Grok 4.6 and Grok 4.7 are arriving within weeks of each other, according to Elon Musk's July 2026 posts on X. Grok 4.6 is a 2-trillion-parameter model — a 33% scale jump from the 1.5T Grok 4.5 that launched July 8 — and Musk has stated it will be "better than our 1.5T in every way" while keeping speed and token efficiency close to its predecessor. Grok 4.7 follows roughly two weeks later. No benchmarks or spec sheets exist for either model yet, so this guide separates what xAI has actually confirmed from what is speculation, and gives you a practical framework for evaluating each release against your real workloads rather than against leaderboard headlines.
Last verified: 2026-07-30
- Grok 4.5 launched July 8, 2026 (1.5T parameters, $2/$6 per million tokens, 500K context)
- Grok 4.6: 2T parameters, training completing late July per Musk, public release targeted early August
- Grok 4.7: ~2.1T parameters, "a few weeks" after 4.6 (late August/early September estimate)
- No official benchmarks or pricing published for 4.6 or 4.7 yet — treat any scores you see as guesswork
- xAI (now branded SpaceXAI) is shipping frontier models on a weeks-long cadence, not months
What Did xAI Actually Announce for Grok 4.6 and 4.7?
The announcement came as a two-line reply on X. On July 24, 2026, Musk posted that Grok 4.6 would arrive in two weeks and Grok 4.7 in four weeks — no blog post, no launch event, just a reply to a thread. This followed an earlier July 18 post where Musk confirmed the 2T parameter count and said the model was finishing its initial training run "next week."
Here is what has been confirmed through Musk's own posts and subsequent reporting:
| Model | Parameters | Timeline | Source |
|---|---|---|---|
| Grok 4.5 | 1.5 trillion | Launched July 8, 2026 | xAI official announcement |
| Grok 4.6 | 2 trillion | Training completing late July; public release targeted ~August 7 | Musk on X, July 18–24 |
| Grok 4.7 | ~2.1 trillion | "A few weeks" after 4.6 | Musk on X, July 24 |
The announcement carried no benchmarks, no pricing details, and no technical deep-dive. xAI has not published a model card or spec sheet for either upcoming model. Musk stated Grok 4.6 aims to surpass Moonshot AI's Kimi K3 (a 2.8T open-weights model) while maintaining speed and token efficiency "close to our 1.5T" — meaning Grok 4.5's performance characteristics.
Confidence label: Reported — the timeline and parameter counts come from Musk's social posts; the August 7 date for Grok 4.6 originated in a follow-up post and was surfaced by tbreak.com. These are targets, not released specifications.
How Fast Is Grok 4.5 Right Now?
Grok 4.5 — the current flagship that 4.6 and 4.7 will replace — already has verified performance data from xAI's official announcement and independent benchmarks. Understanding this baseline matters because Musk says 4.6 will match its speed and efficiency while being smarter.
Speed: Grok 4.5 is served at approximately 80 tokens per second, which xAI describes as "flash-model speed." Independent measurement by Artificial Analysis recorded 91.3 TPS — even faster than xAI's own claim.
Context window: 500,000 tokens. This is actually smaller than the previous Grok 4.3's 1 million-token window, a reduction xAI has not publicly explained.
Token efficiency: This is the number that changes the economics. On SWE-Bench Pro (a benchmark for AI coding agents), Grok 4.5 completes tasks using an average of 15,954 output tokens. Claude Opus 4.8 uses 67,020 for the same work — that is a 4.2x gap. Fewer tokens means faster task completion and lower API bills, especially for agent runs that execute hundreds or thousands of tasks.
Pricing: $2 per million input tokens, $6 per million output tokens, with cached input at $0.50. For comparison, Opus 4.8 costs $5/$25 per million tokens. On a single SWE-Bench Pro task, Grok 4.5 costs approximately $0.096 in output tokens versus approximately $1.00 for Opus 4.8 — roughly 10x cheaper per completed task.
Confirmed benchmarks from xAI's official launch post:
| Benchmark | Grok 4.5 | Claude Opus 4.8 | Claude Fable 5 |
|---|---|---|---|
| SWE-Bench Pro | 64.7% | 69.2% | 80.4% |
| DeepSWE 1.0 (provider harness) | 62.0% | 55.75% | 66.1% |
| DeepSWE 1.1 (neutral harness) | 53% | 59% | 70% |
| Terminal Bench 2.1 | 83.3% | 78.9% | 84.3% |
| SWE Marathon (pass@1) | 29.0% | 26.0% | 24.0% |
The picture is nuanced. Grok 4.5 wins on provider-controlled DeepSWE 1.0 and SWE Marathon, and ties near the top of Terminal Bench. But on SWE-Bench Pro and the neutral DeepSWE 1.1 harness, it trails both Opus 4.8 and Fable 5. The honest read: Grok 4.5 is not the frontier reasoning model, but it delivers more intelligence per dollar than competitors.
For a deeper dive into how these benchmarks actually work and why provider-controlled scores can be misleading, see our guide to DeepSWE vs SWE-Bench Pro.
How Was Grok 4.5 Trained?
xAI trained Grok 4.5 across tens of thousands of NVIDIA GB300 GPUs as part of the Colossus cluster in Memphis, Tennessee. The training pipeline involved three key stages:
Data curation: Beyond raw token volume, xAI invested in deduplication, quality scoring, and domain-focused selection to keep the data mixture high-coverage and high-signal rather than just large.
Reinforcement learning at scale: RL training covered hundreds of thousands of tasks, centered on multi-step software engineering work, with automated and model-based grading.
Asynchronous training architecture: The stack was built for highly asynchronous training, meaning agentic rollouts (the model practicing tasks) can run for many hours while learning continues across tens of thousands of GPUs in the background.
Grok 4.5 was co-trained with Cursor (the AI code editor), using trillions of tokens of Cursor developer session data alongside broader STEM and knowledge-work material. One caveat: Cursor disclosed that an earlier snapshot of the Cursor codebase was accidentally included in Grok 4.5's training data, which potentially inflates CursorBench scores specifically. Third-party benchmarks like DeepSWE and SWE-Bench Pro use independent infrastructure and are not affected.
Why Is xAI Shipping This Fast?
The gap between Grok 4.5 (July 8) and the announced 4.6/4.7 timeline is measured in weeks. For context, most AI labs take months between major releases. xAI's release cadence has compressed dramatically:
- Grok 4 launched July 2025
- Grok 4.3 launched April 2026
- Grok 4.5 launched July 8, 2026
- Grok 4.6: targeted August 2026
- Grok 4.7: targeted late August/early September 2026
That is three major model versions in a single summer. The Colossus cluster in Memphis currently houses over 500,000 GPUs across two buildings, with a third facility being converted in Southaven, Mississippi, and xAI's stated ambition to reach one million GPUs. This infrastructure scale is what enables the compressed timeline.
The competitive target is clear: Musk named Kimi K3 (2.8T parameters, launched July 16, 2026) as the model Grok 4.6 aims to surpass. Kimi K3 scores 57 on the Artificial Analysis Intelligence Index versus Grok 4.5's 54, but costs approximately 3x more per task ($0.94 vs $0.31). Grok 4.6's pitch is "close the intelligence gap, keep the cost advantage."
Should You Wait for Grok 4.6 or Start With 4.5 Today?
No — start with 4.5 now, on your real tasks, and build a comparison baseline. Here is why.
The people who will get the most value out of Grok 4.6 are the ones who already know what Grok 4.5 does badly on their specific workflows. If you wait for the release without testing the current model, you cannot tell whether the upgrade is an improvement or just a different set of trade-offs.
This is the core issue with rapid model iteration: a newer model does not fix a vague prompt. The difference between a poor result and a great one is almost always in how specifically you described what you want, not in which version number of the model you used.
The practical approach is to build your own benchmark on real work — pick one task you repeat regularly, run it through Grok 4.5, document exactly what broke and where, then run the same task with the same prompt when 4.6 lands. In ten minutes you know whether the upgrade is worth switching for.
How to Prepare for Fast Model Releases: A 3-Step System
Model releases happen every few weeks now. Your workflow should not break when a new model lands. Here is the system that works regardless of which model ships next.
Step 1: Stop Chasing the Version Number
A newer model does not mean better results on your tasks. What actually changes output quality is prompt specificity. Describe what you want like you are talking to someone who cannot see your screen: what is on the page, what happens when you click, and what the finished result should look like. Same model, same task, completely different result — the only variable was how specific the instruction was.
Step 2: Build a System, Not a Model Dependency
If your entire setup is one chat window and a saved prompt, every new release resets you to the start. If your setup is a system — structured prompts, documented workflows, repeatable tasks, context files — then a new model is a swap, not a rewrite. This is why AI agent operating systems and multi-model orchestration tools matter: they let you change the model without rebuilding the process around it.
Step 3: Baseline Before You Upgrade
Before any new model release:
- Pick one task you do over and over (a report you rebuild weekly, a page you keep redesigning, a document you generate from a template).
- Run it through Grok 4.5 with your best prompt.
- Write down exactly what worked, what broke, and where the output fell short.
- Save the prompt, the input, and the output.
- When Grok 4.6 lands, run the same task with the same prompt.
- Compare side by side. Is the output better, faster, cheaper — or just different?
This habit works for every model that ships after this one, too. For a deeper framework on evaluating models by cost-per-completed-task rather than per-token pricing, see our guide to AI agent loop cost strategy.
What Does Grok 4.5 Already Do Well Beyond Coding?
Grok 4.5 is the default model in Grok Build, xAI's coding agent CLI. But it also handles knowledge work tasks that go beyond software engineering:
- Excel: Builds multi-sheet financial models that pull research data from the web, uses formulas across sheets, and leaves notes for the next person who opens the file.
- PowerPoint: Creates diagrams from native PowerPoint shapes and designs slide content.
- Word: Writes clean prose for documents.
- Structured data extraction: Tops independent benchmarks at pulling structured data out of messy business documents.
These capabilities are available through Microsoft Office add-ins and through the Grok Build CLI. If your work involves document generation or data extraction, Grok 4.5 is already a practical tool — you do not need to wait for 4.6 to start.
For a guide on using Grok inside Google Workspace (Docs, Sheets, Slides), see our Grok AI for Google Workspace guide.
What This Means for You
If you are a developer, builder, or small business using AI:
- Do not wait. Grok 4.5 is available today, fast, cheap, and barely tested on most people's actual tasks. The baseline you build now is the comparison that tells you whether 4.6 is worth the switch.
- Build model-agnostic workflows. With Grok 4.6 and 4.7 landing weeks apart, plus Anthropic's Claude Opus 5 and OpenAI's GPT-5.6 also iterating, your process needs to survive a model swap without breaking. For a detailed framework on this, see our guide to Claude Opus 5 enterprise cost strategy — the effort-dial approach applies to any model with configurable reasoning.
- Watch for the three confirmation signals: When Grok 4.6 launches, look for (1) an official xAI blog post with benchmarks and pricing, (2) a new model ID in the xAI API console or Grok Build, and (3) independent third-party benchmark scores from evaluators like Artificial Analysis. Until you see all three, treat any claims as marketing.
- Price your workloads, not your tokens. Grok 4.5 costs $2/$6 per million tokens, but the number that matters is cost per completed task. At 4.2x fewer output tokens than Opus 4.8 on SWE-Bench Pro, the effective cost per task is roughly 10x lower — even though Grok 4.5 is not the top scorer on that benchmark.
FAQ
Q: When will Grok 4.6 be released? A: Musk posted on July 24, 2026 that Grok 4.6 would arrive in "two weeks," pointing to early August 2026. An earlier post set training completion for late July. These are targets, not confirmed dates — xAI has not published an official launch date.
Q: What is the difference between Grok 4.5 and Grok 4.6? A: Grok 4.5 has 1.5 trillion parameters; Grok 4.6 has 2 trillion (a 33% increase). Musk stated 4.6 will be "better than our 1.5T in every way" while keeping speed and token efficiency close to 4.5's. No benchmarks have been published for 4.6.
Q: How much will Grok 4.6 cost? A: xAI has not announced pricing for Grok 4.6. Grok 4.5 costs $2 per million input tokens and $6 per million output tokens. If Musk's claim about maintaining 4.5-level efficiency holds, pricing may stay in the same range — but this is speculation until confirmed.
Q: Should I switch from Claude or GPT to Grok when 4.6 launches? A: Only if your own benchmark data supports it. On neutral-harness benchmarks, Grok 4.5 trails Claude Opus 4.8 and Fable 5 on several coding evals. Grok's advantage is cost-per-task and speed, not raw benchmark scores. Run the same task on both models and compare the actual output, not the leaderboard.
Q: Is Grok 4.5 available in the EU? A: EU availability was limited at launch on July 8, 2026. xAI expected EU access by mid-July 2026. Check the xAI console for current regional availability before planning on EU production use.
Q: What is Grok Build?
A: Grok Build is xAI's CLI-based coding agent (installed via curl -fsSL https://x.ai/cli/install.sh | bash). Grok 4.5 is the default model in Grok Build, which coordinates terminal-based coding work with parallel subagents. It is available as a beta for SuperGrok Heavy subscribers ($300/month).
Discussion
0 comments