Grok 4.6 Explained: xAI's Flagship Nears GPT-5.6 and Claude Levels
Grok 4.6 lands near GPT-5.6 and Claude Opus 5 on major benchmarks at roughly half the cost. Here's what changed and why it matters.
What is Grok 4.6 and why does it matter?
Grok 4.6 is xAI’s latest flagship language model, released as an incremental upgrade to Grok 4.5 rather than a new model trained from scratch. It matters because it’s the first Grok release that lands within striking distance of GPT-5.6 and Claude Opus 5 on several major independent benchmarks, while pricing in at roughly $2 per million input tokens and $6 per million output tokens, a fraction of what comparable frontier models from OpenAI and Anthropic charge.
TL;DR
- Grok 4.6 is a supplemental training run, not a new base model. It builds on the same roughly 1.5 trillion parameter foundation used for Grok 4.5, refined with curated engineering data, much of it sourced through xAI’s acquisition of Cursor.
- Benchmark placement is genuinely competitive. On the Artificial Analysis Intelligence Index, Grok 4.6 high ties with GPT-5.6 Soul, sitting just behind Claude Opus 5 and GPT-5.6 in the number one and two slots.
- Cost per task is where Grok 4.6 stands out. At roughly 83 cents per task on Artificial Analysis’s cost chart, it delivers intelligence scores close to models that cost multiple times more, most notably Claude Opus 5.
- Results are mixed depending on the benchmark. Grok 4.6 tops GDPval and Harvey Bench (legal use cases) but trails GPT-5.6 Soul Max on Deep Sweet, a coding benchmark many developers consider a better proxy for real-world feel.
- Grok Bot launched alongside the model, an always-on agent platform that runs in cloud virtual machines, delegates tasks to sub-agents, and targets non-technical knowledge workers rather than developers.
- xAI is shipping on an unusually fast cadence. Grok 4.5 arrived recently, 4.6 followed within weeks, and Elon Musk has said Grok 4.7 is already in training with a target release three to four weeks out.
- The Cursor acquisition is the throughline. xAI bought Cursor for its coding data and paired it with the idle GPU capacity from its Colossus data center, replicating the coding-first flywheel that Anthropic and OpenAI already used to pull ahead.
Everyone else built a construction worker.
We built the contractor.
One file at a time.
UI, API, database, deploy.
How does Grok 4.6 perform on major benchmarks?
The picture is strong but not uniformly first place. On GDPval, an OpenAI-created benchmark that tests models against real economic tasks across industries like engineering, finance, AutoCAD, hospitality, and video editing, Grok 4.6 high posted the top score, ahead of both GPT-5.6 Soul and Claude Opus 5 Max. On Harvey Bench, which measures legal work, Grok 4.6 scored 15.8%, well ahead of GPT-5.6 Soul’s 2.5% and Opus 5’s 11.3%.
On Cursor Bench 3.2, an internal coding benchmark from Cursor, Grok 4.6 came in just behind Opus 5 Max rather than topping the chart. On terminal-based agentic tasks, it jumped from around 15% to 26% compared to Grok 4.5, a large single-version gain.
Deep Sweet, a coding benchmark that several reviewers consider closer to how developers actually experience a model day to day, put Grok 4.6 in third place at 65.9%, behind GPT-5.6 Soul Max at 73% and Opus 5 at 70%. That gap matters: benchmark scores and lived coding experience don’t always match, and Deep Sweet is one of the few evals designed specifically to catch that mismatch.
On the Artificial Analysis Intelligence Index, which aggregates scores across many evals into a single composite, Grok 4.6 high ties with GPT-5.6 Soul for third place, sitting just under Opus 5 (first) and GPT-5.6 (second). That’s a jump of roughly four to five points over Grok 4.5’s prior standing in fourth place.
A separate private benchmark, the Val AI Index, placed Grok 4.6 sixth, behind Gemini 3, GPT-5.6 Soul, and Meta’s Muse Park 1.2. Private, third-party benchmarks like this carry weight precisely because the lab publishing them has no incentive to favor any single model.
What does the cost-per-task comparison actually show?
Raw intelligence scores only tell half the story. Artificial Analysis also plots cost per completed task against intelligence score, and that chart reframes the competition. Grok 4.5 high sat around 36 cents per task with an intelligence index near 55 to 56. Grok 4.6 high moved up to roughly 83 cents per task, but its intelligence score climbed to around 60, landing it in the same territory as GPT-5.6 Soul Max at a lower price. Claude Opus 5, by contrast, scores higher but costs dramatically more per task.
The reason cost per task matters more than sticker price alone comes down to token efficiency. A model priced at half the per-token rate of a competitor isn’t actually cheaper if it needs twice as many tokens to reach the same answer. Kimi K3 Max is the example worth noting here: it’s priced at half of GPT-5.6 Soul and Claude Opus 5, but because it consumes roughly twice the tokens to complete comparable tasks, its effective cost per task ends up close to identical. Grok 4.6’s pricing of $2 per million input tokens and $6 per million output tokens, combined with its token efficiency, is what pushes it toward the upper-left “cheap and smart” quadrant that buyers actually want.
A fast variant of Grok 4.6 is also available at twice the price, aimed at latency-sensitive use cases. During the model’s launch window, both Cursor and Grok Build offered double usage allowances for the new model.
Is Grok 4.6 good at real coding and agentic tasks, not just benchmarks?
Early hands-on testing suggests yes, with some caveats. In one demonstration, Grok 4.6 high was used to build a working clone of a Portal-style puzzle game, including physics for momentum conservation through portals, character reflections, 3D models, and a complete puzzle room, largely in a single extended session with minor steering along the way. That’s a meaningfully complex task for a model to handle with limited human correction.
Separately, a test run had Grok 4.6 wire up a Pokémon Red emulator with a control harness, a simulated chat overlay, and text-to-speech narration for an AI streaming setup, again built out substantially in the first hour or two after release.
On design tasks, results were more mixed. In a side-by-side test generating a stylized profile card, GPT-5.6 Soul produced the cleanest, most visually polished output. Grok 4.6’s version was functional but had layout issues, spacing problems, and a button that was cut off. Claude’s output in the same test was notably weak by comparison. This suggests Grok 4.6’s strengths lean toward code correctness and agentic task completion rather than visual design polish.
Independent user reports echo a similar pattern: the model is fast, cheap, and unusually thorough, sometimes to a fault, producing far more output tokens than comparable models and double-checking work more aggressively than Grok 4.5 did. That verification behavior was previously a weak point for Grok models.
What is Grok Bot and how does it relate to Grok 4.6?
Grok Bot is xAI’s agent platform, launched alongside Grok 4.6, aimed at a broader audience than developers. Rather than a chat interface, it functions as a “chief of staff” agent that spins up its own sub-agents to handle delegated tasks, each running in its own cloud-based virtual machine so work continues even when a user’s local machine is off. Users can teach it tasks by recording a walkthrough, which the agent can later replay and adapt.
Notably, Grok Bot hides model selection entirely and doesn’t expose code. Output comes back as documents, presentations, or plain explanations, positioning it for knowledge workers rather than engineers. It’s available on desktop for Windows, macOS, and Linux, with Android support planned. At time of testing, it wasn’t fully confirmed that Grok 4.6 was yet powering Grok Bot everywhere, though it’s expected to roll in fully soon.
What’s next for xAI after Grok 4.6?
Built like a system. Not vibe-coded.
Remy manages the project — every layer architected, not stitched together at the last second.
xAI has reorganized to ship major model updates roughly every two to three weeks. Elon Musk has stated that Grok 4.7 has completed initial training and is expected within three to four weeks of Grok 4.6’s release, with additional supplemental training incorporating SpaceX-related data. Musk has also claimed 4.7 will be the best model available at release, though that claim is unverified and benchmarks will likely have moved by then given the pace of releases from OpenAI, Anthropic, and Google. A further model, referred to as Grok 5, is reportedly targeted before the end of the year, with unconfirmed reports suggesting a jump to a substantially larger parameter count.
Frequently Asked Questions
Is Grok 4.6 better than GPT-5.6?
It depends on the task. Grok 4.6 leads on GDPval and Harvey Bench, ties GPT-5.6 Soul on the Artificial Analysis Intelligence Index, but trails GPT-5.6 Soul Max on Deep Sweet, a coding benchmark many developers treat as a better proxy for day-to-day feel. Grok 4.6 is generally the cheaper option per completed task.
Is Grok 4.6 better than Claude Opus 5?
Claude Opus 5 currently leads the Artificial Analysis Intelligence Index outright and scored higher on Deep Sweet. Grok 4.6 comes close on several benchmarks at a much lower cost per task, in one comparison completing an equivalent 3D scene generation task for around 38 cents versus roughly $2 for Opus 5.
Why did Grok improve so quickly between 4.5 and 4.6?
xAI attributes the jump to a longer supplemental training run using curated engineering and reasoning data, much of it drawn from its acquisition of Cursor, plus an improved training recipe. xAI also says Grok 4.5 was used to help generate training data for Grok 4.6, a recursive self-improvement approach also used by other labs.
How much does Grok 4.6 cost to use?
Grok 4.6 is priced at $2 per million input tokens and $6 per million output tokens. A faster variant is available at double that price. It’s accessible through the API, Cursor, Grok Build, OpenRouter, Vercel, and Cloudflare.
What is Grok Bot used for?
Grok Bot is an always-on agent platform aimed at non-developers, running tasks in cloud virtual machines and delegating work across automatically created sub-agents, with outputs delivered as documents, presentations, or plain-language answers rather than code.
