Grok 4.6 Release, Benchmarks And Agent Loops

xAI released Grok 4.6 on August 12, 2026. It is a direct upgrade to Grok 4.5 at the same 500K context class, aimed at long-running agents, coding, and interactive or visual first passes. The list price starts at $2 / $6 per million tokens under 200k prompt tokens.

The headline vendor numbers (Grok 4.6 High): 1753 on GDPVal-AA v2, 69.9% on CursorBench v3.2, 65.9% on DeepSWE v1.1. Around them sit 57.5% on APEX-Agents and 26% on Terminal-Bench v3.0. The benches move. The more interesting changes are how you run it and how you pay for it.

At a Glance

• Release date: August 12, 2026. Generally available.
• Model ID: grok-4.6 on the xAI API (Responses + Chat Completions).
• Pricing: $2 / $0.50 cached / $6 per 1M under 200k prompt tokens. $4 / $1 / $12 at or above 200k. The higher tier bills the whole request.
• Priority Processing: 2× standard rates when service_tier is priority. Not a separate model ID.
• Context: 500,000 tokens. Knowledge cutoff February 1, 2026. No published text output cap.
• Modalities: Text + image in, text out.
• Reasoning: reasoning_effort low / medium / high (default) / xhigh.
• Tools: Function calling, web search, X search, code execution (server-side tools billed separately).
• Where: xAI API, Cursor, Grok Build (2× included usage for the first week per the launch post), OpenRouter, Vercel, Cloudflare.
• Rate limits (docs): 150 rps, 50M tokens/min. Regions: us-east-1, us-west-2.

What's New in 4.6

None of these are a new context class. Together they push Grok toward longer, less-attended agent work.

Training

Per xAI: a longer supplemental run than 4.5, curated model-generated reasoning data, then Grok 4.5 regenerated SFT trajectories across efforts and harnesses. Agentic RL across knowledge work, general coding, and domain environments (kernel opt, web, CAD).

Product-shaped first passes

xAI's pitch is turning a broad product idea into a working first version (research, structure, implement, refine), with more self-testing on long trajectories, and stronger visual/interactive first passes than 4.5.

Agent-loop plumbing

Use prompt_cache_key (Responses) or x-grok-conv-id (Chat Completions) so multi-turn requests land on the same server. Without it you often pay full input on a cache-cold box. Long loops also get context compaction. Cache keys are the difference between $0.50 and $2.00 on every agent turn under 200k.

Benchmarks

All scores are self-reported by xAI in the launch post (Grok 4.6 High vs Grok 4.5 High). Not LLM Stats verified.

DeepSWE (+11.9), APEX-Agents (+10.4), and Terminal-Bench v3.0 (+10.3) are the movement. CursorBench is a smaller step. Terminal-Bench is still low in absolute terms (26%). On the same vendor table, GPT-5.6 Sol Max and Fable 5 Max still lead several coding suites (DeepSWE 73% / 70%, Terminal-Bench 34.6% / 34.1%). Do not turn 4.6 into a sweep.

Knowledge-work Elo

Harvey LAB (Vals) is a smaller knowledge-work side signal on the same vendor table: 15.8% vs 12.9% (+2.9). Still self-reported, still not a sweep.

Pricing & Availability

The 200k cliff is the cost story people miss: once the prompt crosses 200k, every token in that request bills at the higher rate. Priority stacks on top as a 2× lane for latency. The first-week 2× included usage in Cursor and Grok Build is a promo, not a list-price change. See docs.x.ai/developers/pricing for live rates.

Set cache keys before you scale agent traffic. Server-side tool calls (web search, X search, code execution) add per-invocation fees on top of tokens and are billed separately. Model docs and the grok-4.6 model page cover the API surface; the Grok 4.6 developer guide covers effort levels and agent loop details.

Migrating from 4.5

Grok 4.6 is a drop-in model id swap to grok-4.6. Three checks before you flip agent traffic at scale.

1. Cache keys first

Set prompt_cache_key (Responses) or x-grok-conv-id (Chat Completions) before you scale agent traffic. Without a sticky key, multi-turn loops often miss cache and pay full input on every turn.

2. Re-measure token-per-task

Re-measure token-per-task at default high (and xhigh). Output can rise on longer trajectories. Do not assume 4.5 token budgets still hold.

3. Watch the 200k cliff and Priority stacking

Once a prompt crosses 200k, the whole request bills at the long-context rate. Priority is another 2× on all token types. There is no grok-4.6-fast to flip for latency. You buy Priority on the same model ID.

Outlook

4.6 is an incremental model release wrapped around a clearer agent-loop product. The benches vs 4.5 High move most where trajectories are long. The platform surface is cache keys, compaction, and a 2× priority lane. Same 500K window. Same $2/$6 opening price. You are paying for the model to stay in the job longer, and for the option to buy latency.

For the announcement and live rates, see xAI's launch post and pricing docs.