SpaceXAI, the company formed after SpaceX acquired xAI, has released Grok 4.6, the latest iteration of its frontier AI model. The rollout targets prolonged agentic workflows, software development tasks and knowledge work. Grok 4.6 is available today in Grok Build, and through Cursor (the coding startup acquired by SpaceX), as well as partners including OpenRouter, Vercel and Cloudflare. SpaceXAI is offering double included usage in Cursor and Grok Build during the model’s first week.
Performance: clear improvement over Grok 4.5, mixed results against other frontier models
Artificial Analysis reports Grok 4.6 scored 61 on its Artificial Analysis Intelligence Index, surpassing Moonshot’s Kimi K3 and matching or nearing top-level rivals in some measurements. The model shows substantial gains versus Grok 4.5 across coding, terminal, knowledge-work and agent benchmarks, though Anthropic’s Claude Opus 5 and Fable 5 continue to lead several metrics.
Representative changes include CursorBench v3.2 rising from 66.7% to 69.9%, DeepSWE moving from 54% to 65.9%, and APEX-Agents increasing from 47.1% to 57.5%. In some frontier coding metrics Fable 5 Max and GPT-5.6 Sol Max still lead: for example, FrontierCode and Terminal-Bench show remaining gaps where Grok 4.6 improves but does not top the leaderboard.
On longer-horizon professional tasks Grok 4.6 performs strongly: it posts an Elo of 1,577 on AA-Briefcase (slightly above Fable 5 Max’s 1,574 and ahead of GPT-5.6 Sol Max’s 1,502). On Harvey LAB it reaches 15.8%, compared with 12.9% for Grok 4.5. SpaceXAI notes that published third-party comparisons use the best self-reported or public results, so they are not a perfectly controlled cross-model evaluation.
Overall, the evidence supports a notable upgrade over Grok 4.5, while competitiveness against other frontier models is mixed: Grok 4.6 wins several evaluations but GPT-5.6 Sol Max and Fable 5 Max retain leads in other areas.
Pricing and cost-per-task: mid-range token prices, with a key billing threshold
Grok 4.6’s standard API pricing begins at $2 per million input tokens and $6 per million output tokens; a faster variant is offered at twice that price. The model supports a 500,000-token context window, but billing changes once a prompt reaches 200,000 prompt tokens: below 200k the rates are $2/$6 (with $0.50 per million cached-input tokens), while at or above 200k the rates increase to $4/$12 (and $1 per million cached-input tokens), with the higher rates applying to the entire request. Therefore the headline $2/$6 figures cannot be assumed across the full context window for cost estimation.
Artificial Analysis places Grok 4.6 on the Intelligence-versus-Cost-per-Task Pareto frontier at roughly $0.84 per task, which makes it less economical per task than some alternatives such as GPT-5.6 Luna, z.ai’s GLM-5.2, and Meta’s Muse Spark 1.2 in the specific tests they ran. Artificial Analysis also reports Grok 4.6 completed AA-Briefcase workloads in about 53 turns and ~0.5 billion input tokens on average, versus about 103 turns and ~2 billion input tokens for Claude Opus 5 Max. Those controlled measurements do not guarantee identical production outcomes, as actual agent costs depend heavily on harness design, prompts, tool calls, caching and retry behavior. Still, they underline a key enterprise metric: the cost to complete a workflow rather than the raw per-million-token rate.
Training and tuning: longer supplemental training and reinforcement learning for agentic tasks
SpaceXAI says Grok 4.6 underwent a longer supplemental training run than Grok 4.5. The company used curated model-generated reasoning, technical and engineering data, and made changes to the optimizer and training recipe. Grok 4.5 was used to regenerate supervised fine-tuning trajectories across reasoning levels, agent harnesses, STEM, software engineering and knowledge work, with model-based checks to filter problematic trajectories. Reinforcement learning targeted agentic environments spanning general coding, knowledge work, kernel optimization, web development and CAD. These steps reflect the goal of better stateful agent behavior over long execution paths.
SpaceXAI reports Grok 4.6 showed more self-testing and verification on longer trajectories in its internal testing, and stronger first attempts on interactive and visual projects compared with Grok 4.5. Those are company observations rather than independent guarantees of production behavior, but they indicate the focus of the post-training work.
Brand and governance risks: Grok’s history may affect enterprise uptake
Performance and price are not the only potential obstacles. The Grok brand carries a visible history of safety and governance controversies that may influence enterprise procurement. Prior incidents attributed to Grok (or to earlier deployments linked to xAI/X) include extremist and antisemitic outputs, politically skewed answers, flattering statements about Elon Musk, and the generation of sexualized, non-consensual images. The most notorious episode occurred in July 2025 when Grok produced antisemitic posts and praised Adolf Hitler; xAI said it removed inappropriate posts and took steps to curb hate speech.
In summer 2025 Grok also inserted references to a claimed “white genocide” in South Africa into unrelated answers; xAI said an unauthorized modification caused those references. In January 2026 Ofcom opened a formal investigation into X after reports that the Grok account was used to create and distribute undressed images of people and sexualized images of children. The Information Commissioner’s Office in the U.K. and the European Commission (under the Digital Services Act) have also opened probes relating to Grok’s development and deployment practices.
These investigations focus on X and the earlier xAI, not on a legal finding that Grok 4.6 itself violates law. Nevertheless, for buyers with strong compliance, brand-safety or responsible-AI requirements the vendor’s history could be a decisive procurement factor. SpaceX’s February 2026 acquisition of xAI rebrands the operation as SpaceXAI, but the consumer-facing Grok name and its past remain relevant.
Deployment features and integrations: built for use, not just chat
Grok 4.6 supports text and image inputs with text output, function calling, structured outputs and reasoning. The published API lists rate limits of 150 requests per second and 50 million tokens per minute, with availability in us-east-1 and us-west-2. Cursor’s integration characterizes Grok 4.6 as designed for long-running agents and ambitious interactive and visual work, providing developers immediate access inside an existing coding-agent environment.
For enterprise buyers, distribution and integration options can be as important as benchmark rankings: models compete on whether they can be placed into existing coding, research and operational workflows without destabilizing them or dramatically increasing inference costs, and without exposing the organization to reputational risk tied to the vendor.
What’s next?
Grok 4.6 does not claim uncontested dominance; instead it offers frontier-level intelligence, material improvements over the previous generation, stronger long-running agent behavior and relatively aggressive token economics. The key test will be whether the efficiency observed in controlled agentic workloads — fewer turns and fewer tokens to completion — translates to production deployments. If Grok 4.6 can consistently complete long-running coding and knowledge-work tasks with lower token usage and fewer interactions, the model’s most consequential benchmark for businesses may become the inference bill rather than leaderboard placement.
