Grok 6

• 1. overview (440 chars)

  Product Express refers to xAI's flagship offensive in the second half of 2026 as "Grok 6", but as of July 21, 2026, the latest model actually released by xAI is Grok 4.5 (July 16, 2026). Grok 4.6 (2 trillion parameters) is still in the final stages of training and is expected to be launched in August. This report is based on the real capabilities of Grok 4.5, covering its multi-modal and real-time retrieval features, and supplements 4.6 and subsequent routes for manual correction of naming caliber during background review.

• 2. Product background (659 chars)

  Grok was built by xAI, founded by Elon Musk, and its name is taken from the meaning of "deep understanding" in "Stranger in a Strange Land". In February 2026, SpaceX announced the acquisition of xAI, and Grok became part of the SpaceXAI system; in January of the same year, xAI completed a US$20 billion Series E financing, and in June SpaceX landed on Nasdaq with a valuation of approximately US$1.77 trillion. Grok 4.5 is the first flagship model after the reorganization. The training relies on the Memphis supercomputing cluster code-named Colossus, with a scale of more than 200,000 GPUs. It is jointly trained with Cursor, a programming tool recently acquired by xAI for $60 billion, to inject real developer session data into the model instead of just using a static public code base.

• 3. Product functions and usage experience (1079 chars)

  Grok 4.5 is based on the 1.5 trillion parameter V9 base (about three times the previous generation), and its positioning has shifted from "general intelligence" to "programming and agent". It performed outstandingly on real engineering tasks: SWE Marathon scored 29.0%, slightly higher than Claude Opus 4.8's 26.0%; DeepSWE 1.0 and Terminal-Bench 2.1 led its peers with 62.0% and 83.3% respectively. In terms of multi-modality, Grok supports text and image input, and integrates Grok Imagine video, voice and real-time X data streams under a unified architecture; Musk said that native multi-modality (text/image/video/audio) is the core direction of subsequent generations. Real-time retrieval is Grok’s long-standing strength. It is connected to X real-time streaming by default and is ahead of its competitors in terms of timeliness and delay of breaking news. In terms of use, Grok 4.5 provides three levels of inference intensity: low/medium/high, a context window of 500,000 tokens, and an inference speed of about 80–90 TPS, which is suitable for interactive and high-throughput workflows. The official also previewed the million-context version, and regarded "multi-modal and real-time retrieval capabilities" as the core selling point of this generation.

• 4. Pricing and business model (882 chars)

  Grok 4.5's API pricing is $2 per million tokens for input and $6 for output. Long context switches over 200,000 tokens are $4 for input and $12 for output; cache hit prices are as low as $0.50. The Cursor-specific quick variant is $4 for input and $18 for output. The consumer side is available via SuperGrok with an X Premium+ subscription (a figure of around $30/month is circulated by the community, not officially confirmed by xAI on the home page). The real killer feature is the "single task cost": on the Coding Agent Index, Grok 4.5 costs about $2.49 on average to complete a standard task, while Claude Fable 5 costs about $11.80, a gap of nearly 5 times; for the same SWE-Bench Pro, it only uses an average of about 16,000 output tokens, and Opus 4.8 costs 67,000, a difference of more than 4 times in efficiency. The monetization path of xAI clearly points to developers and enterprises: Grok Build coding agent (already open source), Office plug-ins, Connectors, Skills, voice agents, and distribution through channels such as Bedrock and Databricks.

• 5. User reviews (967 chars)

  Positive feedback focused on "fast, accurate, cheap, and able to do the job." Many developers said on X that they have switched their main model from Opus 4.8 to Grok 4.5; a creator who has written 7 small projects in a row using Grok 4.5 believes that it is close to the level of Opus in basic to intermediate programming tasks, and "cheapness is its most important function." After actual testing, Tencent Cloud engineers found that it is willing to admit incomplete information in uncertain scenarios, single out items to be confirmed, and is particularly good at interface design and troubleshooting. However, it can also lead people into long troubleshooting lines by over-explaining "why something went wrong." There are two main types of negative voices: one is that the experience of creative writing and one-time drafting is average, and 3D web page generation is rated as "rough and half-finished"; the other is trust issues - the community is very divided on Grok's political leanings. Some people think it has been kidnapped by "political correctness", and some people think it is more docile than GPT and Gemini. The controversy is unresolved.

• 6. Industry evaluation (819 chars)

  Independent reviews place Grok 4.5 at the forefront of cost-effectiveness. The Artificial Analysis Intelligence Index gave it 54 points, ranking fourth, behind Claude Fable 5, GPT-5.5, and Opus 4.8, but ahead of all open source and Gemini models. Musk’s own positioning is honest: Grok 4.5 is about equal to Opus 4.7, but much faster and cheaper. The industry generally believes that xAI has shifted competition from "who is the strongest" to "who can complete a single task most cost-effectively in real scenarios." The price war narrative is ignited by open source models such as DeepSeek and pushed to a climax by Grok 4.5. Some analysts also pointed out that Grok 4.5 repeatedly gives points to its own dominant tests and avoids pure reasoning evaluations such as GPQA. It is recommended to use independent scoring as the criterion; its AutomationBench-AA ranked first with 51.4% (the only model to break 50%), indicating that Agent tool invocation is indeed its strength.

• 7. Controversies and Risks (1066 chars)

  The most prominent thing is that the hallucination rate has doubled: the Grok 4.5 accuracy rate on AA-Omniscience has increased from 35% to 52%, but the hallucination rate has increased from 25% to 54%, that is, when the answer is wrong, it is more likely to be confidently fabricated, which is not friendly to high-risk tasks with no one checking. The second is the security compliance gap: Grok 4.5 did not come with a model card/system card when it was released, while Opus 4.8 and GPT-5.6 both issued capability threshold assessments; under the transparency requirements of the EU AI Act, this is an item to be filled. After the release, security concerns arose - a bug in the model jailbreak and the Grok Build CLI silently uploading user local files (packet capture showed that a complete warehouse of about 5.1 GiB was uploaded at one time, far exceeding the approximately 192 KB required by the model). Grok Build was subsequently open sourced and the default retained data was deleted to calm it down. In addition, the overlap of Cursor training data may increase the CursorBench score; Musk's "AGI / ASI" narrative (such as the private timeline of "Grok 6 = ASI") has long been accompanied by delays and unverified claims, and needs to be discounted before implementation.

• 8. Applicable groups and usage suggestions (663 chars)

  Grok 4.5 is suitable for developers, testers, technical products, and people who often work in environments with "incomplete requirements, insufficient logs, and inconsistent interfaces." It is especially suitable for high-throughput programming agents and workflows that require real-time news. It is not suitable for scenarios where one-time refinement of long articles is pursued, or answers are sent directly to end customers - a high hallucination rate means that a verification layer is necessary (running tests, citing sources, and converting low-confidence to manual work). Alternatives: If you want top-level inference, choose Claude Fable 5 / Opus 4.8, if you want balance, choose GPT-5.5, and if you want cheap open source, choose DeepSeek-V4, Qwen3, and Kimi K2.6.

• 9. Summarize (511 chars)

  Putting "Grok 6" back into reality: The current frontier of xAI is Grok 4.5, which relies on 1.5 trillion parameters, Cursor actual combat data and crushing single-task costs to tear a hole in the programming agent track; 4.6 (2 trillion parameters) is expected to take over in August, and the larger Grok 5 (6 trillion) is on the way. It wins in terms of efficiency and price, rather than absolute intelligence - if your workflow can self-verify, it is the most cost-effective choice on the cutting edge; if not, please wait for manual correction of naming in the background and careful evaluation before using it.

No in-depth report yet

• 

  ECox_88

  —

  The model cards have not been issued yet, so the compliance team has to raise questions.

---

• 

  bq4383k

  —

  I used it to write 7 small projects in a row. The biggest advantage is not the code itself, but the cheapness. Basic to intermediate programming tasks are close to the level of Opus, but the gap in creative scenarios of aesthetics and UI polishing is still there. This is the logic of collecting 60 billion Cursor: let the model understand "how to write code" rather than "what the code looks like". It is too early to tell whether this road will work, but the price anchor has already been moved by it.

---

• 

  AnthonyMiller

  —

  Cursor’s training data is indeed valuable. Ordinary code corpus only tells the model what the code looks like. Cursor’s real conversations can allow it to learn how to locate a bug, where to change it, and why users accept a certain diff. This is also the reason why Grok 4.5 is different from simply stacking large models in programming. Combination is very important.

---

• 

  sadlion744

  —

  The hallucination rate has doubled. You have to be wary of this. When answering incorrect questions on AA-Omniscience, you are more likely to make it up confidently. The workflow of running tests and quoting sources is fine, but throwing it directly to customers and waiting for customer service answers will be unreliable.

---

• 

  Jennifer_Lewis168

  —

  Our team has switched all high-throughput programming agents to Grok 4.5. For the same SWE-Bench Pro task, it only uses an average of 16,000 output tokens, while Opus 4.8 costs 67,000. The cost difference for a single task is 5 times faster. For scenarios where a large number of API calls are run every day, thousands of dollars can be saved a month, which is really something that cannot be ignored.

---

• 

  菊花_6

  —

  The community is quarreling over its political leanings. Some people say that it is kidnapped by political correctness, while some people actually think it is more docile than GPT. Both parties are in the same comment area, which shows that this matter is not finalized at all.

---

• 

  RegenRovtRoberts

  —

  I see many people boasting that "Grok beats Opus". In fact, Musk's own positioning is only at the level of Opus 4.7. The truth is that it is not strong in absolute intelligence, but strong in efficiency and price: the same job costs less tokens, runs faster, and is cheaper. My usage is to use it for daily coding. When I encounter complex logic, I temporarily switch to Opus. Mixing it saves half the time compared to pure Opus. This is the correct way to open it.

---

• 

  SAgre

  —

  Grok 4.5 smells so good, I can’t go back to Opus.

---

• 

  Margaret_Moore838

  —

  But sometimes it wants to explain why something went wrong, and will make a lot of guesses about encoding, newlines, and floating points. Novices can easily be led into long lines of investigation, and they have to restrain themselves.

---

• 

  Sophia.Gray369

  —

  After using it for two weeks, my biggest feeling is that it is willing to admit that "this piece of information is incomplete" and then lists the items to be confirmed. It is very convenient for interface debugging, which is more comfortable than the model that forces conclusions.

---

• 

  Donald_Collins_Pro

  —

  The price is $6 per million output, which is much cheaper than Opus. I can just switch to it for daily coding.

---

• 

  PSullivan_77

  —

  Grok Build is really brave to be open source, but it was also forced out by the silent upload of code. Now that I can run and audit locally, I have regained some trust.

---

• 

  DonaldVasquez_Pro

  —

  To be honest, the experience of creative writing and one-time drafting is average, and the 3D web page generation it makes is more like a rough semi-finished product. Some parts are done well, but there are also obvious bugs. If you want to write sophisticated and long articles, you still need Claude or GPT. It is more suitable as a second assistant in engineering rather than as the sole main force.

---

• 

  LUjon2024

  —

  The EU didn’t come online at the same time again, so it’s the same old script.

---

• 

  彭兰

  —

  Stuck in PPT? I can run more than 80 tokens per second without any lag at all.

---

• 

  d3s0h

  —

  The real-time Internet search is really powerful. It can pick out the graphics card parameters sent at three o'clock in the morning.

---

• 

  易建珍

  —

  It's a bit uncomfortable to cut the context from 1M to 500K. After feeding tens of thousands of words of technical documents, the detail recall occasionally jumps. However, the throughput of long text is still stable with Claude.

No reviews yet