--- title: Grok 4.6 launches with a focus on agents and app-building description: xAI has released Grok 4.6, an update focused on long-running agents and app-building rather than raw benchmark gains. The model is better at turning rough... canonical: https://daily.dev/posts/grok-4-6-launches-with-a-focus-on-agents-and-app-building-sloyifykq twitter:card: summary_large_image twitter:site: @dailydotdev og:type: website og:site_name: daily.dev og:title: Grok 4.6 launches with a focus on agents and app-building | daily.dev og:description: xAI has released Grok 4.6, an update focused on long-running agents and app-building rather than raw benchmark gains. The model is better at turning rough... og:url: https://daily.dev/posts/grok-4-6-launches-with-a-focus-on-agents-and-app-building-sloyifykq og:image: https://api.daily.dev/og/posts/SlOYIFyKq.png og:image:alt: Grok 4.6 launches with a focus on agents and app-building og:image:width: 1200 og:image:height: 630 og:locale: en --- > ## Documentation Index > Fetch the complete documentation index at: https://daily.dev/llms.txt > Use this file to discover all available pages before exploring further. # Grok 4.6 launches with a focus on agents and app-building **[Collections](https://daily.dev/sources/collections)** · 5 min read · 4 upvotes · 0 comments ## Summary xAI has released Grok 4.6, an update focused on long-running agents and app-building rather than raw benchmark gains. The model is better at turning rough ideas into working applications with self-testing and stronger first-pass visual design. It scores roughly even with GPT-5.6 Sol on the Artificial Analysis Intelligence Index but pulls ahead specifically on multi-step project work. Grok 4.6 is available now in Grok Build, the xAI API, Cursor, and partner platforms like OpenRouter, Vercel, and Cloudflare, priced at $2 per million input tokens and $6 per million output tokens, with a faster, pricier variant also offered. Cursor and Grok Build are giving 2x included usage for the first week. ## Content xAI (or SpaceXAI, depending on which branding you catch it under) shipped Grok 4.6 this week, and the rollout has been fast and loud — new integrations landing almost hourly across Cursor, Amp, Devin, GitHub Copilot, Perplexity, Warp, Augment, Vercel's AI Gateway, Amazon Bedrock, and v0. If you touch a coding tool, there's a decent chance Grok 4.6 showed up in it within days of launch. ## What actually changed Grok 4.6 is a post-training upgrade, not a bigger model. It's still built on the same 1.5 trillion parameter base as Grok 4.5. What's different is how it was trained: xAI ran a longer training pass, regenerated its supervised fine-tuning trajectories, added model-based trace filtering, and leaned harder into agentic reinforcement learning across coding, web development, CAD, and kernel optimization. The model also does more self-checking during long tasks — it pauses to verify its own work before continuing, which targets the classic failure mode of agents drifting off course over many steps. In plainer terms: instead of making the model bigger, xAI taught the existing one to think and work better. The result is a model that's noticeably stronger at reasoning, following complicated multi-step instructions, coding, and handling long-running agentic work — researching a topic, navigating an entire codebase, or turning a rough idea into a working app. ## Where it lands on the benchmarks Grok 4.6 scores 61 on the Artificial Analysis Intelligence Index, tying GPT-5.6 Sol and sitting just behind Claude Opus 5 (63) and Claude Fable 5 (62). That's a 5-point jump over Grok 4.5 and a 23-point jump over Grok 4.0 — a fast climb for a model that, as one observer put it, wasn't anywhere near the frontier six months ago. The standout numbers are in agentic work specifically: - CursorBench v3.2: 69.9%, ahead of GPT-5.6 Sol's 67.2% (up from 66.7% on Grok 4.5) - GDPVal-AA v2: 1753, ahead of both Sol Max and Fable 5 Max - AA-Briefcase: 1577, also ahead of Sol Max and Fable 5 Max - DeepSWE: jumped from 54% to 65.9% over Grok 4.5, though it still trails Sol and Fable here - Terminal-Bench: rose from 15.7% to 26% On the AA-Briefcase long-horizon benchmark, Grok 4.6 finished work in about 53 turns and 0.5 billion input tokens, versus Claude Opus 5 Max's 103 turns and 2 billion tokens. That efficiency gap shows up in real costs too — in one head-to-head running the same three builds, Grok 4.6 beat GPT-5.6 Sol on agentic loop cost, spending $13.11 versus $20.18, using bigger coding steps (201 model calls versus 338). Worth noting: 93-98% of input tokens in that test were cache reads, and the total spend would've been roughly 4x higher without prompt caching — a reminder of how much caching now shapes agent economics. ## The price story is the real headline Pricing held flat at $2 per million input tokens and $6 per million output tokens — the same as Grok 4.5, with a faster xhigh variant at double that. Cache-hit pricing did rise slightly, from $0.3 to $0.5 per million tokens, but the headline number didn't move. Compare that to Claude Fable 5 Max at $10/$50 per million tokens, and the gap is stark: Grok 4.6 gets 98.4% of Fable 5's Intelligence Index score — a 1-point, 1.6% relative difference — while costing 80% less on input and 88% less on output. Artificial Analysis puts it at $0.84 per task, squarely on the intelligence-vs-cost Pareto frontier. On Perplexity's WANDR benchmark, it matches Fable 5's results at over 60% lower cost. It's also reportedly a bit better and more efficient than Kimi K3, a model almost twice its size. This matters beyond bragging rights. Three frontier models launched within roughly 24 hours of each other — Grok 4.6, Qwen3.8-Max, and DeepSeek V4-Pro — and all three are competing mainly on price. Because open-weight models like Qwen and DeepSeek can be downloaded and self-hosted, they effectively cap what closed labs can charge. That's likely why Grok 4.6's entire improvement came from post-training rather than a bigger model: xAI matched Fable 5's scores at roughly 85% lower cost without touching parameter count. It also lines up with SpaceX's pending $60 billion acquisition of Cursor — pushing 2x usage limits at launch inside Cursor isn't subtle about who they're trying to pull users from. ## Where it actually falls short Not everyone's sold. One reviewer found Grok 4.6 got noticeably more expensive and slower to run than 4.5 in practice, because token efficiency dropped over 30% — which pushes it off that ideal cost-intelligence sweet spot in real usage, even if the sticker price didn't change. Real-world testing in that same review found weak UI and design output and a failed attempt at porting a 3D game, though a security audit and a PR-generation task went fine. That reviewer said they'd keep using other models day-to-day, but is curious about Grok 4.7.// A separate blind test had Grok 4.6, Claude Code, and Codex (GPT-5.6 Sora) each build the same full-stack jobs board app from an identical prompt, with identical tools and time limits. All three finished in under 21 minutes. Codex won on design; functionality was roughly a wash across the board. But the tester said they'd personally keep building with Grok over Claude Code, citing small UX details — a reset button, better placement of featured posts. Codex could generate PNG images with a built-in search feature; Claude Code only produced SVGs; Grok generated JPEGs. Cost comparison was inconclusive, though Grok looked cheaper than Claude Code. Another critique argues that despite the strong benchmarks, Grok 4.6 doesn't quite match Claude Opus/Sonnet or GPT-5.6 in real-world quality — its real advantage is cost per task, landing around Kimi K2.5 pricing. That piece also raises an interesting side note: Claude Opus 5, Anthropic's flagship, reportedly accounts for only 11% of Anthropic's business spend, possibly because of its high price and mandatory data retention. If true, it's a sign that the whole ## Questions this post answers ### What changed in Grok 4.6 compared to Grok 4.5? Grok 4.6 is a post-training upgrade, not a larger model, staying on the same 1.5 trillion parameter base as Grok 4.5. xAI ran a longer training pass, regenerated supervised fine-tuning trajectories, added model-based trace filtering, and increased agentic reinforcement learning across coding, web development, CAD, and kernel optimization, with the model now pausing to self-check during long tasks. _daily.dev helps developers weighing model upgrades track what actually changed under the hood before switching._ ### How does Grok 4.6 pricing compare to Claude Opus 5 and Claude Fable 5? Grok 4.6 pricing stayed flat at $2 per million input tokens and $6 per million output tokens, the same as Grok 4.5, versus Claude Fable 5 Max at $10/$50 per million tokens. Grok 4.6 reaches 98.4% of Fable 5's Intelligence Index score while costing 80% less on input and 88% less on output, putting it on the intelligence-cost Pareto frontier at roughly $0.84 per task. _developers comparing model costs for agentic workloads can track pricing shifts like this via daily.dev._ ### Is Grok 4.6 actually cheaper to run in practice for coding tasks? Not necessarily; one reviewer found Grok 4.6 got noticeably more expensive and slower than Grok 4.5 in real usage because token efficiency dropped over 30%, pushing it off the ideal cost-intelligence sweet spot despite the unchanged sticker price. That same review found weak UI and design output and a failed attempt at porting a 3D game, though a security audit and PR-generation task succeeded. _anyone choosing between Grok, Claude, and Codex for real projects can follow these hands-on tradeoffs through daily.dev._ ## Community take How the wider developer community reacted, aggregated from 3 discussions and 378 comments across hackernews, x (as of 2026-09-13). **TL;DR:** Reactions are sparse and casual, mostly noting a promotional usage discount and one person mentioning daily use, with little substantive critique or praise beyond that. Introducing Grok 4.6: Discussion mostly veers away from the model itself into a broader debate about why every frontier lab seems to leapfrog to similar capability levels around the same time, alongside heavy skepticism about benchmarks and a large amount of controversy over Musk personally and Grok's reputation. SpaceXAI's Grok 4.6 Scores 61 on the Artificial Analysis Intelligence Index: Some developers report Grok 4.6 is fast, cheap, and good enough for daily coding work (especially via Cursor's generous usage limits), while others insist it still trails Claude Opus/Fable and GPT in raw capability; a large chunk of the discussion is dominated by objections to using the product at all due to its parent company/owner. **Sentiment:** 19% positive · 28% mixed · 53% skeptical **The case for** - A promotional period offering 2x usage was highlighted as a reason for good value. - Some say it's genuinely fast and cheap for coding relative to rival models. - Noticeably faster than Claude/Codex, which matters for iterative engineering workflows. - One commenter said they've been using it as their daily driver and enjoy it. - Its release is seen by some as healthy competitive pressure on other labs. **The pushback** - Many suspect heavy benchmark gaming/'benchmaxxing' rather than genuine capability gains. - Several say real-world coding performance lags behind the benchmarks suggest, still trailing Opus 5, Fable 5, and GPT-5.6 Sol. - Strong distrust of the company's leadership colors views of the product, with references to unethical behavior and misuse concerns like CSAM generation. - Cache-hit pricing nearly doubling is seen as a meaningful cost increase for heavy coding sessions where cached tokens dominate the bill. - Some accuse discussion of the model being astroturfed, while others counter that anti-Musk sentiment is just as inflated. **By community** - hackernews (heated): The thread quickly detours from the release itself into a contentious debate about industry-wide benchmark timing and an even more heated argument about Musk's politics and Grok's ethics/trust issues. - x (positive): Light, informal chatter focused on a usage promotion and casual approval, with no real criticism raised. **Hottest debate:** Introducing Grok 4.6: Whether the near-simultaneous capability jumps across labs (including this release) reflect real technical convergence, leaked techniques, or coordinated benchmark gaming. **Open questions** - Why do multiple AI labs seem to hit similar capability milestones within a very short window of each other? - How will Grok 4.6's real-world coding performance compare to Opus 5 and Fable 5 once more people get hands-on access? - Are published benchmark results a reliable signal of real-world coding/agentic capability, or increasingly gamed? - Why is Grok priced so much cheaper than Claude or GPT despite comparable benchmark scores? - Is there a genuine compute/hardware moat, or is model-building becoming a commodity skill available to any well-funded lab? **Highlights** > @trikcode That's because there's an ongoing promotional period! 2x usage on Grok Build/Cursor until Aug 19. Get after it! > — [quentinkasseh on x · 3 points](https://x.com/quentinkasseh/status/2089728534994420085) > @trikcode Don’t telll everyone, now Elon knows and he is going to fix it > — [zunguzungu\_III on x · 1 points, 1 comments](https://x.com/zunguzungu_III/status/2089745157537075477) > @trikcode I’ve been using grok as my daily a lot recently, love it > — [bazfurby on x · 1 points, 1 comments](https://x.com/bazfurby/status/2089751437567828061) > > benchmark hacking I think this is the main one. The benchmarks from this are heavily cherry-picked, and they also widely publicised their performance for 4.5 while downplaying the fact the benchmarks were "accidentally" in their training set > — [chippiewill on hackernews](https://news.ycombinator.com/item?id=49277969) > A bunch of SWEs at my work use it as their primary model. We have Claude, ChatGPT, and Cursor with essentially no cap on spend (top guy is spending over 10K a month on AI at API prices), and he hasn't had his hand slapped. So it's not like they are using it purely because it's cheaper. I think people like to use it for its speaking style, pretty solid performance, and its speed. > — [busch\_j on hackernews · 1 comments](https://news.ycombinator.com/item?id=49275768) **Source threads** - [hackernews](https://news.ycombinator.com/item?id=49274027) · 246 points · 286 comments - [hackernews](https://news.ycombinator.com/item?id=49275385) · 77 points · 80 comments - [x](https://x.com/trikcode/status/2089714507467640939) · 0 points · 12 comments --- Tags: [#llm](https://daily.dev/tags/llm), [#ai-agents](https://daily.dev/tags/ai-agents), [#grok](https://daily.dev/tags/grok) [View this post on daily.dev](https://daily.dev/posts/grok-4-6-launches-with-a-focus-on-agents-and-app-building-sloyifykq) ```json {"@context":"https://schema.org","@graph":[{"@type":"Organization","@id":"https://daily.dev/#organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180},"sameAs":["https://twitter.com/dailydotdev","https://github.com/dailydotdev","https://www.linkedin.com/company/daily-dev-ltd"]},{"@type":"WebSite","@id":"https://daily.dev/#website","url":"https://daily.dev","name":"daily.dev","publisher":{"@id":"https://daily.dev/#organization"},"potentialAction":{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https://daily.dev/search?q={search_term_string}"},"query-input":"required name=search_term_string"}}]} {"@context":"https://schema.org","@type":"TechArticle","headline":"Grok 4.6 launches with a focus on agents and app-building","url":"https://daily.dev/posts/grok-4-6-launches-with-a-focus-on-agents-and-app-building-sloyifykq","mainEntityOfPage":{"@type":"WebPage","@id":"https://daily.dev/posts/grok-4-6-launches-with-a-focus-on-agents-and-app-building-sloyifykq"},"datePublished":"2026-08-12T15:39:17.258Z","dateModified":"2026-09-13T20:01:04.062Z","description":"xAI has released Grok 4.6, an update focused on long-running agents and app-building rather than raw benchmark gains. The model is better at turning rough...","isAccessibleForFree":true,"articleSection":"Collections","inLanguage":"en","publisher":{"@type":"Organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180}},"author":{"@type":"Organization","name":"Collections","logo":"https://media.daily.dev/image/upload/s--fk_6ycEi--/f_auto,q_auto/v1780996001/logos/collections?_a=BAMAMiWQ0","url":"https://daily.dev/sources/collections"},"commentCount":0,"discussionUrl":"https://daily.dev/posts/grok-4-6-launches-with-a-focus-on-agents-and-app-building-sloyifykq","interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":4},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":0}],"keywords":"llm,ai-agents,grok","timeRequired":"PT5M"} {"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev"},{"@type":"ListItem","position":2,"name":"Collections","item":"https://daily.dev/sources/collections"},{"@type":"ListItem","position":3,"name":"Grok 4.6 launches with a focus on agents and app-building"}]} {"@context":"https://schema.org","@type":"FAQPage","@id":"https://daily.dev/posts/grok-4-6-launches-with-a-focus-on-agents-and-app-building-sloyifykq#faq","mainEntity":[{"@type":"Question","name":"What changed in Grok 4.6 compared to Grok 4.5?","acceptedAnswer":{"@type":"Answer","text":"Grok 4.6 is a post-training upgrade, not a larger model, staying on the same 1.5 trillion parameter base as Grok 4.5. xAI ran a longer training pass, regenerated supervised fine-tuning trajectories, added model-based trace filtering, and increased agentic reinforcement learning across coding, web development, CAD, and kernel optimization, with the model now pausing to self-check during long tasks. daily.dev helps developers weighing model upgrades track what actually changed under the hood before switching."}},{"@type":"Question","name":"How does Grok 4.6 pricing compare to Claude Opus 5 and Claude Fable 5?","acceptedAnswer":{"@type":"Answer","text":"Grok 4.6 pricing stayed flat at $2 per million input tokens and $6 per million output tokens, the same as Grok 4.5, versus Claude Fable 5 Max at $10/$50 per million tokens. Grok 4.6 reaches 98.4% of Fable 5's Intelligence Index score while costing 80% less on input and 88% less on output, putting it on the intelligence-cost Pareto frontier at roughly $0.84 per task. developers comparing model costs for agentic workloads can track pricing shifts like this via daily.dev."}},{"@type":"Question","name":"Is Grok 4.6 actually cheaper to run in practice for coding tasks?","acceptedAnswer":{"@type":"Answer","text":"Not necessarily; one reviewer found Grok 4.6 got noticeably more expensive and slower than Grok 4.5 in real usage because token efficiency dropped over 30%, pushing it off the ideal cost-intelligence sweet spot despite the unchanged sticker price. That same review found weak UI and design output and a failed attempt at porting a 3D game, though a security audit and PR-generation task succeeded. anyone choosing between Grok, Claude, and Codex for real projects can follow these hands-on tradeoffs through daily.dev."}}]} ```