업데이트
What is Grok 4.6?
Grok 4.6 is xAI's top language model, released August 12, 2026 as a post-training upgrade over Grok 4.5, aimed at long-running agents, coding, and knowledge work. It accepts text and image input, outputs text, and runs on UnificAlly as xai/grok-4.6. The context window is 500k tokens, the knowledge cutoff is February 2026, and reasoning effort is tunable across four levels: low, medium, high (the default), and xhigh. xAI built the release in collaboration with Cursor, and it shows in the editor-agent results below.
What's new in Grok 4.6
Near the top of SWE-bench Verified
95.6% on Vals AI's independent SWE-bench Verified re-run, fourth of 82 models, behind Claude Opus 5 at 97.0% and ahead of Kimi K3 at 93.4%. Grok 4.5 scored 86.6% on the same harness, so the version bump is a nine-point jump.
First place on CursorBench 3.2
At Extra High effort it scores 70.8%, the top score on CursorBench 3.2, just ahead of Claude Fable 5 Max at 70.5% and Claude Opus 5 Max at 70.0%. One caveat: Cursor runs the benchmark and worked with xAI on this model, so read the board with that in mind.
Frontier-class general intelligence
Artificial Analysis scores it 61 on the Intelligence Index, sixth of 188 models, one point behind Claude Fable 5 Max and level with GPT 5.6 Sol Max. Grok 4.5 sits at 56.
Four reasoning efforts on one model
low, medium, high, and xhigh let you trade latency and token spend against depth per request. The same model handles quick extraction at low and multi-hour agent runs at xhigh.
What Grok 4.6 is best for
Editor coding agents
Top CursorBench 3.2 score at Extra High effort; strong at multi-file edits with tests.
Long-running agents
Trained on agentic RL environments; keeps verifying its own work across long task chains.
Repository-scale bug fixing
95.6% SWE-bench Verified on real GitHub issues, independently re-run.
Office and knowledge work
Document, spreadsheet, and research tasks with structured outputs and function calling.
Visual input tasks
Screenshots, diagrams, and charts read as input for text answers.
Hard problems at xhigh effort
When answer quality matters more than the wait, xhigh buys extra reasoning.
Grok 4.6 use cases
Use Grok 4.6 behind coding agents that work through real issues end to end: read the repo, patch, run tests, and iterate until green. It fits research and office assistants that gather sources, fill spreadsheets, and return structured JSON through function calling. CAD, engineering, and web-dev agents are a stated training focus, so tool-heavy build loops are a good match. With 500k tokens of context, it also handles long design docs and issue threads in one request, with screenshots included as image input.
Grok 4.6 limitations
Grok 4.6 is slow for interactive chat. Artificial Analysis measures 63.9 output tokens per second, below the frontier median, with time to first token near 47 seconds at high effort. It also tends to produce more output tokens than average on the same tasks.
The 500k context window is half of what Claude Fable 5 or GPT 5.6 Sol offer, which matters for the largest single-request workloads.
On xAI's own model card it trails Claude Opus 5 on Terminal-Bench 3.0 (26.0% against 43.5%) and on long-horizon SWE-Marathon runs. Output is text only, and image generation is a separate Grok Imagine family.
Grok 4.6 vs Grok 4.5
Grok 4.6 is the same family trained further, and the gap is wide. SWE-bench Verified rises from 86.6% to 95.6% on Vals AI's harness, and the Artificial Analysis Intelligence Index moves from 56 to 61, passing Grok 4.5's whole peer group. The trade is speed: Grok 4.6 thinks longer and writes more tokens per task. Keep Grok 4.5 for cheap high-volume calls; move agent and coding traffic to 4.6.
When to use Grok 4.6
Use Grok 4.6 when the task is agentic and the finish line matters: multi-step coding, long research runs, office automation, and tool-calling workflows that punish shallow answers. Skip it when you need sub-second first tokens or the cheapest possible bulk calls.
Grok 4.6 API examples
Call Grok 4.6 from any language by POSTing to /v1/chat/completions, the OpenAI-compatible endpoint shared by every language model in the UnificAlly API. Full parameter docs live at docs.unifically.com/models/llm/xai/grok-4.6.
curl -X POST https://api.unifically.com/v1/chat/completions \
-H "Content-Type: application/json" \
-H "Authorization: Bearer YOUR_API_KEY" \
-d '{
"model": "xai/grok-4.6",
"messages": [
{ "role": "user", "content": "Find the race condition in this worker pool and write a failing test for it." }
]
}'
The response comes back synchronously with the completion. Set "stream": true to receive tokens as they generate.
FAQ
Grok 4.6에 대해 자주 묻는 질문
xai/grok-4.6, called through the OpenAI-compatible POST /v1/chat/completions endpoint with your UnificAlly API key.
Text and image input, text output. The context window is 500k tokens, and function calling, structured outputs, and streaming all work through the standard chat completions request.
Yes, clearly. On Vals AI's independent SWE-bench Verified re-run it scores 95.6% against 86.6% for Grok 4.5, and its Artificial Analysis Intelligence Index rises from 56 to 61.
95.6% on SWE-bench Verified in Vals AI's independent re-run, fourth of 82 models, and 70.8% on CursorBench 3.2 at Extra High effort, the top score on that board in August 2026.
Four levels, low, medium, high, and xhigh, with high as the default. Lower efforts answer faster and cost fewer tokens; xhigh spends the most thinking on the hardest problems.
No. Artificial Analysis measures 63.9 output tokens per second, below the frontier median, and time to first token near 47 seconds at high effort. Pick it for depth on long tasks, not for latency-sensitive chat.
Grok 4.6 가이드 및 비교
이 모델을 실행하기 전에 가격, 제한 사항, 테스트된 결과물을 확인하세요.
