Grok 4.6 Release Date: It Shipped August 12 at the Same $2/$6

On Wednesday, August 12, 2026, the question this page was built around stopped being a question: Grok 4.6 shipped. SpaceXAI — the company that has operated Grok since it acquired the startup in February 2026 and rebranded in July — released it at the same $2/$6 price as Grok 4.5, on the same V9 foundation, with the same 500K-token context window, and Musk called it "a banger" on X. The launch settled the two questions the previous version of this page was tracking: the model exists, and it is a same-scale post-training refresh rather than a scale-up. The one thing it did not settle is the parameter count, which remains officially undisclosed.

A week ago the honest headline was that the model had not shipped: the "around August 7" target Musk set on July 27 had slipped, and the same-scale-reading was still a prediction. This page argued the model was best read as a refresh at unchanged scale, and that refreshes at unchanged scale usually ship without a price increase. Launch day confirmed the business half of that call exactly — $2/$6/$0.50 per million tokens, identical to Grok 4.5 — and the first independent measurements have now come in. Below is what shipped, what the numbers say so far, and what is still vendor-stated.

What shipped on August 12

Grok 4.6 is live on the SpaceXAI API as model id grok-4.6, and in Grok Build, Cursor, and the Grok Bot app; Cursor and Grok Build users get double their included usage for the first week. It is a reasoning model with extended chain-of-thought, accepts text and image input and produces text, and adds a new xhigh reasoning-effort setting above the default.

The API keeps OpenAI-compatible endpoints and adds three tools on the server side: web search, X search, and code execution. For anyone already talking to Grok 4.5 over an OpenAI-compatible endpoint, that means the switch is a model-string change rather than a client rewrite.

What this page predicted — and what landed

The previous version's argument was that Musk's "the 1.5T model with significantly improved SFT & RL" pointed to a same-scale refresh, and that the price would hold. The launch confirmed both halves of the business call:

• Price: $2 / $6 / $0.50 (cache-hit) per million tokens — identical to Grok 4.5. The "no price change" prediction held.

• Context: 500K tokens, identical to Grok 4.5.

• Scale: still officially undisclosed. No parameter count appears in the model card, so Musk's 1.5T figure remains the best available claim, not a verified spec. The 2T number that circulated in early coverage was a misattribution of Grok 4.7's 2.1T.

The independent number — 61 on the Artificial Analysis Intelligence Index

The first third-party measurement is in, and it is a good one. Artificial Analysis scores Grok 4.6 at 61 on its Intelligence Index — tied with GPT-5.6 Sol's max-reasoning configuration, one point behind Claude Fable 5 (62), two behind Claude Opus 5 (63), and ahead of Kimi K3 (60). That is a five-point jump over Grok 4.5's 56, which is a large gain for a post-training refresh.

Two caveats come with the number. The Index is a composite, and the individual dimensions are more mixed. And Grok 4.6 is slow to start: a 42-second time to first token against a ~2.9-second median for its price tier, and roughly 68 output tokens per second, slower than the class average. That is noise for a long-running agent and the whole experience for interactive chat.

On the intelligence-versus-cost Pareto frontier the picture is flattering: roughly $0.84 per task, completing a typical AA-Briefcase workload in about 53 turns and ~0.5B input tokens against ~103 turns and ~2B for Claude Opus 5 Max. That makes it a legitimate frontier bargain, not merely a cheap one.

What is still vendor-stated

SpaceXAI's launch claims — "intelligence comparable to GPT-5.6 Sol and Claude Fable 5," improvement over Grok 4.5 on every listed evaluation, attributed to "a longer supplemental training run, stronger engineering data, and expanded reinforcement learning" — are vendor-stated and not yet independently verified. The launch benchmark table (CursorBench, FrontierCode, APEX-Agents, Terminal-Bench, and the promoted 1,753 Elo on GDPval-AA v2) is xAI-reported; none of those figures has yet been reproduced by a lab that does not sell the model.

The first third-party read is in (preliminary)

The one independent arena result so far is preliminary but real. LMArena lists Grok 4.6 (as grok-4.6-high) at #43 on its Text overall leaderboard with 1,464 Elo on 2,448 votes — flagged preliminary, and four points below Grok 4.5's 1,468 debut. The sample is too small for LMArena's tracked frontier table, which needs a two-week sustained score.

The Agent leaderboard of live sessions has not published yet. When it does, it will be the first read that cannot be optimised against in advance.

The price math

At $2/$6, Grok 4.6 is the cheapest frontier API of the current generation by a wide margin. Against GPT-5.6 Sol at $5/$30, input is 60% cheaper and output 80% cheaper — and because the gap is wider on output, the "half the price" marketing framing understates the saving on output-heavy coding workloads.

The fine print: requests above 200K input tokens step up to a $4 / $12 / $1 tier, and a faster variant costs double. For switching, the arithmetic barely moves — Grok 4.5 and Grok 4.6 share price and context, so moving between them is a configuration change, not a budget decision.

Because OrcaRouter passes provider list prices straight through at 0% markup, the $2/$6 on a grok-4.6 endpoint is SpaceXAI's own number, not a resold rate — a vendor price cut would land here the same day, and failover to a second provider is a routing rule rather than a procurement exercise.

Grok 4.7 and Grok 5 are still ahead

Grok 4.7 is the next step, and Musk's description is unchanged: "the 2.1T model," better than Grok 4.6 "in every way, except slightly slower to serve." On August 5 he put it three to four weeks out, which lands in late August or early September. That timeline is Musk-stated, and it has drifted later at each telling, so treat it as directional.

Grok 5 sits at year-end, trained — per the August 4 earnings call — on "basically all the data SpaceX has ever generated."

What to check next

Four things will firm up the picture, and each currently reads as open:

• The LMArena Agent leaderboard, unpublished as of August 13.

• A third-party reproduction of the coding and agentic numbers — DeepSWE, CursorBench, APEX-Agents from someone not selling the model.

• The faster variant's real price and latency once it is widely available.

• Whether Grok 4.7's late-August timeline holds, given that each specific date so far has drifted a few days later.

And the one that actually matters for a decision: your own workload. A leaderboard position is a hypothesis about your task, not a result for it.

Frequently asked questions

When did Grok 4.6 actually launch? August 12, 2026, on the SpaceXAI API and in Grok Build, Cursor, and the Grok Bot app. Musk's July 27 "around August 7" target slipped five days; the "next week" phrasing from the August 4 earnings call was the accurate window.

What does Grok 4.6 cost? $2 / $6 / $0.50 per million tokens — identical to Grok 4.5 and far below GPT-5.6 Sol's $5/$30. Above 200K input tokens it steps up to $4 / $12 / $1, and a faster variant costs double.

Is Grok 4.6 a 2T model? No parameter count has been published. Musk's 1.5T figure remains the best available (unverified) claim; the 2T number that circulated was Grok 4.7's 2.1T misattributed to 4.6.

Is Grok 4.6 as good as GPT-5.6 Sol? On the Artificial Analysis Intelligence Index, the two are tied at 61 (max-reasoning configuration). The vendor claims it is comparable to GPT-5.6 Sol and Claude Fable 5. It is markedly slower on time-to-first-token and far cheaper; the agentic gains await independent verification.

Where this leaves you

The practical move is unchanged from the pre-launch version of this page. If your code already talks to Grok 4.5 via an OpenAI-compatible endpoint, Grok 4.6 is a same-price configuration change — flip the model string, keep the budget line, and re-run your evals because the build is new. If it does not, a single pass-through-pricing endpoint makes a new frontier model switchable without betting a production path on its first-week benchmarks, which is exactly what those preliminary arena numbers are.