Grok 4.5: What Changed, Pricing, and Coding Use

Grok 4.5 is xAI’s mid-July 2026 flagship for coding, agentic tasks, and knowledge work, launched 16 July 2026 with API list rates of $2 / $6 per 1M tokens, day-one presence in Cursor and Grok Build, and a later path into GitHub Copilot (28 July). xAI states the model was trained alongside Cursor.

This is a PromptHive launch analysis. Official claims come from xAI’s Grok 4.5 post, API docs, and the Copilot availability post. Vendor benchmarks are their charts — we translate them into decisions for people choosing models inside Cursor, GitHub Copilot, or raw APIs.

Verified 2026-08-05 against xAI’s public Grok 4.5 materials and the 28 July Copilot announcement. Rates, free promos, and picker availability can move; re-check the vendor before you budget.

The short answer

• What launched: Grok 4.5 (grok-4.5 in xAI’s API sample) on 16 July 2026, positioned as xAI’s smartest model for coding, agents, and knowledge work.
• Price: $2 input / $6 output per 1M tokens (xAI list). xAI also claims strong token efficiency vs named competitors on SWE-style tasks.
• Training angle: Built alongside Cursor — distribution and data partnership, not a Cursor SKU rename.
• Where you can use it: xAI API / console, Grok Build, Cursor (all plans per launch post), consumer Grok surfaces (grok.com / X / mobile per xAI’s usual Grok path), then GitHub Copilot from 28 July 2026.
• PromptHive take: treat Grok 4.5 as a coding-cost challenger in multi-model editors. Measure against Claude Opus 5 and GPT-5.6 Sol on your tickets before changing team defaults.

Related: Cursor review, how to use Cursor, GitHub Copilot, complete AI coding guide, Codex vs Claude Code, Claude Opus 5.

What was officially announced

16 July 2026 — Introducing Grok 4.5

From x.ai/news/grok-4-5:

xAI’s post shows competitive bar charts (DeepSWE, SWE Marathon, Terminal Bench, SWE Bench Pro) with competitor figures attributed to other vendors’ materials. Read those as marketing context, not as PromptHive scores.

Office / Grok Build angle

Beyond pure coding, xAI highlights Grok 4.5 as default in Grok Build, with demos around Excel models, PowerPoint diagrams, and Word prose, plus Microsoft Marketplace plugin links. That is a knowledge-work expansion, not a reason to replace your spreadsheet process unattended.

28 July 2026 — GitHub Copilot

xAI’s Copilot post and GitHub’s changelog state Grok 4.5 rolls out in GitHub Copilot for Pro, Pro+, Max, Business, and Enterprise, selectable in VS Code, Visual Studio, and Copilot CLI (orgs may need to enable the model). GitHub’s note also references a large context window and reasoning-effort controls — confirm current limits in GitHub’s docs for your SKU.

What’s new vs the previous Grok mental model

What did not change: multi-file agents still need git discipline and review. A cheaper, faster model that lands plausible bugs still costs engineer time — our constant theme in the coding guide.

Features explained (without the brochure)

Coding and agent loops

xAI stresses multi-step software engineering RL, long agentic rollouts, and “one prompt” full apps. For teams, the useful question is narrower: does it finish your tickets with fewer interventions than your current default? Benchmarks that mix harnesses are not interchangeable with Claude Code or Codex runs.

Token efficiency vs sticker price

$2 / $6 is already aggressive against many Opus-class list rates. xAI’s efficiency charts try to double the argument: fewer output tokens per resolved task. Only your traces prove it — log tokens for the same ticket on Grok 4.5 vs Claude Opus / GPT Sol.

Cursor partnership (why the hero is Cursor)

There is no PromptHive product logo for Grok in our set; more importantly, xAI’s own launch ties Grok 4.5 to Cursor training and day-one availability. Practically:

• Cursor users get a first-class model option without leaving the editor they already pay for (Cursor review, how to use Cursor).
• Partnership ≠ automatic best default. Privacy mode, usage pools, and overage still follow Cursor’s billing, not xAI’s raw API card.
• If you are evaluating Grok only as an API, ignore the Cursor UI and use the console sample.

GitHub Copilot path

Copilot availability matters for enterprises standardized on VS Code + GitHub. Grok becomes one picker option among others, subject to org admin enablement and SKU. It does not replace Copilot’s autocomplete product identity — see GitHub Copilot review and how to use Copilot.

Knowledge cutoff and freshness

Plan for a February 2026-class cutoff unless your stack adds browsing or retrieval. For “what shipped this week?” questions, pair any model with search tools or a research product — do not assume chat memory is live.

Pricing and availability

Trap: free launch usage in Grok Build or Cursor is not a permanent $0 coding team. Model the steady-state bill: Cursor Pro overages, Copilot premium requests, or API volume.

PromptHive hands-on analysis

We did not re-run DeepSWE or Terminal Bench. Practical reading:

1. Grok 4.5 is aimed at the same wallet as Claude Opus and GPT Sol — hard coding agents — with a lower list-rate wedge. That is the story worth testing.
2. Cursor co-training is distribution leverage. If you already live in Cursor, the switching cost is a model toggle and a harness, not a new IDE. If you live in Claude Code, the cost is process change (Codex vs Claude Code for the adjacent OpenAI comparison).
3. Copilot inclusion (28 July) is the enterprise legitimacy step. Admins should treat enablement as a policy decision (data handling, allowed models), not only a developer convenience.
4. Speed + efficiency claims matter most for agent farms. For one-off hard design work, review quality still dominates unit token price.
5. Competition moves monthly. Cross-read Claude Opus 5 and GPT-5.6 Sol/Terra/Luna before freezing a 2026 standard model list.

Pros

• Simple, aggressive API list rates ($2 / $6) for a coding-positioned flagship
• Day-one Cursor availability plus later Copilot support covers two huge IDE audiences
• Explicit agentic / SWE training narrative (even if charts are vendor-framed)
• Token-efficiency pitch can compound savings if traces stay short
• Grok Build + Office plugin story broadens beyond pure repo edits

Cons

• Vendor leaderboards are not your production suite
• Multi-model Cursor/Copilot bills are easy to underestimate (overages, premium requests)
• Knowledge cutoff means retrieval is still required for current events and new libraries
• No PromptHive long-running “Grok product review” page yet — this is a launch note, not a full tool scorecard
• Safety, moderation, and enterprise DPA details still need your counsel’s checklist independent of marketing demos

Best use cases

How to trial it (about 45 minutes)

1. Cursor path: on a feature branch, set the agent model to Grok 4.5, clean git status, one real ticket with tests as the success signal (Cursor workflow).
2. API path: call grok-4.5 with the same system prompt and tools you use for another vendor; log tokens and wall-clock.
3. Copilot path (if licensed): enable Grok 4.5 in org settings if required; run the same ticket in VS Code agent/chat.
4. Score: correctness, review minutes, remaining quota/overage, intervention count.
5. Re-run the same ticket on your current default (Opus / Sol / prior Grok) as control.

Keep the winner only if review load does not rise enough to erase token savings.

When to stay on Claude, GPT, or autocomplete-only

What this means for PromptHive clusters

• Coding: multi-model editors (Cursor, Copilot) gain a cheaper flagship option; terminal-first Claude Code users should compare with eyes open, not brand loyalty.
• Agents: reinforces “measure cost × supervision,” the same thesis as our agents guide.
• Comparisons: keep Codex vs Claude Code as process comparison; Grok is another model that can sit under several UIs.

Final PromptHive verdict

Grok 4.5 is a serious coding-and-agents entry at aggressive API list rates, with distribution that matters — Cursor from day one and GitHub Copilot soon after. The Cursor training partnership explains why IDE users will see it immediately; it does not exempt you from reading diffs. If you already pay for a multi-model editor, run a controlled ticket suite this week. If your team is standardized on Claude Opus or GPT Sol with stable quality, treat Grok as a cost/latency experiment rather than a forced migration.

CTA: Pick five real tickets, run them on Grok 4.5 and your current default, and keep the model that minimizes review minutes × dollars. Start from how to use Cursor or how to use GitHub Copilot, then the complete AI coding guide.

Official sources

• Introducing Grok 4.5 (16 July 2026)
• Grok 4.5 in GitHub Copilot (28 July 2026)
• xAI API docs
• Cursor — SpaceX model training (linked from xAI’s launch post)
• GitHub Changelog — Grok 4.5 in Copilot

Further reading on PromptHive

• Cursor review · How to use Cursor AI
• GitHub Copilot review · How to use GitHub Copilot
• Complete AI coding guide
• Claude Opus 5: what changed
• GPT-5.6 Sol, Terra, Luna: what changed
• OpenAI Codex vs Claude Code
• Best AI coding assistants