On September 21, 2026, xAI released Grok 4.7, its most capable model for coding and knowledge work. It arrives just five weeks after Grok 4.6 — and if you asked for 4.6, here is the correction that matters: 4.7 is the current model, served at the same price and speed as 4.6.
New larger base model, longer reinforcement learning on harder multi-hour tasks, better self-verification, and native understanding of the Grok Bot harness. Here is what shipped, what it means for coding agents, and what to do this week.
1. The news in two paragraphs
Grok 4.6 (August 12, 2026) was built with Cursor: supplemental training on anonymized Cursor workflow data, 500k context window, configurable reasoning effort (low, medium, high, xhigh), priced at $2 input / $6 output per 1M tokens. It focused on long-running agents and visual, interactive work.
Grok 4.7 (September 21, 2026) keeps that positioning and changes the engine underneath: a new, larger base model than 4.6, a longer RL run weighted toward problems that take many hours, better verification of its own work and longer-context management, plus native training on the Grok Bot harness for conversational and general knowledge work.
Ficha del lanzamiento
- • Launch: September 21, 2026 — model ID grok-4.7 on the xAI API
- • Price: from $2 input / $6 output per 1M tokens — same as 4.6; fast variant 2x speed at 2x price
- • Training: larger base + longer RL on harder multi-hour task mix (vendor claim)
- • Availability: Cursor, Grok Build, xAI API, third-party harnesses, routers and clouds
- • Supersedes: Grok 4.6 (Aug 12, 2026) — same price, older base; consumer apps come later
2. What 4.7 changes under the hood
Three upgrades, all aimed at long-horizon work. First, the base model itself is new and larger than 4.6’s — this is not a fine-tune, it is a new foundation. Second, the RL phase ran longer on a harder mix, explicitly weighted toward tasks that take many hours to complete. Third, the model was trained to verify its own work and to manage longer context without losing the thread.
The fourth change is easy to miss and very practical: 4.7 natively understands the Grok Bot harness, which makes it better at conversational tasks and general knowledge work, not just code. xAI is clearly pushing Grok from a coding specialist toward the generalist front line.
The three upgrades
Bigger base
New, larger base model vs 4.6 — a new foundation, not a fine-tune of the old checkpoint.
Longer RL
Longer reinforcement run on harder tasks weighted toward multi-hour problems and self-verification.
Bot harness
Native Grok Bot harness understanding — better conversational and general knowledge behavior.
Golden rule
Same price and speed as 4.6, newer base and longer training. There is no reason to start new work on 4.6 — point new agent loops at 4.7 and keep 4.6 only where you already validated it.
3. Evidence and context
xAI’s framing is capability-per-dollar at the frontier: 4.7 is highly competitive in its class at $2/$6, with a fast variant for latency-sensitive loops. The 4.6 model card documents the family’s safety posture across cyber, bio, jailbreak robustness and CBRN refusals, with a stated policy of never silently downgrading intelligence or falling back to other models.
What we know from the 4.6 generation carries over as the baseline: 500k context, text and image input with text-only output, no text output limit, function calling, web and X search, code execution, and reasoning effort levels. Treat 4.7’s capability deltas as vendor claims until independent coding benchmarks land.
Verified facts
- • Grok 4.7 launched Sep 21, 2026, same price/speed as 4.6 (xAI announcement)
- • New larger base + longer RL on harder multi-hour mix (xAI announcement)
- • 4.6 baseline: 500k context, $2/$0.50 cached/$6 under 200k tokens (xAI docs)
- • Available day one in Cursor, Grok Build, API, routers and clouds (xAI announcement)
Read the fine print
No independent benchmarks of 4.7 exist yet — it launched three days ago, within 48 hours of Opus 5.5 and GPT-6 Sol/Luna. The knowledge cutoff for 4.7 was not published in the launch post (the 4.6 card lists Feb 1, 2026). Verify before citing dates in production work.
4. Quickstart: call grok-4.7 today
If you already call the xAI API, the migration is a model-string swap plus a decision about reasoning effort. Grok 4.7 supports low, medium, high (default) and xhigh — start at the default, raise to xhigh only for the hard multi-hour tasks it was trained for, and measure tokens per task before and after.
In Cursor and Grok Build it is available today with no code changes. On the API it speaks both Responses and Chat Completions, with function calling, web search, X search and code execution as tools.
Minimal API call
Chat Completions against the xAI API — swap the model string, keep your tools.
POST https://api.x.ai/v1/chat/completions
{
“model”: “grok-4.7”,
“reasoning_effort”: “high”,
“messages”: [
{“role”: “system”, “content”: “You are a senior engineer.”},
{“role”: “user”, “content”: “Refactor this module...”}
],
“tools”: [ ... your function definitions ... ]
}Honest limits
- • Text-only output — image in, no image out; visual and interactive work means code artifacts, not renders.
- • Consumer surfaces (web, mobile, Grok-in-X) get 4.7 later — today it is a builders’ release.
- • Not for autonomous high-stakes decisions (medicine, law, finance, safety-critical) without expert oversight — stated in the model card.
- • 4.6-era context above 200k tokens doubles API price ($4/$1/$12) — confirm 4.7’s long-context pricing before big runs.
5. What to do this week
Because 4.7 is priced identically to 4.6, the experiment is nearly free: point a copy of your hardest agent loop at grok-4.7 and compare completion rate, steps and output tokens against your 4.6 baseline. xAI’s own claim — fewer steps, fewer output tokens than other frontier models — is directly measurable.
Adoption checklist
- 1. Swap one staging agent loop from grok-4.6 to grok-4.7; keep prompts and tools identical.
- 2. Sweep reasoning effort (high vs xhigh) on your hardest task class and log tokens-per-completion.
- 3. Test the conversational path too — the Bot harness training should show up outside pure code.
- 4. Re-check long-context pricing above 200k tokens before scaling document-heavy workloads.
- 5. Hold the frontier comparison open: run the same eval on Opus 5.5 and GPT-6 Sol when your router supports them.
Why this launch matters
Three frontier releases in 72 hours (Opus 5.5, Grok 4.7, GPT-6 Sol/Luna) with two of them cutting effective cost. The bottleneck for agents is shifting from model price to eval quality — the teams with good task suites capture the gains first.
Verdict
Grok 4.7 is a straightforward upgrade: newer and larger base, longer training on exactly the tasks agents struggle with, same price as 4.6. If you build coding agents on xAI, there is no case for staying on 4.6 for new work.
The open question is the three-way comparison nobody has run yet — 4.7 vs Opus 5.5 vs GPT-6 Sol on real repos. Run that eval on your codebase and you will know more than any launch post can tell you.
Sources for this post
Primary and official docs
- • xAI announcement: Introducing Grok 4.7 (Sep 21, 2026)
- • xAI announcement: Introducing Grok 4.6 (Aug 12, 2026)
- • xAI docs: Grok 4.6 model page + release notes
- • Grok 4.6 model card (PDF, rev Aug 17, 2026)
- • xAI: Grok 4.6 on Microsoft Foundry (Aug 26, 2026)
