AI models · Review

Grok 4.7 Review (Sep 2026): What's New for Melbourne SMEs — Without the Hype

xAI launched Grok 4.7 on 21 September 2026 as the successor to 4.6 and their recommended model for code and chat. If you are a Melbourne business already using AI agents, Cursor, or the xAI API, you may be wondering what actually changed, whether the upgrade is worth it, and what the caveats are. This is a practical review — what the official announcement confirms, what it costs in real money, where you can use it today, and when it makes sense for Australian SMEs.

What shipped on 21 September

xAI published the official announcement at x.ai/news/grok-4-7 on 21 Sep 2026. Here is what they confirmed:

  • Successor to Grok 4.6: Recommended for code, chat, agentic tasks, and knowledge work.
  • New larger base model: Compared to 4.6, Grok 4.7 uses a larger base model, longer reinforcement learning, better self-verification, and improved long-context handling.
  • Trained for Grok Bot harness: The model is optimised for agentic workflows — tools, function calling, code execution, search integration.
  • 500k context window: Enough for large codebases, multi-document analysis, or long transcripts.
  • Knowledge cut-off: May 2026. Without Web Search or X Search enabled, Grok 4.7 cannot surface events after mid-year.
  • Text + image input, text output: Supports JPG and PNG up to 20 MiB.
  • Reasoning modes: Low, medium, high, and xhigh — choose based on task complexity vs latency.
  • Tools: Function calling, web search, X search, and code execution built in.
  • Encrypted reasoning: The Responses API always returns reasoning in encrypted form — you get the answer, not the working.

That is what xAI stated in the launch announcement. We are not adding invented benchmarks, speculated parameter counts, or unverified features.

Where you can use it today

Grok 4.7 is available now via:

  • Cursor: The model slug is grok-4.7. If you use Cursor for coding, it is already there.
  • Grok Build: xAI's own developer platform for building agents and workflows.
  • xAI API: Standard endpoint api.x.ai/v1 and US regional endpoint us.api.x.ai/v1 (the US regional endpoint adds a 10% token premium).
  • Grok 4.7 Fast: A faster, more expensive variant available in Cursor and Grok Build only — not on the public API and not on the free tier. Priced at 2× the standard rates.

Important: We have not verified whether Grok 4.7 is live on X Premium, grok.com, or the iOS/Android Grok app. The launch announcement does not state consumer rollout the way the Grok 4.5 announcement did. If you use those platforms, check directly before assuming 4.7 is available.

What the numbers mean in plain English

xAI published vendor benchmarks in the launch post. These are xAI's own numbers — not independent third-party evals. Treat them as indicative, not gospel. Here is what they claim:

  • CursorBench 4.0: Grok 4.7 scored 46.3% vs Grok 4.6's 40.4%. This benchmark measures code completion and generation in real IDE workflows.
  • DeepSWE: 71.0% vs 65.2% — tests whether the model can solve software engineering tasks in multi-file codebases.
  • EEBench: 64% vs 53% — measures coding task accuracy across languages and frameworks.
  • Terminal-Bench: 38% vs 20.3% — tests whether the model can generate correct shell commands and scripts.

In plain terms: xAI claims Grok 4.7 is better than 4.6 at code generation, debugging, and agentic tasks. The gaps are meaningful — 5 to 18 percentage points depending on the benchmark. But these are vendor numbers. Run your own tests on real work before migrating production systems.

Grok 4.7 vs 4.6: what actually changed

According to xAI, the key differences are:

  • Larger base model: More parameters means better understanding of complex instructions and longer context retention.
  • Longer reinforcement learning: The model was trained for more iterations to improve reasoning and self-correction.
  • Better self-verification: Grok 4.7 is more likely to catch its own mistakes before returning an answer.
  • Improved long-context handling: The 500k context window is the same as 4.6, but 4.7 is better at using the full window without losing track of earlier instructions.
  • Optimised for agentic workflows: The model was trained specifically for tool use, function calling, and multi-step tasks — the Grok Bot harness xAI uses internally.

In short: 4.7 is a refinement, not a revolution. If 4.6 works for you, there is no urgent reason to switch. If you hit edge cases where 4.6 struggles with long context or multi-step reasoning, 4.7 may help.

Pricing (with AUD context)

Grok 4.7 pricing matches Grok 4.6. Here are the rates in USD and rough AUD equivalents (at ~1.50 AUD per USD, September 2026):

Standard Grok 4.7

  • Prompts under 200k tokens: $2.00 / $0.50 / $6.00 per million (input / cached / output) = ~$3 / $0.75 / $9 AUD
  • Prompts 200k or more: $4 / $1 / $12 per million = ~$6 / $1.50 / $18 AUD

Grok 4.7 Fast (Cursor/Grok Build only)

  • Under 200k tokens: $4 / $1 / $12 per million = ~$6 / $1.50 / $18 AUD
  • 200k or more: $6 / $1.50 / $18 per million = ~$9 / $2.25 / $27 AUD

What that means for a Melbourne SME

If you are running an agent that handles 500 customer queries a day, each with a 2,000-token prompt and 500-token response:

  • Daily cost: 500 × (2,000 × $3 / 1M + 500 × $9 / 1M) = ~$5.25 AUD per day
  • Monthly cost: ~$160 AUD

That is in line with other frontier models. Pricing alone is not a reason to switch — pick the model that works best for your workflow.

Our workshop tests (so far)

iWizz is testing Grok 4.7 on code-heavy tasks for client projects — config file generation, log parsing, and script reviews. Early observations:

  • Code generation: Noticeably better than 4.6 at generating correct Python and JavaScript for multi-step workflows. Fewer syntax errors, better edge-case handling.
  • Long context: Handles 20-30k token prompts (multiple files + instructions) more reliably than 4.6. Less likely to lose track of earlier requirements.
  • Reasoning: The encrypted reasoning output is still a limitation — we cannot inspect the model's working. For debugging or audit trails, this is a problem.
  • Speed: Standard 4.7 is slightly slower than 4.6. Grok 4.7 Fast is faster but costs 2× as much — only worth it for latency-critical tasks.

These are early tests, not formal evals. We will update this post if our conclusions change after more production use.

Caveats before you spend

Verify all outputs

Grok 4.7, like all LLMs, can confidently generate wrong answers. If you are building customer-facing agents or business-critical automation, verify outputs before they go live. This is not unique to Grok — it applies to every model.

Search tools required for live data

Without Web Search or X Search enabled, Grok 4.7 knows nothing after May 2026. If your use case needs current stock prices, recent regulations, or live event schedules, you must pass search tools or integrate external data yourself.

Encrypted reasoning is a black box

The Responses API returns reasoning in encrypted form. You cannot see the model's step-by-step working. If your workflow requires inspecting logic for debugging, compliance, or quality assurance, this is a deal-breaker. Other models (Claude, GPT) expose reasoning traces more transparently.

US regional endpoint adds 10% cost

If you use us.api.x.ai/v1 instead of the standard endpoint, you pay a 10% token premium for lower latency. Only worth it if response time matters more than cost.

Fast variant is Cursor/Grok Build only

Grok 4.7 Fast is not available via the public API. If you need speed and you are not using Cursor or Grok Build, you are stuck with standard 4.7.

Who should upgrade to 4.7

Upgrade to Grok 4.7 if:

  • You use Cursor for coding and want better completions, fewer errors, and improved long-context handling.
  • Your agent workflows involve multi-step reasoning, function calling, or tool orchestration — 4.7 is optimised for this.
  • You hit edge cases with 4.6 where the model loses context or produces vague answers.
  • You are already paying for 4.6 and the cost structure is identical — there is no financial penalty for testing 4.7.

Stay on Grok 4.6 (or another model) if:

  • Your current setup works reliably and you have no complaints.
  • You need transparent reasoning outputs for compliance or audit — 4.7's encrypted reasoning is a blocker.
  • You use a different ecosystem (OpenAI, Anthropic, Google) and switching adds integration overhead without clear upside.
  • You need verified, independent benchmarks for procurement or regulatory approval — xAI's vendor numbers are not enough.

What about consumer availability?

We cannot confirm whether Grok 4.7 is live on X Premium, grok.com, or the iOS/Android Grok app. The launch announcement does not mention consumer rollout. If you rely on those platforms, check directly before assuming 4.7 is available. For API and Cursor users, it is live now.

Bottom line for Melbourne SMEs

Grok 4.7 is a solid upgrade for businesses using agents, Cursor, or the xAI API — especially for code generation, agentic workflows, and long-context tasks. The vendor benchmarks look good, but run your own tests on real work before migrating production systems.

Pricing is the same as 4.6. Speed is slightly slower unless you pay 2× for Fast. Encrypted reasoning is still a limitation for debugging and audit. And consumer availability (X Premium, grok.com) is unconfirmed.

For businesses considering AI automation for the first time, we cover practical examples in How AI can free up your time in business. For businesses wanting to show up in AI-powered search, we wrote about structured data and citability in Agentic search for local business.

How iWizz helps Australian businesses with AI

iWizz builds and runs agent workflows for Australian SMEs — automating bookings, receipt processing, content drafts, email triage, review replies, and more. We use whichever models fit the job: GPT, Claude, open-source, and now testing Grok 4.7 for code-heavy tasks.

We do not sell AI strategy decks or run twelve-month pilots. We scope one workflow, put it into production, and measure whether it saves time or money. If it works, we expand. If it does not, we stop. That is how SMEs should approach AI — practical, measurable, reversible.

Talk to iWizz

If you want to automate part of your business — customer support, receipts, content, dashboards — and you want straight advice on which models and tools actually fit the job, call us. We work with hospitality, retail, clinics, trades, clubs, and swim schools across Australia.

No fake testimonials, no invented metrics, no twelve-slide roadmaps. Just practical systems that either save you time or get turned off.

Talk to iWizz