Grok V9-Medium: What xAI's 1.5T Coding Model Means for Free Users

If you have been using Grok for free and wondering whether xAI's big new model announcement affects you, the short answer is: not right away, but more than you might think over the next few months. On May 25, 2026, Elon Musk confirmed that Grok foundation model V9-Medium had finished training at 1.5 trillion parameters -- three times the size of the model currently serving every Grok user. Reinforcement learning and supervised fine-tuning are now underway, and a mid-June launch window is expected. What that means for free tier limits, API costs, and the promised open-source release of an earlier model is worth unpacking carefully.

Grok V9-Medium's defining feature is not just its scale but what it was trained on: real developer workflows from Cursor.

What Is Grok V9-Medium and How Is It Different From What's Running Today?

Every Grok query you run today -- on grok.com or through the API -- is handled by the v8-small foundation model, which operates at approximately 500 billion parameters. Musk has described it candidly as "just 0.5T" and acknowledged it is missing important training data. V9-Medium is the correction: a 1.5 trillion parameter model that triples the underlying scale of the system.

The naming convention is worth understanding. "Grok 4" is the consumer-facing product name. "V9-Medium" is the internal foundation model identifier. When V9-Medium launches and clears safety evaluation, it will replace the v8-small underneath Grok 4 -- so free users on grok.com will be getting the upgraded model automatically, without changing plans or paying anything extra. The transition happens at the infrastructure layer, not at the subscription level.

To put the scale in context: the current v8-small sits at roughly GPT-3.5 territory in terms of raw parameter count. V9-Medium enters the same weight class as models competing for the top spots on major benchmarks. That does not guarantee better performance -- architecture, training data quality, and alignment tuning matter enormously -- but the raw material for improvement is there in a way that was not previously the case. For free tier users who have found Grok useful but occasionally frustrating on complex tasks, the upgrade is worth paying attention to.

Why Was V9-Medium Trained on Cursor Developer Data?

The most technically significant detail in Musk's announcement was not the parameter count. It was the training data. V9-Medium was trained with supplementary data from Cursor, the AI-augmented code editor that has become one of the most widely used development tools among professional engineers. Musk noted that "a lot of Cursor data" was incorporated and that more will continue to be added during fine-tuning.

Why does this matter? Most coding models are trained primarily on public GitHub repositories -- code that represents finished, polished output. Cursor data is different. It captures how developers actually work: the back-and-forth of debugging, the context given when asking for a refactor, the way a coder describes a problem before they know how to solve it. That behavioral signal -- what engineers do, not just what clean code looks like -- is increasingly recognized as the differentiator separating the top coding models from the rest.

The business context reinforces the technical point. In April 2026, SpaceXAI reached an agreement with Anysphere, Cursor's developer, giving SpaceX the right to acquire the company for up to $60 billion later in 2026, or pay $10 billion for collaborative work. The V9-Medium training pipeline is the technical rationale for that deal -- the Cursor data access is not a one-off experiment, but the beginning of an ongoing data relationship that is expected to feed into future Grok versions as well. For users who care about coding assistance specifically, this is xAI's clearest statement yet about where it is choosing to compete.

How Does Grok V9-Medium Stack Up Against Claude and GPT-5?

The honest answer is that we do not know yet, because V9-Medium has not been publicly tested. What we do know is where Grok currently stands and why xAI decided a major upgrade was urgent.

According to Enterprise Technology Research data from March 2026, Grok holds only 6% enterprise AI adoption compared to 55% for OpenAI and 47% for Anthropic (which jumped from 20% just a year earlier). On SWE-bench Verified, the industry-standard benchmark for real software engineering tasks, Grok 4 scores around 75% while Claude Opus 4.7 leads at 87.6% and GPT-5.5 reaches 88.7%. That is not a gap you close by scaling parameters alone -- it requires the kind of training data quality that Cursor workflows are intended to supply.

Analysts have consistently noted that raw parameter scale does not produce proportional performance gains. Mixture-of-Experts architectures, of the kind used by DeepSeek and increasingly by other frontier labs, can match much larger dense models at a fraction of the active parameter count. Whether xAI is using a dense or sparse architecture for V9-Medium has not been officially confirmed. What has been confirmed is that the model is specifically targeting coding performance, and will be evaluated against coding benchmarks as the primary measure of success at launch.

xAI's shift from 0.5T to 1.5T parameters is one of the largest single-version jumps among major frontier model providers in 2026.

What Happens to Grok's Free Tier When V9-Medium Launches?

This is the question most readers on a free or trial plan will care about most. The current free tier on grok.com allows 10 prompts every 2 hours. That limit has not changed since the current Grok 4 rollout and there has been no official announcement that it will change when V9-Medium deploys. The most likely scenario based on how xAI has handled previous model transitions: V9-Medium slots in as the new production model underneath Grok 4, free users continue on the same access terms, and the experience improves without a billing change.

For API developers, the picture is slightly more complex. xAI's developer program has offered up to $150 per month in free API credits through its data-sharing program, though availability has been subject to change and should be verified in the xAI console before building against it. When V9-Medium launches, API pricing for Grok models may be updated to reflect the larger model's cost profile -- inference on a 1.5T parameter model is computationally more expensive than on a 0.5T model, and that cost generally gets passed to API consumers at some point.

One useful thing to watch: xAI's pattern with previous model launches has been to keep the consumer grok.com interface on the flagship model while offering older models at lower API price points. If that pattern holds, the free web tier will improve for no cost, while paid API tiers will see pricing adjustments as V9-Medium moves into the standard tier and v8-small potentially becomes a lower-cost legacy option.

You can compare current Grok free tier limits against competitors on our AI free tier tracker, which is updated as providers make official changes.

Will xAI Actually Open-Source a Grok Model This Year?

Musk confirmed in comments alongside the V9-Medium announcement that the existing v8-small model -- the 0.5 trillion parameter model currently handling all Grok production traffic -- is planned for open-source release by the end of 2026. This would follow the precedent set by Grok 1, whose weights were released under Apache 2.0 in March 2024, and would give developers a free, locally runnable version of today's Grok 4 foundation.

If the release happens on schedule, it would be one of the more significant open-source drops of the year. A 0.5T model is large enough to run competently on a well-equipped workstation with multiple high-VRAM GPUs, and small enough that quantized versions would be accessible on consumer hardware. The AI open-source ecosystem -- already rich with Llama 4 Scout, Qwen 3.6, Mistral Small 4, and others -- would gain a model trained on a meaningfully different data diet, including the Cursor developer workflow data, which could make it particularly useful for coding tasks compared to similarly sized alternatives.

The caveat is "planned." Open-source timelines have a habit of slipping when commercial considerations intervene. Musk's track record on Grok open-source releases is mixed: Grok 1 was released as promised, but subsequent versions have not been. With V9-Medium representing a significant leap in capability, there is no guarantee xAI will want the v8-small in the public domain while V9-Medium is still being evaluated commercially. Developers interested in this release should watch the xAI model documentation for official announcements.

For context on the broader open-source AI landscape, our open-source AI section tracks model releases and weight drops as they happen.