Grok 4.6, released Aug. 12, 2026, is not a new foundation model. It sits on the same V9 foundation as Grok 4.5 with post-training layered on: longer reinforcement learning runs, regenerated supervised trajectories, expanded agentic-environment RL. The result moved its Artificial Analysis Intelligence Index from 56 to 61. Elapsed time, roughly five weeks. Price, unchanged.
That single sentence contains most of what has happened to frontier model competition.
The stated target is long-running agentic work — whole-repository refactors and migrations. For context, the preceding Grok 4.5 ran on the V9 foundation at roughly 1.5 trillion parameters, about 3x the prior generation, while its context actually regressed from 4.3's 1 million to 500K.
The Artificial Analysis Intelligence Index (v4.1.1) is a weighted sum of nine evaluations across four categories: agentic 34%, coding 24%, scientific reasoning 24%, general 18%.
Grok 4.6 (high) scores 61, ranking 6th of 182 models. The top of the board:
Six models effectively inside two points. What matters here is not the rank but the cost spread among models on the same line. GPT-5.6 Sol, tied at 61, runs $1.23 per task. Grok 4.6 runs $0.84.
The Agentic Index measures tool use, planning, autonomy, and complex problem solving, as a weighted average of GDPval-AA v2 and τ³-Banking.
Grok 4.6 (high) scores 59, tied for the top spot with Claude Opus 5 (max) and GLM-5.3 (max). Qwen3.8 Max and GPT-5.6 Sol follow at 58, Claude Fable 5 at 57, Kimi K3 at 54. A model sitting 6th on overall intelligence joins the leading group on agentic work — that inversion is the point of the metric. On GDPval-AA v2 specifically it recorded 1753 Elo, on a scale where the human expert baseline is set at 1000.
The Coding Agent Index rates an agent-tool-plus-model combination rather than a model alone. Version 1.3 averages pass@1 across DeepSWE (113 tasks), Terminal-Bench v2 (84), and SWE-Atlas-QnA (124).
Three points conceded for one-third the cost and 30% less wall-clock time. On raw coding benchmarks Grok 4.5 trailed the Claude line — Terminal-Bench 2.1 at 83.3%, SWE-Bench Multilingual at 78.0%. Efficiency was xAI's chosen axis from the start.
Artificial Analysis computes the real cost of running its full index by combining API-reported token counts with measured cache hit rates.
Grok 4.6 came in at $0.84 per task, $1,068.47 to run the whole index. GPT-5.6 Sol, tied at 61 points, costs $1.23 — about 1.5x. Index leader Claude Opus 5 costs $2.34, or 2.8x. Claude Fable 5 costs $3.14, or 3.7x.
The instructive comparison is Kimi K3: identical $0.84 per task, but a far larger total because it burns more tokens. Grok 4.6 ranks 58th of 182 on cost.
The time metric is weighted average decoding time per index task, excluding TTFT and overhead. Grok 4.6 (high) takes 6.3 minutes, matching GLM-5.3 — slower than GPT-5.6 Sol (max) at 4.2 minutes, faster than Claude Opus 5 (max) at 7.8, Kimi K3 (max) at 10.2, and Qwen3.8 Max at 10.6.
Read that next to token consumption. Grok 4.6 spends 22K output tokens per task; Claude Opus 5 spends 40K and GLM-5.3 spends 41K. Raw output speed is unremarkable — 57.9 tokens per second, 95th of 182.
So Grok 4.6 is not a fast model. It is a quiet one, and the cost advantage above comes from that token thrift rather than from the per-token price.
Grok Build started in mid-May 2026 as an early beta limited to the $300/month SuperGrok Heavy tier, opened in late May to SuperGrok ($30/month) and X Premium+ ($40/month), reached v1.0.0 on Aug. 7, and switched its default model to Grok 4.6 on Aug. 12.
It is a terminal-based agent rather than an IDE, offered as TUI, headless, and ACP. The workflow runs Plan → Search → Build, executing after plan approval, with up to eight subagents (16 on Heavy) running in parallel, each in its own Git worktree.
The most strategic decision is the skill format. Official documentation states that "Grok is fully compatible with Claude Code with zero configuration needed," adopting Anthropic's skill format and driving switching cost to effectively zero.
On June 16, 2026, SpaceX announced it would acquire Anysphere, maker of Cursor, for $60 billion in an all-stock deal. It closed Aug. 14 as the largest startup exit on record. At acquisition, Cursor had roughly $4 billion in annual recurring revenue, 50,000 enterprise customers, and about two-thirds of the Fortune 500.
Four stated rationales: a data flywheel of some 150 million lines of enterprise code per day, Colossus compute, talent replenishment after founding-team departures, and enterprise distribution.
The most consequential change happened quietly. In current Cursor documentation, Grok 4.6 and 4.5 are classified alongside Composer as "first-party" models, while Claude, GPT, and Gemini are "third-party." First-party models are effectively unlimited; third-party usage draws down a dollar budget by plan. Model choice remains in the UI but has left the price list. Worth recalling that Grok 4.5 was the first model trained on Cursor developer data. The data came first, the company second.
The original author tested three website-building prompts: a museum exhibition site themed on kinetic sculpture, a deep-sea documentary site where scrolling becomes descending depth, and a typography experiment built entirely in Helvetica. All three carried specific interaction-design and art-direction requirements.
The verdict was brief. On simple prompts, output quality was hard to distinguish from Claude or GPT models, and given the price gap, using Grok became a reasonable choice.
Three things.
First, the performance gap at the top has narrowed. When six of 182 models sit within two points, model selection stops being decided by a benchmark point or two and starts being decided by cost and workflow fit.
Second, cost advantage now comes from token consumption, not price per token. At identical per-million pricing, a model spending 22K output tokens per task and one spending 40K differ by nearly 2x in real cost. Comparing rate cards alone hides this. The gap widens with call volume — content production pipelines feel it most.
Third, removing switching cost is a distribution strategy. Skill-format compatibility and Cursor's plan reclassification are moves in distribution, not performance. When evaluating AI tooling for a team, treat a "free and unlimited" first-party model as the output of that strategy rather than a neutral technical fact.
On accumulation mattering more than tool choice, see What Socar and Samjjeomsam Learned the Hard Way. On wiring a model into an actual production pipeline, see How to Build an AI Content Workflow From Scratch.
No. It runs on the same V9 foundation as Grok 4.5 with post-training added. That lifted its Artificial Analysis Intelligence Index from 56 to 61 in roughly five weeks, at unchanged pricing.
500,000-token context, $2 per million input tokens and $6 per million output, with a 75% cache discount. Knowledge cutoff is Feb. 1, 2026.
Output token consumption, not rate card. It spends 22K output tokens per task versus 40K for Claude Opus 5 and 41K for GLM-5.3, landing at $0.84 per task against Opus 5's $2.34 — 2.8x cheaper.
Cursor documentation now classifies Grok 4.6 and 4.5 with Composer as first-party models with effectively unlimited use, while Claude, GPT, and Gemini are third-party and draw down a plan-based dollar budget.
To apply what you just read to your own site, start with a free audit of where things are now.
A strategist replies within 24 hours on business days.
The internet had Google, mobile had its own winners. So where is the AI era's giant? A Korean operator argues the answer is structural: this time the screen didn't change, the price of tools did.
Opening an AI before a search engine is becoming the default. As the path back to original creators narrows, where does competitive advantage in the AI era actually come from?
Meta expanded its vibe coding app Pocket to all U.S. users. Describe what you want and it builds a small app or game — then that output circulates in a social feed.