xAI 发布 Grok 4.6：性能追平 GPT-5.6 Sol，价格维持不变

SpaceXAI released Grok 4.6 on August 12, 2026. It scores 61 on the Artificial Analysis Intelligence Index, tying GPT-5.6 Sol Max and sitting just behind Claude Opus 5 and Claude Fable 5. Pricing stayed flat from Grok 4.5 at $2 per million input tokens and $6 per million output tokens, roughly a fifth of what GPT-5.6 Sol charges for output. It's live now in the xAI API, Grok Build, Cursor, and through OpenRouter, Vercel, and Cloudflare. That's the headline every outlet is running today. It's also not the number that should decide whether you switch. ## The gap Grok 4.5 closed in five weeks [...] Sources: Artificial Analysis's Grok 4.6 benchmark article and xAI's own developer documentation, cross-checked against Cursor's launch post. Cache pricing is worth flagging on its own: Grok 4.6's cache-hit rate rose to $0.50 per million tokens, up from Grok 4.5's $0.30. It's still cheap, but if your workload leans hard on prompt caching, the effective savings versus 4.5 are smaller than the he

Key specs: Context window: 131,072 tokens Modality: Text and code (vision input in select configurations) Primary use case: Multi-step code generation, debugging agents, repository-level tasks API availability: xAI API with OpenAI-compatible endpoints ✗ VIBE-CODED APP Tangled. Half-built. Brittle. ✓ AN APP, MANAGED BY REMY UIReact + Tailwind✓ APIValidated routes✓ DBPostgres + auth✓ DEPLOYProduction-ready✓ Architected. End to end. ### Built like a system. Not vibe-coded. Remy manages the project — every layer architected, not stitched together at the last second. RemyThe world's most powerful product manager agentTry Remy today [...] These figures reflect typical pricing for the capability tier these models occupy. Check each provider’s current pricing page for the latest rates, as both xAI and OpenAI adjust pricing regularly. ### Cost Per Agent Run For a typical SWE-bench-style task (reading a repository, understanding an issue, generating a patch, verifying against t

🏆 Fable 5 still leads overall, but GPT-5.6 Sol closes a lot of the gap at a lower cost. 👍 Overall, GPT-5.6 is not a full paradigm shift, but it is a strong and well-priced improvement over GPT-5.5. [...] 🚀 OpenAI has started previewing the GPT-5.6 series with three models: Sol, Terra, and Luna. 🧠 Sol is the flagship model, Terra is the balanced everyday model, and Luna is the fast and cheaper option. 📊 On KingBench 3, Sol scored 78.57%, Terra scored 62.9%, and Luna scored 44.3%. ✅ Sol and Terra performed extremely well on the hard math task, both getting perfect scores. 🛠️ All three GPT-5.6 models scored perfectly on the long-horizon agentic fine-tuning task. 🎨 GPT-5.6 still struggles more with frontend and visual tasks compared to Fable 5 and Opus 4.8. 💸 The pricing is very competitive, with Sol at $5 input and $30 output per million tokens. ⚡ Luna could be one of the best budget agentic models based on the benchmark results. [...] I'm making this video. It comes with a 500K t

computer use, and what they're calling ultra mode, where multiple agents run in parallel on the same task. In plain English, Grok is positioning itself as the builder's performance machine. Soul is positioning itself as the professional operating system for messy, multi-step work. Decks, research memos, spreadsheets, the stuff that doesn't fit neatly into a terminal window. Neither positioning is wrong, but broader is also how you justify a 5x jump in output pricing. So, keep that in your back pocket for the recommendation section. The benchmarks and the ones that are conspicuously missing. Now, the numbers. On XAI's own launch benchmarks, Grok 4.5 scores 62% on Deep Swe 1.0, 53% on Deep Swe 1.1, 29% on Swe marathon, 83.3% on Terminal Bench 2.1, and 64.7% on Swe Bench Pro. That's a [...] you frame it that way, the differences stop looking random and start looking like two deliberate product philosophies. The specs nobody's arguing about. Let's start with what's actually documented beca

Back to news Aug 12, 2026 # Introducing 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. Try for free Start building 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. It stays with complex tasks across many steps, whether researching a topic, analyzing information, working across a codebase, or turning an idea into a polished application or work artifact. 0:00 / 0:00 Grok 4.6 achieves frontier intelligence across several agentic coding and knowledge work benchmarks. It matches GPT-5.6 Sol on the Artificial Analysis Intelligence Index, which is a composite score of nine benchmarks. [...] Grok 4.6 produces stronger first passes on visual and interactive projects than we typically saw with Grok 4.5. Given a concrete product idea, it is able to establish structure and visual language for an applic

Build首周还提供2倍included usage。 SpaceXAI称4.6做了比4.5更长的补充训练，加入更多推理、工程和技术数据，改进optimizer和训练recipe，并针对知识工作、通用Coding、kernel优化、Web开发、CAD等Agent环境继续做RL。 所以现在的模型（包括Grok4.6和DeepSeek-V4-Pro-0813）都开始出现一个非常适合Agent时代的组合： Frontier级能力 + 更低的执行成本 + 长程Agent能力。 4.5到4.6只涨了0.1。 但这0.1，感觉把模型战争往前推了一大截。 [...] 刚刚，SpaceXAI正式发布Grok 4.6🔥 只从4.5升级到4.6，但这次Grok基本已经杀进GPT-5.6和Fable 5所在的第一梯队。 更关键的是：价格不涨。 SpaceXAI官方称，Grok 4.6主要针对长程Agent、Coding和知识工作进行了强化，并强调它相比Grok 4.5有显著提升，同时保持相同价格。 Benchmark上的提升确实很明显： AA Intelligence Index：56 → 61，追平GPT-5.6 Sol Max GDPVal-AA v2：1526 → 1753，全场第一 CursorBench 3.2：66.7% → 69.9% DeepSWE：54% → 65.9% APEX-Agents：47.1% → 57.5% AA-Briefcase：1313 → 1577，全场第一 Harvey LAB：12.9% → 15.8%，全场第一。 但Grok 4.6最值得看的其实是性能/成本比。 在另一组CursorBench 3.2成本曲线里，Grok 4.6 xhigh跑到70.8%，平均单任务成本约$2.81。 Fable 5达到接近70%的成绩时，单任务成本已经来到约$17。 也就是说，Grok开始同时占住“第一梯队能力”和“便宜”这两个位置。 官方API价格依然是每百万Token $2输入、$6输出，基础价格与Grok 4.5一致。 同时上下文扩大到500K，新增xhigh reasoning档位。模型已经上线Cursor、Grok Build、API、OpenRouter、Vercel和Cloudflare，Cursor和Grok Build首周还提供2倍inc