Grok's next generation is nearly ready to ship. On May 25, SpaceXAI chief Elon Musk announced that Grok foundation model V9-Medium — a 1.5 trillion-parameter model three times larger than the version currently handling all Grok production traffic — has completed training, with evaluation results described as positive. Supervised fine-tuning is underway; reinforcement learning was set to begin within days of the announcement; and a public release is expected approximately two to three weeks from that date, placing it in mid-June 2026. For developers who rely on AI coding tools, that release window matters: V9-Medium was explicitly trained on Cursor data — real-world developer workflows from one of the most widely used AI code editors — and Musk has called it a major step forward specifically for complex programming tasks.

The current model in production is the v8-small, running on roughly 500 billion parameters. Musk has described it candidly as "just 0.5T" and acknowledged it is missing important training data. V9-Medium, at 1.5 trillion parameters, is the correction — and the pipeline behind it points toward something significantly larger still.

Cursor Training Data: Why Source Matters

The most consequential technical detail in Musk's announcement is not the parameter count. It is the training data. Cursor, the AI-augmented code editor used by developers at companies including OpenAI, Stripe, and Perplexity, became the source of supplementary training data for V9-Medium, with Musk noting that more Cursor data will continue to be added. That decision matters because it means V9-Medium was not trained only on public GitHub repositories — the standard corpus for most coding models — but on actual developer workflows, including how real engineers debug, refactor, and extend production codebases.

SpaceXAI reached an agreement with Anysphere, Cursor's developer, giving SpaceX the right to acquire the company for $60 billion later in 2026, or to pay $10 billion for collaborative work. That deal — announced in April 2026 — now has a clearer technical rationale: the Cursor data pipeline feeding into V9-Medium is part of a deeper integration plan, not a one-off training decision.

What V9-Medium Is Competing Against

The model enters a market where Grok currently holds approximately 6% enterprise AI adoption, compared with 55% for OpenAI, 47% for Anthropic, and 39% for Google, according to Enterprise Technology Research data from March 2026. On coding benchmarks, the gap is real: Claude Opus 4.7 leads SWE-bench Verified at 87.6%, while Grok 4 — the product brand currently layered above the v8-small foundation — reaches 75% on the same benchmark. Whether tripling the parameter count and incorporating Cursor workflow data closes that gap is precisely what the mid-June release will test.

Analysts have consistently noted that raw parameter scale does not produce proportionally better performance. Mixture-of-Experts architectures, of the kind used by DeepSeek, can match or outperform dense models at a fraction of total parameter activation. The quality of training data and the precision of the post-training alignment phase matter at least as much as size. V9-Medium's Cursor data strategy reflects an awareness of that constraint.

SpaceXAI Open-Source Pattern Continues

Alongside the V9-Medium announcement, Musk confirmed that the current 0.5T v8-small model is planned for open-source release by the end of 2026. The pattern here is consistent: Grok 1 weights were released in March 2024 under the Apache-2.0 license; Grok 2.5 was published on Hugging Face in August 2025 under a custom community license; Grok 3 was pledged for open-source release in February 2026. Each time, the previous-generation model becomes a public resource once the next generation stabilizes. For the open-source AI community, a 500-billion-parameter model from a frontier lab is a substantial foundation — the kind of resource that has powered entire ecosystems around Meta's LLaMA and Mistral.

The Grok 2.5 community license carried restrictions worth noting. Engineer Tim Kellogg characterized it at the time as a custom license with anti-competitive terms that prohibited using the weights to train competing foundation models. It is reasonable to expect similar conditions on any 0.5T release.

Grok 5 Roadmap and SpaceXAI Talent Drain

V9-Medium is not the endpoint of the current roadmap. Grok 5, targeting 6 trillion parameters in a Mixture-of-Experts architecture, is actively training on Colossus 2 — SpaceXAI's gigawatt-scale supercluster in Memphis, Tennessee, equipped with approximately 550,000 NVIDIA GB200 and GB300 GPUs. That cluster crossed the one-gigawatt power threshold in January 2026, making it the first confirmed gigawatt-scale AI training facility in the world. Grok 5 would be the largest publicly announced AI model by parameter count — roughly six times the estimated size of GPT-4 — though prediction markets currently give it only about a 33% probability of shipping by June 30, 2026, reflecting genuine uncertainty about post-training timelines.

That uncertainty has a specific source. Since February 2026, when SpaceX formally absorbed xAI and renamed the AI division SpaceXAI, more than 50 researchers and engineers have departed, according to reporting by The Information. Among the exits is Juntang Zhuang, who led the pre-training team — the group responsible for building new foundation models. The pre-training group has since shrunk to a handful of people. At least 11 former employees joined Meta; at least seven went to Mira Murati's Thinking Machines Lab. Sources cited by The Information attributed some departures to unrealistic deadlines and a culture of extreme work. Whether that talent loss affects V9-Medium — which completed training before the bulk of the exits — or constrains Grok 5's execution is the more pressing operational question for the lab.

SpaceXAI is preparing for what would be one of the largest IPOs in history, with SpaceX having filed confidentially in early 2026. The combined entity posted revenue of $18.67 billion and a net loss of $4.94 billion in 2025, according to Reuters' review of a draft S-1 prospectus. The AI division — burning approximately $14 billion in cash annually — is the primary driver of that loss.

How Does Grok V9-Medium Compare to Claude and ChatGPT?

V9-Medium has not shipped yet, so no independent benchmark data exists. What is available is Musk's own description of the current v8-small's weaknesses — poor training data balance and completeness — and his specific claim that V9-Medium delivers major improvements on complex programming tasks. The Cursor training data suggests genuine prioritization of developer use cases, placing the model in direct competition with Claude Code and GitHub Copilot rather than general-purpose assistants.

Whether it catches Claude on SWE-bench Verified, where Anthropic's models currently hold the highest published scores, is the benchmark that will matter most to developers when the model ships in mid-June.

Frequently Asked Questions

When will Grok V9-Medium be released?

Elon Musk announced on May 25, 2026, that V9-Medium completed training and would be released publicly approximately two to three weeks from that date, placing the expected public launch in mid-June 2026. Fine-tuning and reinforcement learning were still underway at the time of the announcement.

What makes Grok V9-Medium different from the current version of Grok?

V9-Medium runs on 1.5 trillion parameters — three times the size of the 500-billion-parameter v8-small model currently powering Grok in production. It was also trained on Cursor data, meaning real developer workflows from one of the most widely used AI code editors, rather than solely on public code repositories.

Will xAI open-source a version of Grok in 2026?

Yes. Musk confirmed in a reply on X that the current 0.5T v8-small model is planned for open-source release by the end of 2026. This follows SpaceXAI's consistent pattern: Grok 1 (2024), Grok 2.5 (August 2025), and Grok 3 (pledged for February 2026) were each open-sourced after a newer generation was stable.

How does Grok compare to Claude and ChatGPT on coding tasks?

As of May 2026, Grok 4 scores 75% on SWE-bench Verified, while Claude Opus 4.7 leads at 87.6%. Grok holds approximately 6% enterprise AI adoption compared to OpenAI's 55% and Anthropic's 47%. Whether V9-Medium closes that gap depends on post-training quality and independent benchmarks not yet available for the unreleased model.