4.5 the same as Grok V9-Medium? Yes. Same model. V9-Medium is the internal foundation model designation; Grok 4.5 is the consumer and enterprise name. Two naming systems, one product.

Does Grok 4.5 affect Tesla’s Full Self-Driving? No. Not even slightly. Grok 4.5 is an in-car AI assistant for voice interaction, route queries, and conversational features. Tesla’s Full Self-Driving is a completely separate system built on different architecture entirely. The new model doesn’t touch FSD behavior.

What is Grok 5 and how is it different from Grok 4.5? Different model entirely. Grok 5 targets 6 trillion and 10 trillion parameter variants, still in training on Colossus 2 as of late June 2026, and targets general-purpose reasoning rather than specialized coding. Grok 4.5 is a 1.5 trillion parameter coding-specialized product already deployed at SpaceX and Tesla. Not the same release. Not the same use case.

Will Grok 4.5 be open-sourced? Unknown. xAI hasn’t confirmed open-source plans. What is confirmed: the current 500B parameter production model (v8-small) is planned for open-source release by end of 2026. Whether the same path applies to Grok 4.5 hasn’t been addressed publicly.