Grok 4.5 Unveiled: Affordable AI with Cursor Integration

Grok 4.5 dropped on July 8, 2026, and the numbers deserve a real look before anyone skips past them. $2 per million input tokens, $6 per million output tokens, immediately available in Cursor across every subscription tier, in SpaceXAI's developer CLI called Grok Build, and through the SpaceXAI API console. Elon Musk called it an "Opus-class model" that is faster, more token-efficient, and cheaper than Anthropic's flagship. For a developer wiring a model into an editor or an agent loop, that is a real tailwind — say it before the "but."

The release lands on one of the busiest weeks for AI in 2026. Hours after the Grok announcement, OpenAI confirmed GPT-5.6 Sol, Terra, and Luna will become publicly available on Thursday, after weeks of restricted access at the request of the Trump administration. So any honest comparison has to be done in writing, not vibes. The Grok 4.5 launch at Memeburn is the cleanest place to start.

What Grok 4.5 actually ships with

SpaceXAI built Grok 4.5 as a general-purpose model. The target surfaces, per the launch announcement, are coding, agentic workflows, office productivity, and knowledge work. Three details matter for builders:

1. It scored first on Harvey's Legal Agent Benchmark — meaningful if you build agentic legal tooling, less so if you don't.
2. It was co-trained with Cursor, the AI coding startup SpaceX is acquiring for $60 billion in stock. That is not a partnership badge. The model was shaped on Cursor's telemetry, which is why the integration is deeper than "we added a dropdown".
3. It ships with Microsoft Office plugins for Word, PowerPoint, and Excel — an unusual move for a launch that's otherwise framed around developers, and a clear signal that SpaceXAI wants enterprise office workers as a second persona.

The general-purpose framing matters because most frontier launches push one audience. Grok 4.5 is explicitly straddling dev tooling and office workflows. That's a positioning choice, not a hedge.

How to actually use it today

Three surfaces, three config shapes. Pick the one that matches where you work.

Cursor, every subscription plan. Open Cursor → Settings → Models → select grok-4.5. No API key juggling if you already have a Cursor sub.

Grok Build CLI. SpaceXAI's developer CLI:

# Install
curl -fsSL  | bash

# Authenticate
grok auth login

# Run against the new model
grok run --model grok-4.5 \
  "Write a postgres migration adding a 'last_seen_at' index on users"

SpaceXAI API console. Standard HTTP:

curl  \
  -H "Authorization: Bearer $SPACEXAI_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "grok-4.5",
    "messages": [{"role": "user", "content": "Refactor this component to use a discriminated union for the variant prop."}],
    "temperature": 0.2
  }'

Cost on that call: $2 per million input tokens, $6 per million output tokens. GPT-5.6's published price wasn't in the launch coverage, so the dollar advantage is directional — but a 10× input-token gap to Anthropic's flagship is the part that matters for agent loops.

One operational caveat: Grok 4.5 is not available in the EU yet. SpaceXAI is targeting mid-July for European access, with regulatory compliance cited as the likely cause. If your team sits in Frankfurt, plan for a short delay.

The benchmarks — read them, don't quote the marketing

SpaceXAI published four scores at launch. None of them put Grok 4.5 at the top of any category. That's the part of the launch that's easy to skip and shouldn't be.

Read the two DeepSWE rows together. On each provider's tuned harness Grok 4.5 lands at 62%, comfortably ahead of Opus 4.8. On the standardized harness the score drops to 53%, behind Opus. That's a 9-point spread driven entirely by harness choice — and it's the kind of detail that gets erased by a one-line "X is faster than Y" tweet.

Terminal Bench 2.1 is where Grok 4.5 looks most competitive: 83.3% sits within rounding distance of GPT 5.5's 83.4% and Fable 5's 84.3%. SWE Bench Pro, per the launch article, lists Fable 5 at 80.4% and Opus 4.8 at 69.2%, but the source truncates Grok 4.5's score — that's the column to watch for an update.

So: cheaper than Anthropic, faster per Musk's claim, first on Harvey's legal benchmark, mid-pack on standardized coding evals. A real product, not a marketing prop — but not a category leader either.

The thing that doesn't change when the model does

Grok 4.5 lands a week before GPT-5.6, six months after Opus 4.8 was the new hotness, and probably six weeks before whatever ships next. The model rotates; the editing surface underneath it does not. Cursor shipped support across every plan, the CLI took a new flag, the API console took a new model id. None of those changes touched the components a builder actually ships.

This is the part worth naming explicitly: the durable layer between the model and the user is the UI itself. When the model changes — and it just did — the parts that don't churn are the visual primitives that have to look and behave the same on web, iOS, and Android. The same <Card>, the same <DataTable>, the same keyboard shortcuts, the same focus rings. That's the layer OTF is built to hold steady while the rest of the stack rotates underneath.

The practical move is to use Grok 4.5 today, route it through Cursor or the API, take the price win — and pin your output layer to something that doesn't get rewritten every time a launch lands on a Thursday. The model is the volatile part. The product surface is not.

What this enables

A few concrete places where $2/$6 changes the calculus:

• Long-running agent loops where input token volume is the dominant cost. A multi-step agent against a real codebase stops being a procurement conversation.
• Bulk refactor runs — pass the file tree as context, get a structured diff back, pay per token not per seat license.
• IDE integrations where the cost was previously a deployment conversation. Cursor shipping Grok 4.5 across every plan removes the procurement step entirely.
• Office automation via the Word/Excel/PowerPoint plugins. For teams that already live in those surfaces, the integration is the product.

The mid-July EU unblock is the gating event for any European team — track that date before wiring this into a production loop.

Closing thought

Grok 4.5 is not a category leader on standardized coding benchmarks, but it is faster and cheaper than Anthropic's flagship per the launch claim, first on Harvey's legal benchmark, and immediately available in three places builders already work. That's a real product worth wiring up this week. The model will rotate again before the quarter is out — what doesn't rotate is the surface you ship on top of it.