- Grok 4.6 is SpaceXAI’s new flagship model, announced on August 12, 2026. It is a post-training upgrade on the Grok 4.5 base, not a new foundation model.
- It scores 61 on the Artificial Analysis Intelligence Index: five points above Grok 4.5 from a month ago, level with GPT-5.6 Sol, two points behind Claude Opus 5.
- Pricing did not move: $2 per 1M input tokens and $6 per 1M output. Roughly half what rivals charge.
- A new reasoning tier arrived: xhigh. The model also checks and verifies its own work far more often.
- The context window is 500K tokens, but any request above 200K doubles the price of the whole call.
In the AI race, a month is a long time. In early July we were writing about Grok 4.5. Five weeks later SpaceXAI announced Grok 4.6 and put the model straight back into the frontier conversation.
The interesting part is what is missing: there is no bigger model here. The company kept the base and ran a longer supplemental training pass on top of it. The result is a five-point jump on independent benchmarks with no change to the price tag.
The model is written officially as Grok 4.6, and its API id is grok-4.6. The company name may throw you off too: xAI became SpaceXAI after merging into SpaceX, but the Grok brand stayed as it was.
Let us look at what actually changed. 👇🏻
What Is Grok 4.6?
Grok 4.6 is SpaceXAI’s flagship model, positioned for long-running agents, coding and knowledge work. It does what ChatGPT or Claude do: answers questions, writes code, reads documents, understands images.
Its distinguishing claim fits in one sentence: stay on task for hours without drifting. The scenarios the company highlights are repository-wide refactors, migration agents, research pipelines over large document corpora, and GPU kernel optimization.
Here is the spec sheet:
| Feature | Value |
|---|---|
| Model id | grok-4.6 |
| Context window | 500,000 tokens |
| Input | Text + images |
| Output | Text only (no stated limit) |
| Knowledge cutoff | February 1, 2026 |
| Reasoning tiers | low, medium, high (default), xhigh |
| Open weights | None, no self-hosting |
SpaceXAI did not disclose a parameter count. The trillion-scale figures circulating online are not official; all the company has said is that this is a post-training upgrade on the Grok 4.5 base.
How much is 500K tokens?
A token is roughly a fragment of a word. 500,000 tokens is about 375,000 words: an entire mid-sized codebase, or several hundred pages of technical documentation.
If you want to know what your own text costs, paste it into our token counter and see the number instantly.
What Changed in This Release?
SpaceXAI worked on three fronts:
A longer supplemental training run. Grok 4.6 got a longer top-up training pass than Grok 4.5 did, using curated model-generated data for reasoning and advanced technical concepts.
Reinforcement learning in agentic environments. The model was trained with rewards inside real task environments covering coding, web development, CAD and kernel optimization.
Checking its own work. This is the behavioral change the company points to: the model tests and verifies what it produced before moving to the next step. That is what matters most in long tasks, because agents usually fail not on one hard step but on a small unnoticed error at step fifteen snowballing into everything after it.
There is also a new reasoning tier: xhigh. You now pick between low, medium, high (default) and xhigh. Low is cheap and fast for simple work; xhigh thinks longer on hard problems.
How Does It Compare? 📊
Here is the comparison table SpaceXAI published, with rival figures taken from those companies’ system cards or public leaderboards:
| Evaluation | Grok 4.6 High | Grok 4.5 High | GPT-5.6 Sol Max | Fable 5 Max |
|---|---|---|---|---|
| AA Intelligence Index | 61 | 56 | 61 | 62 |
| GDPVal-AA v2 (Elo) | 1753 | 1526 | 1728 | 1741 |
| CursorBench v3.2 | 69.9% | 66.7% | 67.2% | 70.5% |
| DeepSWE v1.1 | 65.9% | 54% | 73% | 70% |
| FrontierCode v1.1 (Extended) | 61.3% | 56.6% | 60.6% | 63.6% |
| APEX-Agents | 57.5% | 47.1% | 56.7% | 59.2% |
| Terminal-Bench v3.0 | 26% | 15.7% | 34.6% | 34.1% |
| APEX-SWE | 56.4% | 53.6% | - | 58.8% |
| AA-Briefcase (Elo) | 1577 | 1313 | 1502 | 1574 |
| Harvey LAB (Vals) | 15.8% | 12.9% | 2.5% | 11.3% |
The summary is straightforward: Grok 4.6 beats its own previous version in every column, while the picture against rivals is mixed.
It leads on real-world tasks across 44 occupations (GDPVal-AA v2), long document work (AA-Briefcase) and the legal-focused Harvey LAB evaluation. It trails on pure software engineering: GPT-5.6 Sol Max is seven points ahead on DeepSWE v1.1 and more than eight points ahead on the newer, harder Terminal-Bench v3.0.
On the independent Artificial Analysis Intelligence Index, Grok 4.6’s 61 ties GPT-5.6 Sol, sits one point behind Claude Fable 5 (62) and two behind Claude Opus 5 (63). It is five points above the month-old Grok 4.5 and ahead of Kimi K3. In the same firm’s separate evaluations the model scored 50.7% on τ³-Banking (tool-using customer service scenarios), among the top two, and 88.4% on the earlier Terminal-Bench v2.1.
The real story: efficiency
The previous release built its pitch on finishing work in fewer steps rather than topping the charts. That pitch still holds in 4.6.
On long-horizon tasks Grok 4.6 averages 53 turns and 0.5B input tokens. On the same tasks Claude Opus 5 averages 103 turns and 2.0B input tokens. Cost per task lands at $0.84.
In other words, the model finishes what the frontier leader finishes in roughly half the steps and a quarter of the tokens. If you are the one paying the invoice, that is the headline.
Pricing and the 200K Trap 💸
API pricing per 1M tokens:
| Item | Under 200K request | Over 200K request |
|---|---|---|
| Input | $2.00 | $4.00 |
| Cached input | $0.50 | $1.00 |
| Output | $6.00 | $12.00 |
For comparison: Claude Opus 5 runs $5/$25 and GPT-5.6 Sol runs $5/$30. Grok 4.6 is roughly half the price of both.
One detail deserves attention. The moment a request crosses the 200K token band the rate doubles, and per SpaceXAI the higher rate applies to every token in that request. A 200K-token prompt bills at $2; a 201K-token prompt bills the whole thing at $4. A single token can double the invoice.
Practical advice if you work with long context: trim inputs to sit just under the 200K line and only cross it when you genuinely have to.
What does that mean for your workload? Enter your input and output token counts in our LLM cost calculator to compare Grok against GPT and Claude side by side.
Where Can You Use It?
As of August 12, 2026 Grok 4.6 is available through:
- Grok Build, where it is the default model.
- Cursor, on all plans.
- The SpaceXAI API under the id
grok-4.6. - OpenRouter, Vercel and Cloudflare routing.
New users get 2x included usage for the first week. A faster variant also exists at double the price, with no separate model id published.
There are no open weights, so you cannot download the model and run it on your own hardware. For that, open-weight models like Kimi K3 remain the only option.
For Developers: Five Lines to Start
The API is OpenAI-compatible. Code you wrote for ChatGPT connects to Grok with two changed lines:
from openai import OpenAI
client = OpenAI(
api_key="YOUR_KEY",
base_url="https://api.x.ai/v1", # the only difference
)
response = client.chat.completions.create(
model="grok-4.6",
reasoning_effort="high", # low | medium | high | xhigh
messages=[{"role": "user", "content": "Describe Grok 4.6 in one sentence."}],
)
print(response.choices[0].message.content)
Three things worth knowing:
reasoning_effortnow takes four values. The default ishigh. Trylowfor simple classification work andxhighfor multi-step agent runs.- Use the cache. If you resend the same system prompt or the same code file, cached input bills at a quarter of the normal rate.
- Watch the 200K line. Measure request size before sending; one oversized call doubles the bill.
What Comes Next: Grok 4.7 and Grok 5
This release reads like a waypoint. Industry sources expect a larger model, Grok 4.7, within weeks, and Grok 5 before the end of 2026. These are not official SpaceXAI statements, they come from routing providers, so treat them as reports rather than facts.
The cadence itself is clear though: SpaceXAI is shipping roughly every five weeks, holding the price flat and pushing the score up each time. Even if that does not take the crown, it puts real pressure on rivals on price-performance.
Frequently Asked Questions
Q: When was Grok 4.6 released? A: SpaceXAI announced Grok 4.6 on August 12, 2026 and made it available the same day through the API, Cursor and Grok Build.
Q: How much does Grok 4.6 cost? A: $2 per 1M input tokens, $0.50 cached input and $6 per 1M output. Requests above 200K tokens double the whole tariff.
Q: Is Grok 4.6 free? A: API use is paid. There is limited free use through the Grok app and grok.com, with a SuperGrok subscription for heavy use. New API users get 2x included usage for the first week.
Q: Is Grok 4.6 better than GPT-5.6 and Claude? A: It ties GPT-5.6 Sol on the Artificial Analysis Intelligence Index (61) and sits two points behind Claude Opus 5. It is very strong on agentic tasks and behind GPT-5.6 Sol on coding evaluations like DeepSWE. On price-performance it leads.
Q: What is the difference between Grok 4.6 and Grok 4.5? A: 4.6 is a post-training upgrade on the 4.5 base. It moved from 56 to 61 on the Intelligence Index, added the xhigh reasoning tier, and verifies its own work more often between steps. Context window and pricing are unchanged.
Q: How many parameters does Grok 4.6 have? A: SpaceXAI did not disclose a parameter count. The company described it as a post-training upgrade on the Grok 4.5 base rather than a new foundation model.
Take care… 🙂