Grok 4.6

--- title: “Grok 4.6” description: “Grok 4.6 (xAI, August 12, 2026): what the release is, why it matters for operators, specs, benchmarks, and the call.” type: guide category: ai pubDate: 2026-08-12 image: /images/abs-model-grok-4.6.png imageAlt: “A retro robot representing Grok 4.6” imagePrompt: “Bold graphic editorial illustration, 1990s comic-book influence, heavy ink outlines, halftone texture, crimson and electric blue on cream. A single retro-futurist robot representing an AI model / a brain-in-a-server, no text, no logos, 16:9.” affiliate: false sources: facts: - label: “Vendor” value: “xAI” - label: “Released” value: “August 12, 2026” - label: “License” value: “Proprietary API” - label: “Type” value: “AI model release” related: - abs-model-claude-opus-5 - abs-model-gpt-5.6-cyber - abs-model-kimi-k3 - abs-model-lfm2.5-vl-3b - abs-model-ling-3.0-flash - abs-model-muse-glimmer-30b - abs-model-muse-spark-1.2 - abs-model-nemotron-3.5-lightning - abs-model-qwen-image-3.0 - abs-model-qwen3.7-flash - abs-model-qwen3.8-2.4t-a95b - abs-model-qwen3.8-max-weights-update - abs-model-solar-pro-4 tags: - grok-4.6 - ai-models - model-release draft: false --- --- title: “Grok 4.6” description: “Grok 4.6 (xAI, August 12, 2026): what the release is, why it matters for operators, specs, benchmarks, and the call.” type: guide category: ai pubDate: 2026-08-12 image: /images/abs-model-grok-4.6.png imageAlt: “A retro robot representing Grok 4.6” imagePrompt: “Bold graphic editorial illustration, 1990s comic-book influence, heavy ink outlines, halftone texture, crimson and electric blue on cream. A single retro-futurist robot representing an AI model / a brain-in-a-server, no text, no logos, 16:9.” affiliate: false sources: facts: - label: “Vendor” value: “xAI” - label: “Released” value: “August 12, 2026” - label: “License” value: “Proprietary API” - label: “Type” value: “AI model release” related: - abs-model-claude-opus-5 - abs-model-gpt-5.6-cyber - abs-model-kimi-k3 - abs-model-lfm2.5-vl-3b - abs-model-ling-3.0-flash - abs-model-muse-glimmer-30b - abs-model-muse-spark-1.2 - abs-model-nemotron-3.5-lightning - abs-model-qwen-image-3.0 - abs-model-qwen3.7-flash - abs-model-qwen3.8-2.4t-a95b - abs-model-qwen3.8-max-weights-update - abs-model-solar-pro-4 tags: - grok-4.6 - ai-models - model-release draft: false --- --- title: Grok 4.6: SpaceXAI’s 1.5T Tuning Release That Stops Quitting Halfway type: guide category: ai --- # Grok 4.6: SpaceXAI’s 1.5T Tuning Release That Stops Quitting Halfway Audience: Agent operators wiring models into long-running sub-agents, research-synthesis pipelines, and mid-tier coding workflows. Not for: Anyone shopping for the absolute reasoning ceiling (use GPT-5.6 Sol or Fable 5) or waiting for the bigger Grok 4.7 (2.1T) in a few weeks. Last verified: 2026-08-12 Evidence weight: documentation-verified --- ## The release — and the 1.5T-vs-2T myth SpaceXAI (the company formerly known as xAI, absorbed into SpaceX in February 2026) shipped Grok 4.6 on August 12, 2026 — simultaneously to the x.ai blog, the API, Grok Build, and Cursor. > “Today we are releasing Grok 4.6. Grok 4.6 builds on Grok 4.5 with a particular focus on long-running agents and more ambitious interactive and visual work.” One correction up front, because the press got this wrong: Grok 4.6 is a 1.5T-parameter post-training refresh of the Grok 4.5 foundation, not a 2T rewrite. Musk confirmed it himself on July 27: > “Grok 4.6 releases around August 7, it is the 1.5T model with significantly improved supervised fine-tuning and reinforcement learning, and Grok 4.7 will be the 2.1T model released a few weeks later, better in every way except slightly slower to serve.” The “August 7” date was Musk’s loose estimate; the real public launch landed August 12, with the developer API going live late August 11. Both dates are right at their layer. The 2T-scale model everyone guessed at is Grok 4.7, still weeks out. --- ## Why it matters for operators Grok 4.6 is the release where SpaceXAI stops chasing a bigger brain and ships reliability instead. The strongest read on agentic coding — Terminal-Bench v3.0 — jumps from 15.7% (Grok 4.5 High) to 26%, the single biggest delta in the launch. APEX-Agents climbs 10 points too (47.1% → 57.5%). And the model now checks its own work mid-trajectory, which directly answers the “Grok quits halfway through the task” complaint that dogged 4.5. The price is the other hook. At $2 in / $0.50 cached / $6 out per million tokens below 200k prompt (doubling above), effective OpenRouter input lands at $0.77/1M after cache — destructively cheap for long-context agent loops. That’s GPT-5.6-Sol-tier agentic capability at roughly half the per-token cost. Distribution breadth is real too. Cursor co-launched day one: > “Today we are releasing Grok 4.6 together with SpaceXAI” And there’s a house angle: Hermes Agent is already the #4 app routing tokens to Grok 4.6 on launch day — 213M tokens, behind Claude Code (305M), bcode (266M), and pi (261M). --- ## Specs that matter | Spec | Value | |---|---| | API model ID | grok-4.6 | | Context window | 500,000 tokens | | Knowledge cutoff | February 1, 2026 | | Modalities | Text + image input; text-only output | | Output limit | None | | Parameters | 1.5T (per Musk, July 27; same as Grok 4.5) | | Reasoning effort | low / medium / high / xhigh (new) | | API surfaces | Responses API, Chat Completions, function calling, web search, X search, code execution | | Cache | Supported; prompt_cache_key recommended for multi-turn | | Pricing (≤200k) | $2.00 in / $0.50 cached / $6.00 out per MTok | The new xhigh reasoning effort is a real addition — 4.5 only had low/medium/high. No output limit removes a constraint that bit long-writing and long-coding use cases on prior Groks. --- ## Benchmarks Headline numbers from xAI’s Aug 12 announcement: - AA Intelligence Index: 61 — ties GPT-5.6 Sol Max, one point behind Fable 5 Max (62) - GDPVal-AA v2: 1753 — up from 1526 on 4.5 - CursorBench v3.2: 69.9% — up from 66.7% - DeepSWE v1.1: 65.9% — up from 54%, but behind GPT-5.6 Sol (73%) and Fable (70%) - Terminal-Bench v3.0: 26% — the +10pp swing from 15.7% - FrontierCode v1.1: 61.3% - AA-Omniscience non-hallucination: 65.7% vs accuracy 48.2% — it declines to answer rather than confabulate Third-party (OpenRouter pulling Artificial Analysis): Intelligence Index 60.9 (better than 96% of models), Agentic Index 58.7 (99th percentile), GPQA Diamond 94.9%, AA-LCR long-context reasoning 75.0%. The honest read: Grok 4.6 ties Sol on overall intelligence, leads on agentic percentile, but loses DeepSWE v1.1 to GPT-5.6 Sol (73%) and stays well behind on Terminal-Bench (Sol 34.6%, Fable 34.1%). “Frontier agentic coding” is generous on the hardest SWE-bench-style evals. --- ## What changed vs Grok 4.5 Per xAI’s official post, the gains come from method, not scale: > “Grok 4.6 underwent a longer supplemental training run than Grok 4.5, with curated model-generated data for reasoning and advanced technical concepts, high-quality engineering data, and an improved optimizer and training recipe.” The post-training recipe added three things: Grok 4.5 itself regenerated the SFT trajectories across reasoning efforts and agent harnesses (filtered by model-based checks); agentic RL expanded into kernel optimization, web development, and CAD; and the model picked up self-verification on long runs: > “On longer trajectories, we also started to see more self-testing and verification, with the model checking its own work before moving on.” On visual and interactive first passes, xAI says: > “Grok 4.6 produces stronger first passes on visual and interactive projects than we typically saw with Grok 4.5.” --- ## Strengths - Best $/agentic-coding-value in its tier. $2/$6 with $0.50 cache matches GPT-5.6-tier capability at roughly half the cost. AA Agentic 58.7 ranks 99th percentile. - Real reliability lift on long agents. +10pp Terminal-Bench, +10pp APEX-Agents, plus self-testing before moving on. - 500k context, no output limit. Competitive with Claude (200k) and Gemini (1M+ with caveats); AA-LCR 75.0% confirms long-context reasoning is real. - Day-1 distribution breadth. Cursor + Grok Build + OpenRouter + Vercel + Cloudflare + AWS Bedrock + Databricks — the widest simultaneous Grok rollout yet. - Hallucination behavior is genuinely better. 65.7% non-hallucination vs 48.2% accuracy — the right asymmetry for operator use. --- ## Weaknesses - Loses the hard-SWE eval. DeepSWE v1.1 65.9% vs Sol’s 73% — third-tier behind GPT-5.6 and Fable on real-PR-fixing. - Parameter scale unchanged. 1.5T base, same as 4.5. The “bigger Grok” is Grok 4.7 (2.1T) in a few weeks; if 4.7 lands hot, 4.6 ages fast. - Terminal-Bench still trails the leader. 26% vs Sol 34.6% / Fable 34.1% — the weakest measurable area for the “long-running agents” pitch. - xhigh reasoning is uncosted. No published per-token price; pin budget caps before testing. - Text-only output. Native multimodal output (image/audio/video) still routes to separate Grok Imagine / Voice APIs. --- ## The operator call Put Grok 4.6 in the rotation for long-context research synthesis. 500k context + $0.77 effective input + AA-LCR 75% makes it a strong pick for “process lots of source material and synthesize” workloads. Pair it with Claude Opus 5 and GPT-5.6 Sol for publish-grade deep research. Route mid-tier agentic coding to it on cost, but don’t brand-route. On benchmarks: GPT-5.6 Sol on DeepSWE, Fable on CursorBench, Grok 4.6 on Terminal-Bench / APEX-Agents. Stay multi-model. Use it as a hallucination-aware second opinion. For citation-sensitive work where made-up citations are the top failure mode, 4.6’s 65.7% non-hallucination rate makes it a strong check even if you don’t pick it first. Try it free before paying. On X Premium+ or SuperGrok,

Sources

• xAI — Today we are releasing Grok 4.6 — primary launch post; full benchmark table, training-methodology quotes.
• xAI release notes — confirms grok-4.6 slug, 500k context, pricing, reasoning_effort, Feb 1 2026 cutoff.
• Grok 4.6 model doc — APIs, tools (web/X search), partner surfaces (OpenRouter, Vercel, Cloudflare, Cursor).
• Cursor co-launch blog — Grok 4.6 with SpaceXAI — day-1 availability, flagship framing.
• OpenRouter — x-ai/grok-4-6 — independent Artificial Analysis benchmark, live pricing, launch-day traffic.
• Orcarouter — Grok 4.6 release date — Musk’s July 27 clarification that 4.6 = 1.5T (not 2T).