Grok 4.6 for Coding Agents: What Changed and How to Run It
Published August 13, 2026 · by the AQ team
Grok 4.6 is xAI's new flagship model, released August 12, 2026. It is a post-training upgrade over Grok 4.5 rather than a larger base model: a longer supplemental training run plus reinforcement learning in agentic environments, aimed squarely at long-running agents, coding, and knowledge work. It keeps Grok 4.5's headline API price ($2 per million input tokens, $6 per million output below 200K prompt tokens), carries a 500K-token context window, and adds a fourth reasoning-effort level, "xhigh", above the default "high". The biggest published gain is agentic: xAI reports APEX-Agents jumping from 47.1% to 57.5%. One naming note so nothing below confuses you: xAI merged into SpaceX and rebranded as SpaceXAI in July 2026, but the API, docs, and model IDs still live under x.ai.
What xAI says changed
All numbers in this section are xAI's own, published with the launch on August 12, 2026, and run at "high" reasoning effort:
| Benchmark | Grok 4.5 | Grok 4.6 |
|---|---|---|
| Artificial Analysis Intelligence Index | 56 | 61 |
| APEX-Agents | 47.1% | 57.5% |
| APEX-SWE | 53.6% | 56.4% |
| DeepSWE v1.1 | n/a | 65.9% |
| CursorBench v3.2 | n/a | 69.9% |
| Terminal-Bench v3.0 | n/a | 26% |
The shape of that table is the story. The ten-point APEX-Agents jump and the launch framing ("more self-testing and verification", "especially strong at turning a broad product idea into a working first version") say xAI optimized for the long-running agent loop: plan, execute, verify, keep going. The modest APEX-SWE move says raw patch-writing skill changed less. Notably, xAI published no comparison against Claude Opus 5, and no classic SWE-bench Verified number appeared in the launch material.
The independent read, one day in
Artificial Analysis, measuring independently on August 12, 2026, scored Grok 4.6 at 61 on its Intelligence Index, sixth overall among the models it tracks and roughly tied with GPT-5.6 Sol, one point under Claude Fable 5. Three of its measurements deserve as much attention as the score. Output speed averaged 65 tokens per second, slightly below the median of the models it tracks. Time to first token averaged 32 seconds in its measurement, so interactive use feels deliberate. And the model is verbose: it burned more output tokens running the index than most frontier peers, which matters because output is what you pay for.
On coding specifically, the same day's numbers put Grok 4.6 behind the frontier leaders: 65.9% on DeepSWE against GPT-5.6 Sol's 73% and Claude Fable 5's 70%, and 26% on Terminal-Bench v3 against their 34.6% and 34.1%. Early hands-on reports from Cursor's forum (where it shipped day one, capped at a 256K window) split the same way the benchmarks do: praise for long multistep runs and first-pass builds of interactive UIs, complaints about slowness relative to 4.5 and excessive tool calls on simple tasks. One day of anecdotes is one day of anecdotes; treat them as directions to test, not conclusions.
Two caveats the launch post leaves out. Artificial Analysis measured a non-hallucination rate of 65.7%, meaning a meaningful share of the wrong answers it sampled were confident fabrications rather than admissions of uncertainty; verify claims the model makes about your own codebase. And cached input pricing rose from $0.30 to $0.50 per million tokens, a 67% increase on the token class agent loops consume most, so long agentic sessions cost more than the unchanged headline price suggests. Above 200K prompt tokens, all rates double.
How to run it
Four ways, as of August 13, 2026:
The xAI API. Model ID grok-4.6, available since launch day. $2 input, $0.50 cached input, $6 output per million tokens below 200K prompt tokens; $4, $1, and $12 at or above it. Text and image input, 500K context, reasoning effort selectable from low to xhigh.
Grok Build, xAI's coding CLI. Grok 4.6 is now its default model. Grok Build is the terminal agent xAI shipped in May 2026: plan mode with approvals, parallel subagents that use git worktrees, headless mode, MCP support. The CLI itself is open source under Apache-2.0, but signing in requires a SuperGrok ($30/month) or X Premium+ ($40/month) subscription, which is the cheapest way to run 4.6 heavily without metering yourself per token.
Third-party surfaces. Cursor added it day one on all plans (256K window there), and it is routable through OpenRouter, Vercel, and Cloudflare.
Side by side with other agents in AQ. More on that below.
When to pick it, when not to
Pick Grok 4.6 when the work matches what it was tuned for: long-running agentic tasks where the model plans, executes, and self-verifies over many steps, first versions of product ideas, and large-context work that actually uses the 500K window. At $6 per million output tokens it undercuts most frontier peers per unit of intelligence, and the SuperGrok flat rate makes heavy CLI use predictable.
Reach for something else when latency dominates (32-second first tokens compound painfully in tight interactive loops), when the task is terminal-heavy (its Terminal-Bench score trails the frontier), or when you need the strongest raw patch-writer available and cost is secondary. The honest summary of the first day's evidence: a real step up for agentic endurance at a strong price, not a new state of the art for coding skill.
Test it against your incumbent, on your own repo
Benchmark deltas are somebody else's repo. The only number that matters is how a model handles your codebase, your test suite, your conventions. The cheap way to get it: give the new model and your incumbent the same real task from your backlog, in parallel isolated checkouts, and compare the sessions, not just the diffs. Our guide to reviewing an AI coding session covers what to look for beyond the patch: how the agent explored, what it verified, where it guessed.
Where AQ fits
AQ is the multiplayer coding harness where engineering teams run AI coding agents like Claude Code and Codex together: shared live terminals, a code editor, and app previews, in your own cloud. Grok runs in AQ as the real xAI CLI, next to Claude Code, Codex, Cursor Agent, Kimi, Antigravity CLI, and plain shells, each in its own workspace on an isolated git worktree. That makes model-release weeks like this one cheap: open two workspaces on the same repository, give Grok 4.6 and your current default the same task, and let the whole team watch both live sessions from the browser to judge the difference on your own code. You sign into the CLI with your own xAI subscription, and AQ never marks up your model usage. Sessions run on your team's VM and survive closed laptops, which suits exactly the long-running agent work this model was tuned for.
Frequently asked questions
What does Grok 4.6 cost through the API?
As of August 13, 2026: $2 per million input tokens, $0.50 per million cached input tokens, and $6 per million output tokens for prompts under 200K tokens. At or above 200K prompt tokens, the rates double to $4, $1, and $12. The cached-input rate is 67% higher than Grok 4.5's, which matters for long agent loops where cached context dominates.
Is Grok 4.6 better than Grok 4.5 for coding?
For long-running agentic coding, the evidence says yes: xAI reports APEX-Agents rising from 47.1% to 57.5% and APEX-SWE from 53.6% to 56.4%. For quick interactive edits the picture is mixed: early users report it running slower than 4.5 with more tool calls on simple tasks. If your workflow is short prompt-and-review cycles, test before switching your default.
How big is Grok 4.6's context window?
500K tokens through the xAI API, with text and image input. Third-party hosts may serve less: Cursor runs it with a 256K window as of launch. Pricing also steps up once a prompt reaches 200K tokens, so the biggest contexts cost double per token.
Are Grok 4.6's weights open?
No. Grok 4.6 is proprietary, like every xAI release since Grok 3. The lineage: Grok-1's weights were published under Apache-2.0 in 2024, Grok-2 came under a restrictive community license, and everything after runs only through xAI's API and products. The Grok Build CLI is open source; the model behind it is not.
How can my team try Grok 4.6 next to the agent we already use?
Run both on the same task and compare. In AQ, each agent CLI runs in its own workspace on an isolated git worktree of the same repository, so you can give Grok Build and Claude Code the identical backlog item, watch both live sessions in the browser, and review the two sessions side by side. Everyone signs in with their own subscriptions, and AQ adds no markup on model usage.