Grok 4.5 Is Good: xAI Enters the Coding-Agent Race

Grok 4.5 is the first Grok release that made a lot of skeptical builders pause.

xAI's official claim is not small: Grok 4.5 is built for coding, agentic tasks, and knowledge work; it was trained alongside Cursor; it is available in Grok Build, Cursor, and the SpaceXAI API; and it is priced far below the most expensive frontier coding routes. The creator tests are loud too: Theo calls it unexpectedly good, Wes Roth runs visual/game demos, and Bijan Bowen throws it at browser OS, simulations, frontend, C++, and 3D model tests.

My read: Grok 4.5 is not automatically "the best model." But it is now serious enough that model routing has to include it.

Video credit: Theo - t3.gg. Additional embedded testing videos below are credited to Wes Roth and Bijan Bowen.

Source Note

This post uses xAI/SpaceXAI's official Grok 4.5 launch page, model docs, pricing docs, API page, and Grok Build page as the factual spine. The three YouTube videos and the supplied transcripts are treated as hands-on commentary. The X links are included as public examples and launch reactions, not as standalone proof.

I attempted to capture a screenshot from the official xAI launch page, but the public page blocked headless local capture through Cloudflare. Instead, the visual below is a JQ AI SYSTEMS chart built from the official xAI numbers and clearly labelled that way.

Link Map

What Is Confirmed

The official launch page says Grok 4.5 launched on July 8, 2026 and is SpaceXAI's strongest model for coding, agentic tasks, and knowledge work. It also says Grok 4.5 was trained alongside Cursor.

The model docs list the model ID as grok-4.5, with a 500k token context window, configurable reasoning, and image input support. The pricing docs list chat API pricing at $2 per 1M input tokens, $0.50 per 1M cached input tokens, and $6 per 1M output tokens.

Availability matters for us in Portugal: xAI's launch page says Grok 4.5 is not yet available in the EU in SpaceXAI products or the API console, with EU availability expected in mid-July. So this is worth tracking, but do not build a client promise around access until your account actually has it.

Official Numbers In Context

The xAI launch page reports strong coding benchmark results. The most useful numbers are not one single leaderboard row, but the pattern:

The token-efficiency claim is the part builders should care about most. xAI says Grok 4.5 resolves SWE Bench Pro tasks with 15,954 output tokens on average, versus 67,020 for Opus 4.8 max in the same comparison. If that pattern holds in real work, Grok 4.5 is not only cheaper per token; it can also need fewer tokens per completed task.

That is why this launch matters. Cost per completed task beats price per million tokens.

Creator Tests

The creator videos all point in the same direction: Grok 4.5 is not a joke model anymore. But they also show why demos are not enough.

Wes Roth's video focuses on vivid demos: a 3D sailing game, RPG-style tests, SVGs, and an "ultimate test" style comparison. That is useful because coding agents are now judged by end-to-end output, not just code snippets.

Bijan Bowen's test is useful because it includes rough edges: Browser OS, skydiving simulations, sprite-sheet games, Linux driver work, frontend design, C++, 3D model creation, and a subway FPS test. These broad tests show range, but they also remind us that agent performance is messy. A model can be strong and still fail in weird places.

Why It Matters

Three things changed with Grok 4.5:

1. xAI is now credible in coding-agent conversations. The official benchmark claims and creator tests put Grok 4.5 into the same routing table as GPT, Claude, GLM, and Fable.
2. Price pressure is real. At $2 input and $6 output per 1M tokens, the default question becomes: why spend the expensive route unless the task truly needs it?
3. Cursor matters. xAI says Grok 4.5 was trained alongside Cursor and is available in Cursor. That makes the model feel less like generic chat and more like a coding-agent product move.

This is the same pattern we have been tracking across Fable 5, GPT-5.6, GLM 5.2, and local models: the winning stack is not one model. It is routing.

How To Test Grok 4.5

Do not start with a toy benchmark. Start with a real but bounded task from your own work.

1. Pick one repo or project. Use a non-critical branch or throwaway copy.
2. Give it one clear task. Example: fix a bug, refactor one module, improve one UI, add one test suite, or build one prototype.
3. Run the same prompt against your current model. Compare Grok 4.5 against the route you already trust.
4. Measure completed work, not vibes. Count tests passed, files changed, regressions, review time, token cost, and how many follow-up prompts were needed.
5. Keep a fallback route. If Grok stalls, try Fable, Opus, GPT-5.5, GPT-5.6, GLM 5.2, or your current preferred model.

xAI's official quickstart uses the Responses API:

curl -s https://api.x.ai/v1/responses \
  -H "Authorization: Bearer $XAI_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "grok-4.5",
    "input": "Find and fix the bug, then explain it: function median(a){a.sort();return a[a.length/2]}"
  }'

The Grok Build page also shows a one-line CLI install:

curl -fsSL https://x.ai/cli/install.sh | bash

Again, EU caveat: if you are in Portugal, check whether your account is eligible before assuming the API or product route works today.

Where To Route It

Watch-Outs

• Benchmark framing matters. xAI's official page reports impressive numbers, but some rows still show Fable, GPT, or Opus ahead. Do not flatten that into "Grok beats everything."
• Availability is not global yet. The EU caveat is directly relevant for Portugal.
• Tool costs can creep. xAI prices Web Search, X Search, Code Execution, file search, and other tools separately from tokens.
• Priority processing is a premium route. The pricing docs say priority processing bills at 2x standard rates when served at that tier.
• Creator demos are useful but not conclusive. Games, SVGs, and one-prompt apps are good stress tests, but your business workflow needs its own eval.
• Cursor claims should stay precise. xAI says Grok 4.5 was trained alongside Cursor and is available in Cursor; do not add stronger corporate claims unless you have a verified source.

Sources

• xAI / SpaceXAI: Introducing Grok 4.5
• xAI docs: Models
• xAI docs: Pricing
• xAI API
• Grok Build Beta
• Grok plans and pricing
• Cursor
• Official SpaceXAI X post about Grok 4.5
• Theo - t3.gg: Oh no, the new Grok model is good
• Wes Roth: Grok 4.5 just cooked Claude and OpenAI
• Bijan Bowen: Grok 4.5 is insane, is this a GPT and Opus competitor?
• Public X example from @anshuc
• Public X example from Daniel Farina
• Public X example from @grok
• Public X example from GMI Cloud
• Public X example from cb_doge