Grok 4.6: 500K-Context Post-Training Upgrade for Agents

SpaceXAI shipped Grok 4.6 on August 12, 2026 — a post-training upgrade over Grok 4.5, not a new base model. The context window extends to 500,000 tokens, a new xhigh reasoning-effort tier is added above the existing ladder, and reinforcement learning ran specifically in agentic environments. The model is live via the xAI API as grok-4.6, is the default in Grok Build, and ships in Cursor across all plans. For teams running long-running coding or knowledge-work agents, the two immediately actionable changes are a context window large enough to cover most full-repository workloads, and a pricing cliff at 200K tokens that needs to be engineered around from day one.

What Changed in Post-Training

SpaceXAI held the base model constant and concentrated improvement in three places. First, a longer supplemental training run using curated model-generated data targeting reasoning and advanced technical concepts, plus high-quality engineering data, with an improved optimizer and training recipe. Second, Grok 4.5 was used to regenerate supervised fine-tuning trajectories across reasoning-effort levels, agent harnesses, and domains spanning STEM, software engineering, and knowledge work — with problematic traces filtered via model-based checks. Third, reinforcement learning ran in agentic environments covering knowledge work, general coding, web development, computer-aided design, and kernel optimization.

The behavioral consequence xAI reports internally — a vendor observation, not an independently audited result — is that on longer trajectories the model exhibits more self-testing and verification before advancing. Artificial Analysis data provides a partial proxy: Grok 4.6 completed AA-Briefcase workloads in roughly 53 turns and approximately 0.5B input tokens on average, compared to roughly 103 turns and approximately 2B input tokens for Claude Opus 5 Max. Fewer turns does not straightforwardly equal better work, but the efficiency gap is substantial enough to matter for cost modeling on high-volume agent pipelines.

The reasoning_effort parameter now supports four levels: low, medium, high (the API default, and the level at which xAI reports its launch benchmark numbers), and the new xhigh — appropriate for the hardest steps in a trajectory such as complex debugging and multi-turn plans, with lower levels recommended for routing, extraction, and short tool calls. The model accepts text and image input and produces text-only output, with no stated output token limit. Knowledge cutoff is February 1, 2026. SpaceXAI has not disclosed a parameter count.

Benchmarks: Wins and Losses

Grok 4.6 (High) scores 61 on the Artificial Analysis Intelligence Index, five points above Grok 4.5's 56 and tied with GPT-5.6 Sol Max. The comparison set on xAI's launch table does not include Claude Opus 5, which currently leads the index. On GDPval-AA v2, Grok 4.6 reaches 1753 Elo against Grok 4.5's 1526 and GPT-5.6 Sol Max's 1728 — though xAI's bolded wins on this row and AA-Briefcase sit inside Artificial Analysis's published confidence intervals and are better read as statistical ties. The 264-point generational jump on AA-Briefcase (1577 versus 1313) is the more reliable signal.

The 11.9-point generational gain on DeepSWE v1.1 is real progress; the 7.1-point deficit to GPT-5.6 Sol Max at 73.0% is the clearest disclosed loss for teams running repository-level coding agents. Terminal-Bench v3.0 is starker: nearly doubling Grok 4.5's 15.7% to reach 26.0% still leaves it last of the four listed models, 8.6 points behind GPT-5.6 Sol Max at 34.6%. Agents that spend significant time in a shell environment should run their own evals before switching.

Pricing, the 200K Cliff, and Access

Grok 4.6 bills at two tiers. At or below 200,000 prompt tokens: $2 / $0.50 / $6 per 1M tokens (input / cached input / output). Above that threshold every rate doubles to $4 / $1 / $12. The launch page references a faster variant priced at double those rates, but no separate model ID has been published. Grok Build and Cursor are providing 2× included usage during the first week post-launch.

Cache behavior is operationally critical: without a prompt_cache_key — or the x-grok-conv-id header on Chat Completions calls — requests scatter across servers and cache hits become unreliable, meaning the full input rate applies to every token. For high-volume pipelines sitting near the 200K boundary, the difference between $0.50 and $4 per million cached tokens is the kind of cost exposure that warrants prompt optimization instrumentation before scaling.

The model is routable through OpenRouter, Vercel, and Cloudflare in addition to the xAI API directly. There is no open-weights release and no self-hosting path, ruling out air-gapped deployments. Given the security exposure patterns documented in agentic deployments, teams integrating multi-step pipelines should audit credential handling before production promotion.

Grok 4.6's APEX-Agents score of 57.5% — a 10.4-point generational gain and the row most directly tied to the long-agent RL training — confirms that the post-training investment landed where xAI aimed it. The unresolved gap is on terminal and repository-level tasks against the current coding leaders. The model's strongest near-term fit is knowledge-work synthesis at scale and first-pass code scaffolding rather than drop-in replacement for specialized SWE agents.