TL;DR
On the evening of August 12, xAI officially released Grok 4.6. On paper it looks like a small bump from 4.5 to 4.6, but the numbers are anything but small: 61 on the Artificial Analysis Intelligence Index — tied with GPT-5.6 Sol; 95.9% on MedAgentBench, global #1; 500K context; and API pricing that stays at $2/$6 with doubled free quota in Grok Build and Cursor for the first week.
I installed the official CLI and ran it hard: one prompt produced a playable CS-style FPS in under 30 minutes; it repaired code that wouldn’t even launch; it correctly identified hyperkalemia in a complex patient case and prioritized treatment; and on the same prompt, its Airbus H145 3D model was a full tier above Grok 4.5’s. My conclusion: Grok 4.6 has moved from “frontier contender” to “first-tier participant” — and this time the focus is on finishing work, not just answering questions.
Quick Specs
| Metric | Data |
|---|---|
| Release Date | August 12, 2026 (Beijing evening) |
| Model Family | 1.5T parameters (official Model Card) |
| Context Window | Up to 500,000 tokens |
| API Pricing | $2 / 1M input · $6 / 1M output · $0.5 cached input |
| AA Intelligence Index | 61 (tied with GPT-5.6 Sol, global #3) |
| MedAgentBench | 95.9% (global #1) |
| DeepSWE v1.1 | 65.9% (high) / 67.0% (xhigh) |
| Terminal-Bench 3.0 | 26.0% |
| GDPVal-AA / AA-Briefcase | 1,753 Elo / 1,577 Elo (both first-tier) |
| Code Arena WebDev | 1,618 — ranked #7 |
| Free Access | Official CLI + 2x quota in Grok Build / Cursor week one |
Why This Launch Matters
I’ll be honest: model launches have been so dense this month that I was starting to go numb. Grok 4.6 is the one that woke me up — not because of a parameter count, but because xAI is pushing the model from “good at chatting” toward “good at doing”: long-horizon agents, coding, knowledge work and interactive project development.
First, general intelligence is now first-tier. The AA Intelligence Index v4.1.1 is a weighted composite of 9 evaluations including GDPVal-AA, Terminal-Bench, SciCode and GPQA Diamond. Grok 4.6 scores 61 — up 5 points from Grok 4.5 High’s 56 — tied with GPT-5.6 Sol (max), behind only Claude Opus 5 (max) at 63 and Claude Fable 5 at 62.
Second, medical agents are the sleeper hit. MedAgentBench isn’t a trivia test — it runs 10 categories and 300 real clinical tasks inside electronic health records, testing reasoning, tool use and multi-step execution. In the independent eval published August 17, 2026, Grok 4.6 hit 95.9%, edging out GPT-5.6 Sol (94.7%). Grok 4.5 was already a top medical agent; 4.6 pushes that lead further.
Third, the price didn’t move and the free quota doubled. Grok Build and Cursor both get 2x free usage in week one, and the API stays at $2/$6. For anyone running agents daily, that’s real goodwill.
One-line take: Grok 4.6 has graduated from “frontier contender” to “first-tier participant,” with clear strengths in long-horizon agents, coding, knowledge work and interactive project development.
If you missed the previous generation, start with our Grok 4.5 review. The usual comparison targets — Claude Opus 5, Kimi K3 and Qwen 3.8 Max — all have full hands-on reviews here too.
Benchmark Scorecard
Here’s the full head-to-head from the official Model Card, comparing Grok 4.6, Grok 4.5, GPT-5.6 Sol and Fable 5 (bold = best in group):
| Benchmark | 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 | 1,753 | 1,526 | 1,728 | 1,741 |
| 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% | 64.9% |
| APEX-Agents | 57.5% | 47.1% | 56.7% | 59.2% |
| Terminal-Bench v3.0 | 26.0% | 15.7% | 34.6% | 34.1% |
| APEX-SWE | 56.4% | 53.6% | — | 58.8% |
| AA-Briefcase | 1,577 | 1,313 | 1,502 | 1,574 |
| Harvey LAB (Vals) | 15.8% | 12.9% | 2.5% | 11.3% |
Three things stand out in this table:
- Knowledge work is the real strength: GDPVal-AA 1,753 beats Fable 5 (1,741) and GPT-5.6 Sol (1,728); AA-Briefcase 1,577 is also first. These long-cycle “turn these materials into a deliverable report” tasks are exactly what Grok 4.6 was trained for.
- Medical is on another level: Harvey LAB 15.8% crushes the field, and MedAgentBench 95.9% makes the “best medical agent” claim easy to defend.
- The weaknesses are honest ones: DeepSWE (65.9%) trails GPT-5.6 Sol (73%), and Terminal-Bench (26.0%) trails both GPT-5.6 Sol (34.6%) and Fable 5 (34.1%). Long-horizon software engineering and terminal agents aren’t the top yet.
Free Install & Setup
The most practical part of this launch: there’s an official free trial path. No subscription required — install the CLI and run it locally. Here are the one-liners (I tested the Windows route):
irm https://x.ai/cli/install.ps1 | iexcurl -fsSL https://x.ai/cli/install.sh | bash⚠️ From my testing: if you’re not in a supported region, enable a proxy in global mode during install and login or it will stall. After install, run grok, complete the browser authorization, and you’re inside the Grok Build terminal.
Test 1: One-Prompt FPS Game
First test: write me a CS-style shooting game. It started by scoping the current state and matching game-asset skills, then created a project folder on my desktop — done in under 30 minutes, in one shot.
The result is called “CS Raid Operation” (WEB OPERATION · offline demo v1.1). The mission brief sets the scene: terrorists have taken a warehouse district, waves escalate, headshots kill instantly, and each cleared wave refills ammo and health. Controls cover WASD, mouse aim, R reload, 1/2 weapon switch and Shift walk. The weapon archive lists an AK47 (36 dmg, 600rpm, 30/90 rounds) and a USP tactical pistol (42 dmg, 12/48), with $400 per headshot.
In combat the HUD shows “Wave 1 | 3 enemies left”, a 1:06 timer, $1,500 and a 270 score; the AK47 reads 10/90; kill feeds pop “You [AK47] Terrorist”; headshots float “+150 HEADSHOT”. The audio is genuinely atmospheric. I survived to wave three before getting killed and seeing “Action failed — redeploy.” This is a generated, playable game — not a mockup. It shoots, scores and kills you.
Test 2: Code Repair
Second test: fixing broken code. I dropped in a furnace brick-breaker game that auto-closed on launch. Grok 4.6’s approach was clean: phase one fixes the feel-killing issues — paddle collision center offset and items that could barely be caught — then strengthens the theme UI and removes the hidden bugs that could break the game.
After the fix, the file on my desktop just worked: score 7,710, high score 7,710, level “1 · First Flame”, an 8-combo chain and a live “Flame Pierce 5.1s” power-up, plus a full legend (copper 1 hit, iron 2, gold 2, lava 3, and power-ups for widen, multi-ball, pierce, life, slow and narrow). One pass, running normally, with level progression included.
Test 3: Medical Case Analysis
Third test — the one that best shows agentic reasoning: case analysis. A 68-year-old male on five medications, latest labs in hand, and the patient feels fine. The blood glucose and other numbers were deliberately added as confounders; the real issue is hyperkalemia, driven by classic dual RAAS blockade.
Grok 4.6 nailed the diagnosis and — crucially — didn’t get pulled off by the confounders: potassium first, stop the spironolactone, blood sugar is not the main battle right now. That’s not memorized medical trivia; it’s understanding a case, connecting findings, and prioritizing multi-step treatment. Exactly what agentic reasoning should look like.
I then dragged in four medical reports and let it analyze meds and labs together. About a minute later it concluded: this is not leukemia — it’s a myeloproliferative neoplasm, with routine-check and medication suggestions. That’s a round-the-clock analysis assistant living on your desktop.
Test 4: 3D Helicopter Comparison
Last up: visual capability. With the exact same prompt, I asked Grok 4.6 and Grok 4.5 to build an Airbus H145 helicopter in 3D. The gap is obvious at a glance:
- Grok 4.6: white fuselage with red accents, five-blade main rotor, Fenestron tail, skid landing gear — plus a full TYPE DATA panel (rotor Ø 10.8 m, length 13.54 m, MTOW 3,800 kg, 2 × Arriel 2E) and Rotors / Turntable / Reset View buttons. It’s close to Airbus’s official showcase page;
- Grok 4.5: a low-poly blocky sketch — rotor blades like sticks, a loose tail ring, closer to a rough 3D-editor draft.
A simple “make a chemistry 3D microscopic demo” also produced a genuinely satisfying result. The community consensus matches my experience: Grok 4.6 is finally legit — no longer a third-tier chatbot toy.
Agent, Coding & Knowledge Work
Beyond the visible tests, the training story matters. xAI says Grok 4.6 got longer continued training than 4.5, with model-generated data, high-quality engineering data, and data aimed at reasoning and advanced technical concepts. They also generated SFT trajectories from Grok 4.5 at different reasoning intensities, across agent harnesses and domains, filtered problematic traces with model checks, then ran Agentic RL on knowledge work, coding and engineering environments.
The result is that model competition is shifting from “who chats better” to “who actually gets the job done.” Grok 4.6’s 500K context window makes long-horizon work practical — code, logs, requirements, tool results and error traces all stay in memory, so the agent doesn’t get lost mid-task.
The developer ecosystem is live too: GitHub Copilot lets you pick Grok 4.6 in VS Code, Copilot CLI and cloud agents; Amazon Bedrock offers the 500K context window with configurable reasoning effort; OpenRouter, Vercel and Cloudflare are onboard; and the official API model id is grok-4.6.
Price & Value
| Item | Detail |
|---|---|
| API input | $2 / 1M tokens |
| API output | $6 / 1M tokens |
| Cached input | $0.5 / 1M tokens |
| Context | 500K tokens |
| Week-one offer | 2x free usage in Grok Build / Cursor |
| AA measured cost per task | $0.84 — lowest among frontend models |
On Artificial Analysis, Grok 4.6 resolves complex tasks in about 53 turns and roughly 500M input tokens, and the measured cost per task is $0.84 — the lowest in the frontend tier. Cheaper than most rivals while delivering equal-or-better results is the combination that should worry the competition.
Value verdict: same price, much more capability, doubled free quota in week one. If you’re choosing a primary model for agent work, Grok 4.6 belongs on the shortlist.
How to Get Started
1. Official CLI (free)
Run irm https://x.ai/cli/install.ps1 | iex in Windows PowerShell, or curl -fsSL https://x.ai/cli/install.sh | bash on Linux/macOS/WSL, then run grok and sign in. Use a global proxy outside supported regions.
2. Grok Build / Cursor (2x quota in week one)
Switch the model picker to Grok 4.6 in Grok Build or Cursor and start working — the easiest route for developers.
3. Call the API
The xAI API exposes grok-4.6 with up to 500K context and configurable reasoning effort. It’s also available through GitHub Copilot, Amazon Bedrock, OpenRouter, Vercel and Cloudflare.
FAQ
What is Grok 4.6?
Grok 4.6 is xAI’s flagship model released on August 12, 2026, part of a 1.5T-parameter family with up to 500K tokens of context. It shifts focus from chat toward long-horizon agents, coding, knowledge work and interactive project development, and scores 61 on the AA Intelligence Index, tied with GPT-5.6 Sol.
How much does Grok 4.6 cost?
Grok 4.6 API pricing is $2 per million input tokens and $6 per million output tokens, with cached input at $0.5 per million, on a 500K-token context window. In the first week, Grok Build and Cursor offer 2x free usage, and the official CLI is free to try.
Where can I use Grok 4.6?
Grok 4.6 is available in Grok Build, Cursor, GitHub Copilot (VS Code, Copilot CLI and cloud agents), Amazon Bedrock, OpenRouter, Vercel and Cloudflare, plus the xAI API as the grok-4.6 model with configurable reasoning effort.
How does Grok 4.6 compare with GPT-5.6 Sol?
Both score 61 on the AA Intelligence Index. Grok 4.6 leads on MedAgentBench (95.9% vs 94.7%) and GDPVal-AA (1,753 vs 1,728), but trails on DeepSWE (65.9%/67.0% vs 73%) and Terminal-Bench (26.0% vs 34.6%).
How strong is Grok 4.6 at agentic tasks?
Very strong in practice. In my tests it generated a playable CS-style FPS game from one prompt in under 30 minutes, repaired broken code in one pass, identified hyperkalemia in a complex patient case with correct prioritization, and combined four medical reports into a myeloproliferative neoplasm diagnosis. It tops MedAgentBench with 95.9% across 300 real clinical tasks.
How can I try Grok 4.6 for free?
Install the official CLI: run irm https://x.ai/cli/install.ps1 | iex in Windows PowerShell, or curl -fsSL https://x.ai/cli/install.sh | bash on Linux/macOS/WSL, then run grok and sign in. Grok Build and Cursor also give 2x free usage during the first week; a proxy may be needed outside supported regions.
Verdict
| Category | Score | Comment |
|---|---|---|
| Performance / Benchmarks | 4.6 / 5 | AAII 61 ties GPT-5.6 Sol; MedAgentBench global #1 |
| Agent / Long-horizon | 4.8 / 5 | Agentic RL + 500K context; games, code and medical tests all land |
| Coding | 4.5 / 5 | DeepSWE 65.9/67.0; Code Arena WebDev #7 (1,618) |
| Price / Value | 4.8 / 5 | $2/$6 unchanged + 2x quota; AA cost per task $0.84, lowest |
| Ecosystem / Usability | 4.4 / 5 | Copilot / Bedrock / OpenRouter all live; free official CLI |
One-line verdict: Grok 4.6 is the step where xAI stops being a chatbot company and starts shipping a real AI agent — general intelligence tied with GPT-5.6 Sol, the world’s best medical agent score, and hands-on results that include a playable FPS, fixed code and a correct clinical diagnosis, all at $2/$6. Keep expectations in check on DeepSWE and Terminal-Bench, but the model has officially earned its seat at the first-tier table.
If you’re choosing a primary agent model, my advice is direct: run the free CLI or the Cursor quota for a week, stress it with long tasks and interactive builds, then decide on pay-as-you-go.
Sources
Facts and figures in this review were cross-checked against official public materials and independent evaluations:
- xAI official docs — models & API pricing (Grok 4.6)
- xAI official CLI — free trial channel
- Artificial Analysis — Grok 4.6 intelligence index & cost data
- Medical Sphere — MedAgentBench independent eval (Aug 17, 2026)
- GitHub Copilot — Grok 4.6 in the model picker
- Amazon Bedrock — Grok 4.6 with 500K context