Product Launch · XiaoHu Explains

xAI Launches Grok 4.5: Rivals Opus 4.8, ~2x Faster, Quarter the Price

Ranks 3rd on DeepSWE, 4th overall — faster than Opus 4.8 and notably cheaper.
60-Second Rundown
  • xAI launches Grok 4.5: official and Musk's own framing put it near "Opus-class," but the pitch is speed and cost; co-trained with Cursor, positioned for coding, agentic work, and knowledge tasks.
  • DeepSWE 1.0 score of 62.0% ranks third; SWE Marathon 29% ranks first; third-party Artificial Analysis puts it 4th overall (score 54), with a Coding Agent Index of 76 tying GPT-5.5.
  • Output speed ~80 TPS, roughly 2× faster than Opus 4.8; output pricing $6/million tokens, about a quarter of Opus 4.8's $25; AA measures per-task cost at ~$0.31, though hallucination rate rose to 54%.
  • Priced at $2/million input tokens and $6/million output tokens; already the default model in Grok Build, rolled out across all Cursor plans, plus official add-ins for Word, PowerPoint, and Excel.
  • Trained on tens of thousands of NVIDIA GB300 GPUs, with reinforcement learning spanning hundreds of thousands of tasks centered on software engineering; not yet available in the EU, expected mid-July.
Stance note: this piece interprets xAI's official announcement. The benchmark scores, 80 TPS speed claim, ~2× token efficiency, and pricing cited here all come from xAI's own figures — some are self-reported or compared against self-selected competitors, and have not been independently reproduced by a third party.
1 Launch

xAI Launches a New Flagship: Grok 4.5

On July 8, 2026, xAI launched Grok 4.5, positioning it as their strongest model yet, built for coding, agentic tasks, and knowledge work — trained in partnership with the code editor Cursor.

Grok 4.5 packs flagship-level reasoning into the speed and cost of a fast model: running at 80 TPS, with official claims of roughly half the token consumption of mainstream flagship models on comparable tasks, priced at $2/million input tokens and $6/million output tokens.
Why it matters: flagship-level reasoning, fast-model speed, and low pricing — stacked onto one model, officially claimed to deliver the most intelligence per unit of time and cost. Let's start with a live demo.
2 Live Demo

One Sentence, One Spinning Solar System

Before the benchmarks, look at what it can actually build. In one official example, a single prompt was enough for Grok 4.5 to generate an end-to-end, fully working three.js solar system simulator — adjustable speed, realistic orbits, a starfield, and a styled HUD. Officials emphasize that even with a bare-bones prompt, the output is a complete, ready-to-use application.

Input · Original Official Prompt
Make a beautiful simulation of the universe and solar system. should be sped up with adjustable time, realistic motion, orbits, stars. use threejs. Make the HUD well styled and conform to modern design principles.
Make a beautiful simulation of the universe and solar system, with adjustable, sped-up time, realistic motion, orbits, and stars, built with threejs; style the HUD to conform to modern design principles.
Grok 4.5's solar system simulation demo (source: live footage posted by @DogeDesigner on X). The official launch page has a similar interactive demo too — see the Training section of the original post.

One natural-language instruction, all the way to a complete, interactive, real-time-running front-end app. Beyond demos like this, officials also note it can handle challenging Rust and C / C++ tasks — end to end, from prompt to finished product.

2b Third-party test

Same challenge: one prompt + one reference image to rebuild a 3D globe dashboard

Developer @hqmank ran the same challenge he previously used on Fable 5: one prompt plus one reference image, task = rebuild a Three.js 3D globe dashboard.

His take: lighting, glass panels, depth, and spacing all matched; the Three.js scene rendered correctly on the first try; frontend quality was better than expected. On this task he ranks Grok 4.5 above Opus 4.8, with only Fable 5 still ahead.

Full demo from @hqmank. Source: original post on X. Single-task, single-tester comparison — not an official leaderboard.
3 Benchmarks

Where It Actually Lands on Coding Tests

Here's the DeepSWE 1.0 score officials released. DeepSWE 1.0 is a benchmark for how well an AI handles real-world software engineering tasks, scored on pass@1 — solving the task correctly in a single attempt (like an exam with no do-overs). On this leaderboard, Grok 4.5 scores 62.0%, ranking third.

Coding benchmark comparison table for Grok 4.5 vs. Opus 4.8, GPT-5.5, Composer 2.5, and Fable 5
A common official/ecosystem comparison table: Terminal-Bench 2.1, SWE-Bench Multilingual, DeepSWE 1.0, SWE-Bench Pro. The Grok 4.5 column is highlighted; some competitors are labeled with high/max/xhigh tiers. SWE-Bench Multilingual is a supplementary comparison beyond the launch page's five main charts.
0%20%40%60%DeepSWE score (pass@1)66.1%Fable max64.31%GPT 5.5 xhigh62.0%Grok 4.555.75%Opus 4.8 max40.12%Opus 4.7 maxEval created by Datacurve, run with each model provider's harnesses by AA
0%20%40%60%DeepSWE score (pass@1)70.0%Fable max67.0%GPT 5.5 xhigh59.0%Opus 4.8 max53.0%Grok 4.544.0%GLM 5.2mini-swe-agent harness run by Datacurve
0%20%Resolution rate (pass@1)29.0%Grok 4.526.0%Opus 4.8 max24.0%Fable max16.0%Opus 4.7 max
0%20%40%60%80%Terminal Bench 2.1 score84.3%Fable max83.4%GPT 5.5 xhigh83.3%Grok 4.578.9%Opus 4.8 max78.9%Opus 4.7 max
0%20%40%60%80%SWE Bench Pro resolve rate80.4%Fable max69.2%Opus 4.8 max64.7%Grok 4.564.3%Opus 4.7 max62.1%GLM 5.258.6%GPT 5.5 xhigh
The chart above is taken directly from xAI's official launch-page SVGs (switch between 5 charts). Benchmarks were authored by Datacurve; competitor scores are pulled from each vendor's public system cards or leaderboards.

The same launch page lists several other coding-related benchmarks. Different scope, different rankings — some first place, some further down:

SWE Marathon · Resolution rate pass@1
29.0%
Ranks 1st (Opus 4.8 max 26.0% · Fable max 24.0%)
Terminal Bench 2.1 · Terminal Agent Tasks
83.3%
Close behind the top tier (Fable 84.3% · GPT 5.5 83.4%)
SWE-Bench Pro · Resolve Rate
64.7%
Behind Fable max 80.4%, Opus 4.8 max 69.2%
DeepSWE 1.1 · mini-swe-agent
53%
Further behind than 1.0 (Fable 70% · GPT 5.5 67% · Opus 59%)
All figures above come from charts xAI itself listed on the launch page. DeepSWE 1.0 was run with each vendor's own harness, while 1.1 uses mini-swe-agent uniformly — switch the framework on the same task family, and rankings shift.
Third Party · Artificial Analysis

Independent evaluator Artificial Analysis published composite scores right after launch (not from xAI's own charts):

  • Intelligence Index score of 54, 4th overall — behind only Fable 5, GPT-5.5, and Opus 4.8; up 16 points from the previous-generation Grok 4.3.
  • Coding Agent Index (run in Grok Build) score of 76, tying GPT-5.5 in Codex, slightly behind Fable 5 in Claude Code.
  • Per-task cost: roughly $0.31/task for the Intelligence Index; roughly $2.49–2.59/task for the Coding Agent Index (versus ~$11.80 for Fable 5 / Claude Code, ~$5.07 for GPT-5.5 / Codex).
  • Average total tokens per Coding Agent task is about 1.9M, significantly less than Fable 5 (7.2M) and GPT-5.5 (6.2M).
  • The weak spot is noted too: AA-Omniscience accuracy rose from 35% to 52%, but its hallucination rate rose from 25% to 54% — it knows more, and is more confident when it's wrong.

Third-party and surrounding coverage add some more context: Cursor says it co-trained Grok 4.5 with xAI, calling it their "first model that isn't only for software engineering." Some reports, citing Artificial Analysis and Musk-related disclosures, put Grok 4.5's parameter count at roughly 1.5 trillion — about 3× Grok 4.3 (the official launch page itself doesn't state a parameter count). Musk's public framing is "Opus-class, but faster, leaner, and cheaper." SpaceX's roughly $60 billion stock acquisition of Cursor back in June has been widely cited by outlets as the industry backdrop for this joint training effort.

4 Fast and Cheap

Why It's Both Fast and Cheap

Grok 4.5's differentiation is concentrated on speed and cost. It didn't top the benchmarks, but it puts "flagship-level brains" together with "fast-model speed and low price."

Fast and Cheap

It outputs at 80 TPS, reaching what the industry calls "fast model" territory; officials claim that on comparable tasks, its token efficiency is roughly double that of the latest flagship models — solving the same problem in under half the steps and tokens.

Official token efficiency comparison chart: Grok 4.5 ~15,954 tokens vs. Opus 4.8 max ~67,020
Token efficiency chart from the official launch page: average output tokens per SWE-Bench Pro task — Grok 4.5 at roughly 15,954, Opus 4.8 max at roughly 67,020, about a 4.2× gap.
80 TPS
Output speed reaches "fast model" territory. TPS is tokens generated per second — higher means faster output.
Opus 4.8 max
SWE-Bench Pro
67,020
Grok 4.5
15,954
Scope: average output tokens to complete comparable tasks on SWE-Bench Pro. About a 4.2× gap. The official text elsewhere summarizes it as "roughly 2× token efficiency / half the steps" — the two framings use different scopes; treat the figure caption as authoritative.

For agentic use cases, this means sustaining longer autonomous runs and more back-and-forth within the same token budget and time.

5 How It Was Trained

What the Training Actually Involved

How was this capability built? Grok 4.5 was trained on tens of thousands of NVIDIA GB300s (NVIDIA's latest-generation high-performance GPU chip for large-scale AI training). Beyond just piling on compute, xAI says a big chunk of the effort went into curating and cleaning data: deduplication, quality scoring, and domain-based selection, so the training data is both broad in coverage and dense in signal.

The reinforcement learning stage covers hundreds of thousands of tasks, centered on multi-step software engineering, judged with a mix of automated scoring and model scoring. Their training architecture is highly asynchronous: a single agent can run autonomously for hours at a stretch, while tens of thousands of GPUs keep training in parallel.

Massive Raw Data
Dedup · Quality Scoring
Domain Filtering
Tens of Thousands of GB300s
Pretraining
Hundreds of Thousands of Engineering Tasks
RL (Async Rollout)
Grok 4.5
6 Office Automation

From Writing Code to Writing Excel, PPT, and Word

Grok 4.5's abilities aren't limited to coding. It's now the default model in Grok Build (xAI's command-line coding tool), and that same skill set now extends into the Excel, PowerPoint, and Word trio.

Office Automation

Each of the three has its own specific approach: Excel can research online while building multi-sheet models, and leaves notes for later reference; PowerPoint uses native shapes to draw complex diagrams and lay out clear slide content; Word writes well-organized, formal copy.

📊

Excel

  • Researches online, then builds models
  • Uses formulas across multiple sheets
  • Leaves notes / comments for later reference
📽️

PowerPoint

  • Draws complex diagrams with native shapes
  • Designs clear, intuitive slide content
  • Builds out a full structure from a prompt
📄

Word

  • Writes clearly organized, formal copy
  • Handles official/formal written expression
Example PowerPoint Prompt
Outline a 5-slide quarterly business review
Outline a 5-slide quarterly business review

Word, PowerPoint, and Excel all now have official add-ins available for install through the Microsoft Marketplace.

7 Pricing and Access

What It Costs, and Where to Use It Now

Pricing and access: Grok 4.5 charges $2/million input tokens and $6/million output tokens. Layered on top of the token efficiency mentioned above, the actual bill for comparable coding tasks comes down even further.

$2
Per million input tokens
$6
Per million output tokens
4.2×
Fewer output tokens on SWE-Bench Pro
80 TPS
Output speed

Here's how it compares to peer flagships' public pricing (per million tokens):

ModelInputOutput
Grok 4.5$2$6
Opus 4.8$5$25
GPT-5.5 / 5.6$5$30
Fable 5$10$50
Competitor pricing sourced from each vendor's public price lists and media roundups (The Decoder, etc.); actual bills also depend on caching discounts and effort-tier settings.

On specs, Artificial Analysis recorded a current context window of 500K tokens (down a tier from Grok 4.3's 1 million); Musk later said on X that it would "probably be back to 1 million next week or so." Modality is text + image input, text output.

Where you can use it now: Grok Build, all Cursor plans, and the xAI console (console.x.ai) are all live — grab an API key and you're wired up in a few lines of code. OpenRouter, Hermes Agent, and others also integrated it on launch day.

Official API Call Example (curl)
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]}"
  }'

Artificial Analysis also noted a few more specs: cache hits run about $0.5/million tokens (roughly a 75% discount off list price); pricing doubles once input exceeds about 200K tokens; the current context window is 500K tokens (down a tier from Grok 4.3's 1 million, with Musk saying it'll probably be back to 1 million next week or so); vision input and configurable reasoning are retained. Cursor says usage doubled in the first week, and stresses that Grok 4.5 and its own Composer line are models at different scales — Composer 2.5 is staying.

⚠️ It's currently unavailable in the EU across any of xAI's products or APIs; officials expect a mid-July rollout there. Separately, Grok 4.5 is free to use in Grok Build and Cursor for a limited time.

8 What It Means

What This Update Actually Delivers

Breaking this down by who benefits, here's roughly what's on the table right now.

  • Independent Developers

    A single-sentence prompt produces a complete, working app directly — the path from idea to demoable prototype gets faster, and you can switch to it directly inside Grok Build or Cursor while coding.

  • Office Knowledge Workers

    One more agent that can automatically build models, draw diagrams, and write copy: Word / PowerPoint / Excel add-ins are live, capable of multi-sheet Excel modeling that needs online research, PowerPoint diagramming, and formal document writing.

  • API Developers

    Pricing of $2/$6 per million tokens, combined with roughly 2× token efficiency, means running more — and longer — autonomous agentic tasks on the same budget.

Overall, Grok 4.5 delivers the highest intelligence per unit of time and cost. From xAI's official "Introducing Grok 4.5"
This piece is based on xAI's official launch page, "Introducing Grok 4.5" (July 8, 2026, x.ai/news/grok-4-5). The coding benchmarks, 80 TPS figure, SWE-Bench Pro token counts, and $2/$6 pricing shown in the official charts come from that page; Artificial Analysis's Intelligence / Coding Agent indices, per-task cost, and hallucination rate come from AA's public release (cross-checked against secondary reporting from The Decoder, Latent Space, etc.); competitor public pricing is compiled from each vendor's price lists and media coverage. The benchmark bar charts and token-efficiency chart are taken from the official launch page's SVGs/screenshots; the solar-system demo is live footage of Grok 4.5 posted by @DogeDesigner on X; see the official post for the full interactive demo.