SpaceXAI ships Grok 4.6: it ties the GPT-5.6 Sol tier on the composite index, for a third of the price — but the coding benchmarks tell a narrower story

What shipped

SpaceXAI released Grok 4.6 on 12 August, roughly five weeks after Grok 4.5 went public. The framing matters: this is a post-training refresh, not a new base model. SpaceXAI held the foundation constant and spent the improvement on a longer supplemental training run, regenerated supervised fine-tuning across reasoning levels and domains, and reinforcement learning in agentic environments. One behavioural change stands out — the model now does more self-testing and verification, checking its own work on extended tasks.

The specs, confirmed across Artificial Analysis, MarkTechPost and SiliconANGLE:

• Context window: 500,000 tokens. Text and image input, text-only output. Knowledge cutoff of 1 February 2026. Parameter count unpublished.
• Reasoning levels: four now — low, medium, high (the default), and a new xhigh for harder problems.
• Availability: the SpaceXAI API (model ID grok-4-6), Cursor on all plans, Grok Build as the default model, and routing via OpenRouter, Vercel and Cloudflare. No open-weights release, no self-hosting.
• Pricing: $2 input / $0.50 cached / $6 output per million tokens under 200K; $4 / $1 / $12 above 200K. A faster-serving variant costs roughly double.

The one number everyone will quote — and what it hides

Grok 4.6 scores 61 on Artificial Analysis’s composite Intelligence Index, up from 56 for Grok 4.5. That level ties GPT-5.6 Sol Max and, per VentureBeat, places it among the top three on the index — overtaking Moonshot’s Kimi K3 and closing most of the gap to the frontier. It still sits behind Fable 5, the ceiling model in Anthropic’s current lineup.

A single composite score is exactly the thing to distrust, though, because it averages away the shape of the gains. Break it apart and Grok 4.6 is two different models depending on the workload:

• Knowledge and analysis work — strong. It leads the table on GDPval-AA v2 (1753 Elo, up from 1526), AA-Briefcase (1577, up from 1313), and Harvey’s legal benchmark. If your use is research, drafting, briefing, financial or legal analysis, this is where the model is genuinely competitive with the frontier.
• Agentic coding — middling. DeepSWE v1.1 lands at 65.9% — a real 11.9-point generational jump, but behind GPT-5.6 Sol Max at 73%. Terminal-Bench v3.0 reaches 26%, nearly double Grok 4.5’s 15.7% and still last of the four models Artificial Analysis lists. CursorBench v3.2 is a healthier 69.9%.

Artificial Analysis itself flags that some of the wins fall inside statistical confidence intervals — “statistical ties, not leads.” So the honest summary is: Grok 4.6 has closed the gap to the GPT-5.6 Sol tier on average, leads it on knowledge work, and still trails it where autonomous coding is measured.

The real story is distribution, not the benchmark

The benchmark line will get the headlines, but the more consequential fact is where Grok 4.6 runs. It is the default model in Cursor, one of the most widely used AI coding environments. That placement is not a neutral integration — SpaceX signed a $60 billion deal to acquire Cursor’s maker, Anysphere, in June 2026, and while that acquisition has not yet cleared regulatory review, Grok is already the model millions of Cursor users touch first.

That is how a value-tier model becomes a default without winning a benchmark war: not by beating Claude or GPT-5.6 on Terminal-Bench, but by being wired into the tool at a price the platform can afford to serve broadly. For an AI-tools buyer, the distribution question — what model does my IDE reach for by default, and can I change it — is now as important as the leaderboard.

Why this matters

The frontier is getting cheaper faster than it’s getting better. Grok 4.6 matches a tier that cost far more a few months ago, at $2/$6. The generational story across 2026 has been less about a new ceiling and more about the price of near-ceiling capability collapsing. For most buyers that is the more useful trend: the model you could not justify on cost last quarter is now a third of the price.

“Matches on the index” is not “matches on your task.” The split between Grok 4.6’s knowledge-work strength and its agentic-coding gap is the whole point. A composite score is a marketing artefact; your workload is not the composite. Teams doing research, analysis and drafting will find it genuinely frontier-adjacent; teams running long-horizon autonomous coding agents will still feel the distance to the top Claude and GPT-5.6 tiers.

Latency is the unglamorous catch. Grok 4.6’s time-to-first-token is about 42 seconds — far above the sub-three-second median for models in its price band — and output runs around 67 tokens per second, slightly below median. For batch and agentic jobs that is tolerable. For interactive back-and-forth it is a real friction, and it is the kind of number that never appears in a launch headline.

No open weights keeps this a rental. Unlike the open-weight tier led by Kimi K3, Grok 4.6 is API-and-product only. There is no self-hosting, no offline option, and no way to pin a version you control. For regulated or air-gapped environments that rules it out regardless of the benchmark.

Honest caveats

The benchmarks are early and partly vendor-adjacent. Artificial Analysis is an independent evaluator, which is why it anchors this piece, but several figures come from SpaceXAI’s own reporting on its API and some wins are statistical ties. Treat the composite score as directional, not decisive, and wait for a broader independent read on real tasks.

“Optimised for long-running agents” is a claim, not a measurement. The marketing positions Grok 4.6 for autonomous coding and multi-step work, yet the agentic-coding benchmarks are precisely where it trails. The self-verification behaviour may help in practice, but that is exactly the kind of improvement that needs independent, task-level testing to confirm.

Pricing has a cliff. The attractive $2/$6 rate doubles above a 200K-token prompt. Long-context work — the thing a 500K window invites — is billed at $4/$12, and per xAI’s own documentation the higher rate applies to all tokens in the request, not just those past the threshold. That materially changes the cost calculus for exactly the workloads the big context is meant to serve. It is also unresolved on Amazon Bedrock, whose model card publishes flat rates with no long-context band — see our breakdown of Grok 4.6’s Bedrock listing and its data-residency premium.

The verdict

Grok 4.6 is a clean, credible iteration: a five-point index gain, real progress on coding, unchanged pricing, and a smart new xhigh setting for hard problems. On the composite number it has pulled level with the GPT-5.6 Sol tier, and on knowledge work it is genuinely competitive with the frontier at a fraction of the cost.

But the composite number oversells it. This is a value and distribution play, not a new capability leader. It trails Fable 5 overall, trails GPT-5.6 Sol Max where autonomous coding is measured, sits last on Terminal-Bench, and carries a 42-second cold-start that interactive users will feel. Its biggest advantage is not a benchmark at all — it is being the default inside Cursor, at a price that makes broad deployment cheap.

Recommendation: if you already live in the SpaceXAI or Cursor orbit, or you run knowledge and analysis workloads where cost per task matters, Grok 4.6 is worth putting into your own evaluation now — it may be the best value at its capability level. If you run long-horizon autonomous coding, or you need open weights, offline use, or low interactive latency, the proven frontier options — the top Claude and GPT-5.6 tiers — remain the safer default. As always: benchmark on your own tasks before you move anything you depend on.

Questions or corrections? Email Pick Right. Want the full list? See all news.