Grok 4.6 holds the $2/$6 rate, then doubles it above 200,000 tokens

SpaceXAI announced Grok 4.6 on Wednesday, August 12, and left the base API rate where Grok 4.5 had it: $2 per million input tokens and $6 per million output tokens ^[1][2]. A higher score at an unchanged entry price is a procurement non-event and an engineering one, because it makes rerunning an existing eval harness roughly free while making a migration decision no more urgent than it was last week.

The number being sold is 61 on the Artificial Analysis Intelligence Index, five points above Grok 4.5, according to VentureBeat, which puts Grok 4.5 at 56 ^[3][19]. VentureBeat reported that the move takes Grok past Moonshot AI's Kimi K3 and level with OpenAI's GPT-5.6 Sol Max ^[4]. SpaceXAI's own published evaluation table also shows 61 ^[5]. Five index points on a composite is enough to justify replaying your own tasks; it is not a substitute for them.

The pricing needs the qualification the headline rate omits. The $2 and $6 rates apply only while a prompt stays below 200,000 tokens; at or above that threshold, the rates double to $4 and $12 ^[6]. The doubled rate then applies to all tokens in the request rather than to the overage alone ^[7]. The documentation lists a 500,000-token context window ^[8], which means 300,000 tokens of the advertised context sit on the expensive side of the boundary ^[20]. A 201,000-token request is billed at twice the per-token rate of a 199,000-token request ^[21].

That structure interacts badly with the thing the model is being positioned for. SpaceXAI pitches Grok 4.6 as an agent that researches unfamiliar subjects, navigates a codebase, operates tools and turns a broad product idea into working software over a sequence of steps ^[22]. As the source notes, for an agent that repeatedly reads a repository, tool results and its own prior work, the 200,000-token boundary can materially change the bill ^[23]. SpaceXAI remains below several premium frontier models in VentureBeat's comparison, but the useful measurement is a full trajectory, not the first line of the pricing table ^[18]. There is also a faster Grok 4.6 variant at twice the standard price, which is a second axis to test rather than a default ^[10].

On provenance: SpaceXAI says Grok 4.6 got a longer supplemental training run than 4.5, using curated model-generated reasoning and technical data, engineering data, an updated optimizer and a revised recipe ^[11]. Grok 4.5 then generated new supervised fine-tuning trajectories across reasoning, software engineering, STEM and knowledge work, with model-based checks filtering problematic traces ^[12], and reinforcement learning focused on coding ^[13]. A pipeline that heavily grades its own output is another argument for weighting your held-out tasks over the vendor's table.

Distribution was Cursor and Grok Build at launch ^[9]. Cursor also serves models from labs that compete with SpaceXAI directly ^[17], and SpaceX has agreed to acquire Cursor parent Anysphere for approximately $60 billion in stock, a deal AP reported is expected to close in the third quarter of 2026 subject to conditions and regulatory approval ^[16].

Watch three things: whether the 200,000-token cliff pushes teams into context compaction work they had stopped doing, whether the Anysphere close changes which models Cursor surfaces by default, and what SpaceXAI publishes on the terminal-work gap that runtimewire flags but does not quantify in the available text ^[24].