What Is Grok 4.6? xAI's 1.5T-Param Model Explained
Sofia Marenco
Model Evaluation Lead
TLDRGrok 4.6 launched August 7, 2026 as xAI's 1.5T-parameter frontier model with improved SFT and RL. Confirmed specs, access, and how it stacks up.
Grok 4.6 Is xAI's 1.5-Trillion-Parameter Successor to Grok 4.5 — Built to Match Kimi K3 at Half the Weight
Grok 4.6 is a 1.5-trillion-parameter frontier language model from xAI, launched on August 7, 2026 as the direct successor to Grok 4.5. Rather than stepping up in scale, Grok 4.6 reuses the same 1.5T V9 foundation and delivers its gains through significantly improved supervised fine-tuning (SFT) and reinforcement learning (RL). It is positioned to challenge Moonshot's ~2.8T-parameter Kimi K3 and Claude Opus 4.8 while preserving the speed and token efficiency of Grok 4.5. A larger 2.1T Grok 4.7 is scheduled to follow a few weeks later.
Key Takeaways
- 1.5T parameters, the same V9 foundation as Grok 4.5 — the upgrade is in post-training (SFT + RL), not raw scale. Confirmed by Elon Musk on July 28, 2026.
- Launched August 7, 2026, the date xAI publicly targeted the week before.
- Capability targets: Kimi K3 (~2.8T) and Claude Opus 4.8. Musk's bet is stronger post-training over parameter count.
- Speed and token efficiency held at Grok 4.5 levels — the same throughput envelope that hit ~80 transactions per second.
- Distribution: xAI API, Grok app, grok.com, SuperGrok / X Premium+ tiers, plus third-party evaluation surfaces like arena.ai.
- Grok 4.7 follows in a few weeks as the 2.1T step-up, all-around better than 4.6 but slightly slower to serve, with even better token efficiency.
What Is Grok 4.6?
Grok 4.6 is the latest iteration of xAI's Grok family of frontier chat and reasoning models. It sits above Grok 4.5 in xAI's version ladder and is defined by three confirmed facts: a 1.5-trillion-parameter scale (the same V9 base as Grok 4.5), a post-training overhaul focused on supervised fine-tuning and reinforcement learning, and a stated capability target of challenging Kimi K3 and Claude Opus 4.8 while preserving Grok 4.5's inference economics. Musk's own post earlier framed the next model as "better than our 1.5T in every way," and the July 28 clarification landed the version as the 1.5T model "with significantly improved SFT & RL."
Positionally, Grok 4.6 is the model xAI expects to hold the frontier through August 2026 before Grok 4.7 (2.1T) takes the parameter-count leap. It succeeds Grok 4.5, which shipped on July 16, 2026 as a 1.5T-parameter model with strong SWE Marathon coding scores and ~80 transactions-per-second throughput. Our earlier analysis of the Grok 4.5 canary traces covers the V9 foundation and Cursor-derived training data that Grok 4.6 is built on top of.
Grok 4.6 at a Glance
| Attribute | Value |
|---|---|
| Developer | xAI |
| Type | Frontier chat / reasoning LLM |
| Parameters | 1.5 trillion (same V9 foundation as Grok 4.5) |
| Predecessor | Grok 4.5 (1.5T, V9 foundation) |
| Successor | Grok 4.7 (2.1T, releasing a few weeks later) |
| Modality | Text (multimodal not yet confirmed) |
| Context window | Not yet confirmed |
| Pricing | Not separately published (Grok 4.5: ~$2/M input, $6/M output) |
| Key upgrade | Significantly improved SFT and RL post-training |
| Release date | August 7, 2026 |
| Availability | xAI API, Grok app, grok.com, SuperGrok, X Premium+, arena.ai |
| License | Closed weights |
| Confirmed by | @elonmusk on X, July 28, 2026 |
How Grok 4.6 Works and What Makes It Different
The distinguishing bet in Grok 4.6 is capability gains through post-training, not scale. Where most frontier releases lean on a bigger base model, xAI held the 1.5T V9 foundation constant and poured the delta into supervised fine-tuning and reinforcement learning. That keeps Grok 4.5's throughput envelope intact while — per Musk's framing — pushing task quality into Kimi K3 and Opus 4.8 territory.
Three factors plausibly drive that claim:
- The V9 foundation is unchanged. Grok 4.5 was built on a 1.5T V9 base with training data that included a substantial pull from Cursor's AI coding platform. Grok 4.6 sits on the same base, so the inference-time cost profile carries directly forward.
- SFT and RL are the whole story. Musk explicitly described 4.6 as "the 1.5T model with significantly improved SFT & RL." Stronger reinforcement learning typically closes gaps on instruction-following, agentic reliability, and safety alignment — the axes where Grok has historically trailed Anthropic.
- Closed-loop engineering data. xAI has referenced feedback from Tesla, SpaceX, Neuralink, and The Boring Company as a training-signal source, and Musk flagged SpaceX's engineering corpus (excluding ITAR-restricted material) as being added during supplemental training for the larger 2T run. Whether the same corpus fed 4.6's post-training is not fully broken out.
The Kimi Capability Parity target, the V9 Efficiency Envelope, and the SFT/RL post-training step are the three anchors that separate Grok 4.6 from a routine version bump. If xAI hits all three, the release matters. If it only hits post-training polish, it still improves on Grok 4.5 — but the real parameter-count leap waits for Grok 4.7.
What You Can Do With Grok 4.6
The practical use cases inherit Grok 4.5's strengths at higher post-training capability:
- Agentic coding workflows. Grok 4.5 posted 29.0% on SWE Marathon, ahead of Claude Opus 4.8. A same-base successor with stronger RL is expected to push agentic and long-horizon coding tasks further without a speed penalty.
- Long-context research and document analysis. Grok already supports PDF and dataset ingestion through the xAI Files API, making it a fit for multi-document synthesis.
- Multi-model orchestration. Developers building agent systems that route between coding, reasoning, and search-heavy tasks are the primary early adopters signaled in community discussion.
- Cost-sensitive high-volume inference. Because Grok 4.6 preserves Grok 4.5's efficiency envelope, it remains attractive for applications where per-token cost dominates model selection.
How Grok 4.6 Compares
Only three peers are relevant given what xAI has said publicly:
| Model | Parameters | Position | Confirmed pricing |
|---|---|---|---|
| Grok 4.6 | 1.5T | Frontier, launched Aug 7, 2026 | Not separately published |
| Grok 4.5 | 1.5T | Prior xAI flagship | ~$2 / $6 per M tokens (in/out) |
| Kimi K3 | ~2.8T | Frontier, open-weights challenger | Published by Moonshot |
| Claude Opus 4.8 | Undisclosed | Frontier, Anthropic | Anthropic pricing |
The comparisons to watch are Grok 4.6 vs Kimi K3 and vs Claude Opus 4.8 — the two models xAI explicitly named as capability targets. Our Kimi K3 vs Claude analysis covers how the 2.8T open-weights model has been landing against Anthropic's stack, which is the same frontier band Grok 4.6 is aiming at.
Availability: How to Access Grok 4.6
Grok 4.6 launched on August 7, 2026 and is expected on the same channels as Grok 4.5:
- xAI API via the developer console (grok.com / x.ai). Grok 4.6 appears as a new model ID alongside 4.5.
- Grok app and grok.com consumer interface.
- SuperGrok ($30/mo) and SuperGrok Heavy ($300/mo) subscription tiers, plus X Premium+ for X-integrated access.
- Third-party evaluation surfaces including arena.ai, which announced same-week availability.
- EU rollout may still lag other regions, as it did for Grok 4.5.
For teams evaluating chat and reasoning workflows on the same architectural lineage, Grok 4.5 remains a direct comparison point.
What We Don't Know Yet
Several open questions remain even after the August 7 launch:
- Official benchmark numbers. xAI has not yet published SWE Marathon, MMLU, GPQA, or agentic benchmark scores for 4.6. Independent evaluations from surfaces like arena.ai are the first place to watch.
- Context window. Grok 4.5's window has not been formally extended in Grok 4.6 documentation.
- Pricing detail. xAI has not published a separate rate card for 4.6; the assumption is parity with Grok 4.5 given the shared 1.5T base, but this is unconfirmed.
- Multimodal scope. Vision, audio, and video capabilities are unmentioned in 4.6 disclosures so far.
- Full training mixture. Whether SpaceX's engineering corpus fed 4.6's post-training (as it will for the 2T Grok 4.7 run) has not been broken out.
Frequently Asked Questions
What is Grok 4.6?
Grok 4.6 is a 1.5-trillion-parameter frontier language model from xAI, launched on August 7, 2026. It reuses the same 1.5T V9 foundation as Grok 4.5 but ships with significantly improved supervised fine-tuning (SFT) and reinforcement learning (RL). It is positioned to challenge Moonshot's Kimi K3 and Claude Opus 4.8 while keeping Grok 4.5's speed and token efficiency.
When was Grok 4.6 released?
Grok 4.6 launched on August 7, 2026, the date Elon Musk publicly targeted on July 28, 2026. The larger 2.1T Grok 4.7 is scheduled to follow a few weeks later.
Is Grok 4.6 better than Kimi K3?
Musk positioned Grok 4.6 as a challenger to both Kimi K3 (~2.8T parameters) and Claude Opus 4.8, betting on stronger post-training (SFT and RL) rather than raw scale. Independent benchmark comparisons are still arriving; xAI has not yet published its own head-to-head numbers.
How many parameters does Grok 4.6 have?
Grok 4.6 has 1.5 trillion parameters — the same V9 foundation as Grok 4.5. The improvements are in post-training (supervised fine-tuning and reinforcement learning), not scale. The 2-trillion-parameter model earlier rumored under the 4.6 name is actually being held for a later release; Grok 4.7 will step up to 2.1T.
How much does Grok 4.6 cost?
xAI has not published a separate price list for Grok 4.6 at launch. Grok 4.5 is priced at roughly $2 per million input tokens and $6 per million output tokens through the xAI API; because Grok 4.6 reuses the same 1.5T foundation, that band is the reasonable working baseline until xAI publishes official 4.6 rates.
Is Grok 4.6 open source?
No. Grok 4.6 shipped as a closed-weights model with API access. Prior Grok generations have been API-only or partially released long after launch, and xAI has not signaled open weights for 4.6.
How can I access Grok 4.6?
Grok 4.6 is available through the xAI API, the Grok app, grok.com, and the X Premium+ / SuperGrok subscription tiers, following the same distribution pattern as Grok 4.5. It is also being made available in third-party evaluation surfaces such as arena.ai.
What to Watch Next
Three signals will now define how Grok 4.6 lands: an official xAI model card publishing benchmarks, independent SWE Marathon and long-horizon agentic scores from third-party evaluators, and the Grok 4.7 (2.1T) release a few weeks out — which will make it clear how much of the frontier push came from post-training versus scale. The Kimi K3 and Opus 4.8 head-to-heads are the ones that matter most; anything short of parity would undercut the "1.5T at 4.5-speed" pitch.
Building similar chat and reasoning workflows today? On kie.ai you can try Grok 4.6, Claude Opus 5, and Gemini 3.6 Flash.
About Sofia Marenco
Sofia stress-tests new models on coding and reasoning benchmarks and reports what holds up.
View all posts by Sofia Marenco