What We Know About Grok 4.5

Grok 4.5 is built on a proprietary architecture, V9 with 1.5 trillion parameters, which xAI developed internally. Data from the Cursor developer tool were also included in the model's supplementary training. This is no coincidence — SpaceX announced in early June the acquisition of Anysphere, which develops Cursor, for 60 billion dollars. The transaction is expected to be completed in the third quarter of 2026. The integration of Cursor data into Grok's training is a clear strategic move: xAI needs to catch up with competitors in programming capabilities, where it currently lags behind both Claude and GPT.

Musk admitted on X that the model is still undergoing optimization — specifically, reinforcement learning from human feedback (RLHF) is not yet complete. Furthermore, the internal benchmark "Grok Build" is updated daily, so results may still change. Musk did not disclose a public release date. However, according to xAI's current practices, it can be expected that the model will be gradually released to subscribers of the X platform after internal tests are completed.

One Model Per Month: Unprecedented Pace

The biggest surprise of Musk's announcement does not come from Grok 4.5 itself, but from the plan to release a completely new model trained from scratch every month. In traditional large language model development, it is typically about incremental improvements — a new version builds on the previous one and adds improved data or techniques. Musk's vision is radically different: to build a completely new foundation model every month.

If the plan is realized, xAI will release six independently trained models by the end of 2026. This is a research density that far exceeds anything ever created in the field. For comparison: OpenAI needed roughly two years between GPT-4 and GPT-5, Anthropic releases new generations of Claude approximately every 6–9 months. A monthly cadence would mean an exponential increase in computational power costs — and this is where SpaceX comes in.

SpaceX as the Financial and Computational Backbone

In February 2026, SpaceX completed the acquisition of 100% of xAI at a valuation of 1 trillion dollars. In May, xAI was officially integrated into SpaceX and renamed SpaceXAI. Thanks to SpaceX's computational power and capital, xAI gained resource support unmatched by any other AI startup.

However, the cash burn figures are as staggering as the ambitions. According to internal documents, xAI's operating loss for the first quarter of 2026 reached 2.47 billion dollars, with capital expenditures totaling 7.7 billion. Although the company closed a 20-billion-dollar Series E round at the beginning of 2026, most of these funds have already been invested in data center construction and research costs. Training a new foundation model every month will put enormous pressure on cash flow.

Real-World Testing: Rockets and Cars

Grok 4.5 is not just tested in the lab — it runs in pilot operation directly within SpaceX and Tesla. Musk repeatedly emphasizes that AI, to be useful in the real world, must be tested by the real world. The engineering complexity and cost of errors in a rocket launch or car manufacturing are orders of magnitude higher than in a typical conversation. If the model encounters problems first in internal operations, it is more valuable for subsequent optimization than any benchmark.

This approach simultaneously gives xAI a unique competitive advantage: direct access to real industrial data and scenarios that neither OpenAI nor Anthropic have. While competitors test on academic benchmarks and synthetic tasks, Grok learns from the real-world operations of factories and space missions.

What This Means for the AI Industry

Musk's "new model every month" strategy — whether it ultimately succeeds or not — pushes the pace of competition to a new level of intensity. For competitors, it's a dilemma. Keeping pace means dramatically increasing research budgets and computational investments. Not keeping pace means risking falling behind in the iteration speed.

OpenAI, Anthropic, Google DeepMind — all continuously release new models, but at an order of magnitude slower pace. If xAI can maintain a monthly cadence for even a few months, it will rewrite the entire market's expectations of what a "normal" AI development speed is.

At the same time, it must be added that pace does not equal quality. Training a model from scratch every month requires not only enormous computational capacity but also sufficient high-quality training data, which is an increasing challenge with the current depletion of publicly available resources. And most importantly — whether the models truly achieve the promised performance will only be shown by independent benchmarks, not internal tests.

Availability for Czech Users

Grok is currently available to subscribers of the X (Twitter) platform as part of the X Premium+ subscription, which costs approximately 400 CZK per month in the Czech Republic. The model is not localized into Czech — communication primarily takes place in English, although Grok can handle basic queries in Czech. xAI offers a separate API for developers through the x.ai platform. API prices for Grok 4.5 have not yet been announced — the current Grok 3 version costs $3 per million input tokens and $15 per million output tokens, which is comparable to GPT-4o.

For Czech companies and developers, it is important to mention that Grok — unlike Claude or Gemini — is not available within EU AI Act compliance programs, and xAI has not yet published detailed documentation on how it handles European user data.