Grok 4.6 Release: Benchmarks & Analysis

Grok 4.6 Ships From SpaceXAI: What Actually Landed on August 12

Grok 4.6 arrived on August 12, 2026, roughly five weeks after Grok 4.5. Not a teaser, not a benchmark leak, but a live model in Cursor, Grok Build, and the API on the same afternoon.

TLDR SpaceXAI released Grok 4.6 with a 61 on the Artificial Analysis Intelligence Index, matching GPT-5.6 Sol and sitting two points behind Claude Opus 5, while holding pricing flat at $2/$6 per 1M tokens. The headline "matches GPT-5.6 Sol" claim is true only on the nine-benchmark composite. Read xAI's own ten-row table and Claude Fable 5 wins the most rows. Grok 4.6 is strongest on knowledge work and legal reasoning, weakest on terminal use, and its main selling point is price-to-intelligence, not raw leadership.

Key Takeaways

• Grok 4.6 scored 61 on the Artificial Analysis Intelligence Index, up 5 points from Grok 4.5 and +23 from Grok 4.3.
• Pricing stayed flat at $2/$6 per 1M input/output tokens, roughly 60%+ below GPT-5.6 Sol's $5/$30, according to Artificial Analysis.
• On xAI's ten-row eval table, Claude Fable 5 Max takes the most first-place rows; Grok 4.6 wins knowledge work and legal.
• It loses Terminal-Bench v3.0 by roughly 8.6 points to GPT-5.6 Sol, a weakness the launch post's prose skipped.
• Context window is 500,000 tokens, unchanged from Grok 4.5.
• No independent third-party replication of the numbers exists yet as of August 13, 2026.

What Actually Shipped

The confirmed facts are unusually clean for a same-day release. SpaceXAI's official announcement describes Grok 4.6 as building on Grok 4.5 with a focus on Long-Running Agents and more ambitious interactive and visual work. Michael Truell called it a "more capable digital colleague", framing it as Opus-class intelligence at low cost and high speed.

Pricing is the anchor fact. Per the Cursor release post, Grok 4.6 starts at $2 per million input tokens and $6 per million output tokens, with a fast variant at twice that. That matches Grok 4.5 exactly. Holding headline pricing flat across a generation is uncommon at the frontier, where intelligence gains usually arrive with price hikes.

Availability was immediate and broad. OpenRouter confirmed it went live the same day, restating the $2/$6 pricing. Lumina noted it live on Grok Build, and Ashutosh Shrivastava reported it inside Cursor. The x.ai post lists Cursor, Grok Build, the SpaceXAI API, and partners including OpenRouter, Vercel, and Cloudflare.

Source: @OpenRouter

NVIDIA also weighed in on the hardware side. It congratulated the SpaceXAI team and stated Grok 4.6 was trained and runs on NVIDIA GB300 NVL72 with NVLink. That is one of the few infrastructure details on record.

The Training Story

The training recipe is more candid than most launch posts. Per the official x.ai writeup, Grok 4.6 ran a longer supplemental training pass than 4.5, with curated model-generated data, high-quality engineering data, and an improved optimizer.

The interesting step is what came next. xAI used Grok 4.5 itself to regenerate the SFT trajectories across reasoning efforts and agent harnesses, spanning STEM, software engineering, and knowledge work, then filtered bad traces with model-based checks. The previous model taught this one. The RL stage was then pointed at agentic tasks: kernel optimization, web development, and computer-aided design.

Self-Verification is the behavior xAI highlights from this process. On longer trajectories, the model reportedly does more self-testing, checking its own work before moving to the next step. That is a claim to test, not a measured number.

Reading the Ten-Row Eval Table

Here is where the "matches GPT-5.6 Sol" headline needs scrutiny. That statement is true on the Artificial Analysis Intelligence Index, a composite of nine benchmarks, where Grok 4.6 scores 61 to GPT-5.6 Sol's 61. But xAI's own table has ten rows, and the marketing reads the winning ones.

A third-party review laid out the full table. Counting first-place finishes tells a different story than the composite:

Evaluation	Grok 4.6 High	Grok 4.5 High	GPT-5.6 Sol Max	Fable 5 Max
AA Intelligence Index	61	56	61	62
GDPVal-AA v2	1753	1526	1728	1741
CursorBench v3.2	69.9%	66.7%	67.2%	70.5%
DeepSWE v1.1	65.9%	54%	73%	70%
FrontierCode v1.1	61.3%	56.6%	60.6%	63.6%
APEX-Agents	57.5%	47.1%	56.7%	59.2%
Terminal-Bench v3.0	26%	15.7%	34.6%	34.1%
APEX-SWE	56.4%	53.6%	—	58.8%
AA-Briefcase	1577	1313	1502	1574
Harvey LAB (Vals)	15.8%	12.9%	2.5%	11.3%

Count the row winners: Claude Fable 5 Max takes five, Grok 4.6 takes three, GPT-5.6 Sol Max takes two. The composite tie with GPT-5.6 Sol holds, but the model that wins the most individual rows is Fable 5.

Grok 4.6's wins cluster in two areas. Knowledge work is its strength: it leads GDPVal-AA v2 at 1753 and AA-Briefcase at 1577. The legal gap is the most dramatic on the board. On Harvey LAB it scores 15.8% against GPT-5.6 Sol's 2.5%, a more than sixfold margin.

The weakness is just as clear. On Terminal-Bench v3.0, Grok 4.6 posts 26% against 34.6% for GPT-5.6 Sol and 34.1% for Fable 5, trailing by roughly 8.6 points. This benchmark measures work inside a terminal. The launch post's prose does not mention this row; it shows up only in the table. On DeepSWE v1.1 it also trails GPT-5.6 Sol, 65.9% to 73%.

The generational jump over Grok 4.5 is real regardless. Terminal-Bench moved from 15.7% to 26%, APEX-Agents from 47.1% to 57.5%, DeepSWE from 54% to 65.9%, and AA-Briefcase from 1313 to 1577. Five points on the Intelligence Index in five weeks is a genuine step, not a version-number bump.

The Cost Story Is the Real Headline

Strip away the leaderboard positioning and the durable claim is about price. Per Artificial Analysis, Grok 4.6 delivers a 5-point Intelligence Index gain at unchanged $2/$6 pricing, with a measured cost per task of $0.84.

The comparison that matters is against the models scoring within two points. Claude Opus 5 lists at $5/$25 and GPT-5.6 Sol at $5/$30. At roughly one-fifth the output token price of GPT-5.6 Sol for effectively the same composite score, Grok 4.6 is the best price-to-intelligence deal at the frontier right now. That is the sentence other analysts will quote.

Artificial Analysis also flagged efficiency. Grok 4.6 reportedly completes tasks in about 53 turns and 0.5B input tokens on average, against roughly 103 turns and 2.0B input tokens for Claude Opus 5 (max). Turn-Efficiency, defined as resolving the same task in fewer agent steps, is a cost lever that compounds in reasoning-heavy workloads. Cache-hit pricing did rise, though: from Grok 4.5's $0.3 per 1M tokens to $0.5 per 1M for Grok 4.6.

If you want to run your own price-to-intelligence check against a current frontier chat model before trusting the $0.84-per-task figure, a comparable option to benchmark against is Claude Opus 5, the model sitting two points ahead on the same index.

Grok 4.6 vs GPT-5.6 Sol: What the Reports Say

GPT-5.6 Sol is the most-cited comparison across the reports, so it is the cleanest head-to-head. Here is what the reports actually support.

• Composite intelligence. Both score 61 on the Artificial Analysis Intelligence Index, per Artificial Analysis. A genuine tie on the headline number.
• Coding and terminal use. GPT-5.6 Sol leads. It takes DeepSWE v1.1 (73% vs 65.9%) and Terminal-Bench v3.0 (34.6% vs 26%) per the third-party table.
• Knowledge work. Grok 4.6 leads. It tops GDPVal-AA v2 (1753 vs 1728) and AA-Briefcase (1577 vs 1502).
• Legal reasoning. Grok 4.6 wins decisively on Harvey LAB, 15.8% to 2.5%.
• Pricing. Grok 4.6 is far cheaper at $2/$6 versus GPT-5.6 Sol's $5/$30, per Artificial Analysis.

The pattern is consistent: GPT-5.6 Sol is the stronger software-engineering model, Grok 4.6 the stronger knowledge-work model, and the two tie on the composite. On the limited evidence so far, Grok 4.6 appears to win on cost-adjusted value rather than raw capability, but no independent replication confirms these numbers yet.

What We Know vs. What We Don't

The clean split matters here because most of the loud benchmark claims circulating are repeated by commentators, not independently tested.

What we know:

• SpaceXAI released Grok 4.6 on August 12, 2026, roughly five weeks after Grok 4.5, per the official x.ai announcement.
• Pricing is $2/$6 per 1M input/output tokens with a fast variant at twice the price, per the Cursor blog.
• The context window is 500,000 tokens, unchanged from Grok 4.5, per Artificial Analysis.
• It scored 61 on the Artificial Analysis Intelligence Index, matching GPT-5.6 Sol and gaining 5 points over Grok 4.5.
• It is available in Cursor, Grok Build, the SpaceXAI API, OpenRouter, Vercel, and Cloudflare, per the official announcements.
• Training used a longer supplemental run, Grok 4.5–regenerated SFT trajectories filtered by model-based checks, then agentic RL, per the official x.ai post.

What we don't know:

• No independent third-party replication of the benchmark numbers exists as of August 13, 2026; the figures are published by xAI or repeated by commentators.
• Whether Grok 4.6 is broadly ahead of competitors or only strong on selected benchmarks remains disputed; on the ten-row table Claude Fable 5 wins the most rows.
• Whether identical 500K context and $2/$6 pricing apply across the API, Grok Build, Cursor, and other providers is unconfirmed.
• A third-party review reported a 65.7% non-hallucination rate from Artificial Analysis, but that figure is not independently corroborated in the reporting available so far.
• Benchmark stability is unsettled; one arena reported results were still tightening confidence intervals and needed more votes.
• Hands-on quality is mixed and early; some testers reported strong broad performance while others reported worse results than GPT-5.6 Sol after short sessions.

Why This Matters for Builders

The takeaway is not "Grok 4.6 is the best model." It is that the frontier now has a genuine price-to-intelligence outlier. Agent Native framed it as SpaceX catching up, and Emily noted Grok 4.6 and DeepSeek V4 Pro shipping the same day as a sign the cadence is compressing.

For a builder, that changes the routing math. If your workload is knowledge work, long-horizon agents, or legal-style reasoning, Grok 4.6's row wins plus its price make it a serious default candidate. If your workload is terminal-heavy software engineering, the DeepSWE and Terminal-Bench gaps suggest GPT-5.6 Sol or Fable 5 may still earn their premium. The verdict flips entirely on workload, which is exactly why the composite score is a weak decision input on its own.

The non-hallucination figure, if it holds, is the one to weigh most. A reported 65.7% non-hallucination rate means roughly a third of outputs may carry a confident error, which matters far more in a customer-facing agent than a two-point Intelligence Index gap. Retrieval boundaries and confidence gates do more for that than any benchmark row.

How to Evaluate It Yourself

Do not trust the launch table as a routing decision. The numbers come from xAI's own materials, with competitor figures drawn from published system cards, and no unified third-party recount exists yet.

Three practical checks. First, run your own coding eval before relying on any single row; the DeepSWE gap to GPT-5.6 Sol is 7.1 points, small enough that harness effects could swamp it. Second, test the Self-Verification behavior directly by asking the model to verify functionality after implementing and iterate until production-ready, then watch whether it actually reopens and re-tests its own output. Third, measure cost per task on your real workload rather than the reported $0.84, since turn count varies wildly by task type.

What to Watch Next

Three signals are worth tracking over the next week. Watch for an independent benchmark recount from a third party that reruns the ten-row table rather than restating xAI's figures. Run your own knowledge-work and terminal evals before trusting the composite tie with GPT-5.6 Sol, since the row-level story diverges sharply. And pin the non-hallucination rate: confirm whether the reported 65.7% figure survives a second measurement pass before putting Grok 4.6 in front of real users.

Try related models on Kie.ai: Grok 4.6, GPT-5.6 and Claude Fable 5.