Where to begin

On Wednesday, August 12, Eastern Time, SpaceXAI, led by Elon Musk, officially launched its next-generation large language model, Grok 4.6. The company stated that the new model delivers significant improvements over Grok 4.5, with a particular focus on strengthening long-running autonomous agents, complex coding, knowledge work, and interactive and visual task capabilities.

The most notable aspect of this release is not simply the model iteration, but that Grok 4.6 has already entered the top tier in multiple frontier model tests announced by SpaceXAI. Additionally, its API starting price is set at just $2 per million input tokens and $6 per million output tokens. SpaceXAI claims this pricing is roughly half that of other frontier models. In the first week of launch, users on Cursor and Grok Build also receive double the included usage.

Test data released by SpaceXAI shows that Grok 4.6 scored 61 points on the Artificial Analysis Intelligence Index, tying with OpenAI's frontier model, GPT-5.6 Sol Max. On the GDPVal-AA v2 benchmark, it achieved 1753 Elo, surpassing the 1728 points of GPT-5.6 Sol Max and the 1741 points of Anthropic's frontier model, Claude Fable 5.

Musk subsequently reposted on social media about Grok 4.6 being a major upgrade, scoring higher than Grok 4.5 High across all benchmarks. He highlighted the 1753 Elo rating and praised the new model as incredibly powerful.

Comprehensive enhancement of complex tasks from knowledge work to coding

The core upgrade of Grok 4.6 is no longer just about answering questions, but enabling the model to consistently complete complex tasks requiring multi-step execution.

SpaceXAI stated that Grok 4.6 is specifically trained for long-running agents, capable of continuously researching topics, analyzing information, processing codebases, and transforming a vague product idea into a runnable application or work output. The company also noted that the new model begins to exhibit more autonomous testing and verification capabilities over longer task trajectories, checking its previous work before proceeding.

In terms of training, Grok 4.6 underwent more extensive supplemental training than Grok 4.5, incorporating curated model-generated reasoning data, high-quality engineering data, and improved optimizers and training schemes. Subsequently, SpaceXAI used Grok 4.5 to generate SFT trajectories covering different reasoning intensities, agent frameworks, and STEM, software engineering, and knowledge work. It then further trained the model via reinforcement learning for tasks like general programming, knowledge work, kernel optimization, web development, and computer-aided design.

SpaceXAI specifically emphasized that Grok 4.6 has improved its ability to turn an "idea" into a "product." For a concrete product concept, the model can first research an unfamiliar domain, design the application structure, implement core interactions, and then perform multiple rounds of iteration based on feedback. In visual and interactive projects, the new model can also form a relatively complete first version more quickly.

Low-cost strategy targeting developer market with doubled usage in first week on platforms like Cursor

Compared to the model's capabilities themselves, the pricing strategy of Grok 4.6 may be more noteworthy for the AI industry chain.

SpaceXAI's announced API prices start at $2 per million input tokens and $6 per million output tokens, with a faster, double-priced version also available. The company also announced that Grok 4.6 has been integrated into Cursor and Grok Build, and is accessible through partners like OpenRouter, Vercel, and Cloudflare.

For comparison, in its announcement, SpaceXAI positioned Grok 4.6 as "half the price of other frontier models." Investment firm Atreides Management's Chief Investment Officer, Gavin Baker, made a more aggressive assessment: he believes Grok 4.6's performance is roughly equivalent to Fable 5 Max, but its input token price is 80% lower and output token price is 88% lower, giving it a clear advantage in the combination of performance and cost.

Tech industry analyst and editor-in-chief of the Superintelligence Newsletter, Kim Monnis, also pointed out that Grok 4.6 is not only cheaper than Opus 5 and GPT-5.6 Sol, but even lower than Sonnet 5, while its performance is at least at the level of top-tier models. In his view, this combination of "frontier performance and low cost" is Grok's true moat.

It should be noted that the API unit price does not fully equate to actual usage costs. Different models have variations in token consumption, reasoning intensity, and task completion paths. Therefore, the claim of being "half the price" is mainly a comparison of announced API prices, not an assertion that the final cost for all specific tasks will be exactly 50% lower.

The first-week promotion further reinforces this low-cost strategy: SpaceXAI stated that Cursor and Grok Build users can receive double the included usage in the first week, allowing developers to trial Grok 4.6 at a low cost.

Strengthening long-range execution with autonomous testing and continuous correction

From the test results, Grok 4.6's improvements are particularly concentrated in the area of agents executing real-world work, which is becoming a core competitive field for AI models.

SpaceXAI notably showcased the GDPVal-AA v2 benchmark introduced by Artificial Analysis. This test is based on OpenAI's GDPval dataset, evaluating model performance on real-world, economically valuable professional tasks across 44 occupations and 9 major industries. Models must complete tasks within an agent environment using tools like Shell and web browsing, with final output files compared in a blind pairwise manner to generate an Elo rating.

In the comparison data released by SpaceXAI, Grok 4.6 achieved 1753 Elo on GDPVal-AA v2, ranking higher than GPT-5.6 Sol Max's 1728 and Claude Fable 5 Max's 1741, and representing a significant jump from the previous generation Grok 4.5 High's 1526.

Simultaneously, Grok 4.6 scored 61 points on the Artificial Analysis Intelligence Index, tying with GPT-5.6 Sol Max. This index is a composite of nine tests, measuring the model's comprehensive capabilities across multiple dimensions including mathematics, science, programming, and reasoning.

In other tests published by SpaceXAI, Grok 4.6 also showed clear progress over the previous generation:

CursorBench v3.2: 69.9%, compared to Grok 4.5 High's 66.7%.

FrontierCode v1.1: 61.3%, compared to Grok 4.5 High's 56.6%.

APEX-Agents: 57.5%, compared to Grok 4.5 High's 47.1%.

APEX-SWE: 56.4%, compared to Grok 4.5 High's 53.6%.

AA-Briefcase: 1577 points, compared to Grok 4.5 High's 1313 points.

Harvey LAB: 15.8%, compared to Grok 4.5 High's 12.9%.

However, Grok 4.6 does not hold the top position in all tests. For example, in DeepSWE 1.1 and Terminal-Bench v3.0, the scores for GPT-5.6 Sol Max released by SpaceXAI remain higher than those for Grok 4.6. Therefore, a more accurate statement is that Grok 4.6 has entered the top tier of frontier models in multiple agent and knowledge work tests, rather than having comprehensively defeated all competitors.

It is important to note that the GDPVal-AA v2 leaderboard from Artificial Analysis itself changes with the addition of new models and evaluation updates. Thus, the claim of "bests" in the title specifically refers to the horizontal test results released by SpaceXAI at the time of this launch, and does not assert that Grok 4.6 ranks first in all evaluations at all times.

From model to ecosystem, SpaceXAI accelerates its capture of the AI development entry point

The release of Grok 4.6 also shows that SpaceXAI is extending the competition from pure model performance comparisons further into developer entry points and application ecosystems.

Grok 4.6 is now simultaneously available on Cursor, Grok Build, and the API, and is integrated into channels like OpenRouter, Vercel, and Cloudflare. Through these platforms, SpaceXAI hopes to allow developers to embed Grok 4.6 directly into coding, web development, and agent workflows, rather than just using the model through the Grok chat product.

Elon Musk himself has been actively supporting the new model on X. He first announced "Grok 4.6 is now out," then reposted reports about the 1753 Elo ranking, emphasizing Grok 4.6's first-place finish on GDPVal-AA v2. In another repost, he listed Grok 4.6's rankings in tests like AA-Briefcase, Harvey LAB, CursorBench, FrontierCode, APEX-Agents, and APEX-SWE, concluding with the remark: "Grok 4.6 is a banger."

Meanwhile, Gavin Baker from Atreides Management further directed market attention to the next-generation product. He predicts that Grok 4.7 will be a significantly larger model and may incorporate Cursor and SpaceX data into its pre-training. This statement is currently an assessment from an investment professional and is not a confirmed product roadmap from SpaceXAI's announcement.

If Grok 4.6 can maintain frontier-level model capabilities while sustaining a lower token price, its significance may extend beyond a single generation's performance upgrade. It could become a new signal that the AI large model price war is extending into the agent era. As the capability gap between models from different vendors narrows, the balance between performance and reasoning cost may increasingly become the key factor for developers and enterprises in choosing a model.