📰 AI News: SpaceXAI Launches Grok 4.5, Its First Model Since Going Public and Buying Cursor 📰

📝 TL;DR 📝

SpaceXAI released Grok 4.5 on July 8, its first model since going public and acquiring the AI coding tool Cursor for $60 billion. Elon Musk is calling it "Opus-class," roughly comparable to Anthropic's Opus 4.7, at a fraction of the price: $2 per million input tokens and $6 per million output, versus Opus 4.8's $5 and $25. Independent testing backs the cost story more than the capability claim: Grok 4.5 lands fourth on a real-world agentic benchmark, behind the newest Claude models, but uses roughly four times fewer output tokens on coding tasks, which meaningfully changes total cost per completed job.

🧠 Overview 🧠

This is SpaceXAI's first release since two major structural changes: going public, and acquiring Cursor, one of the most widely used AI coding tools, in a deal reportedly valued at $60 billion. Grok 4.5 was trained jointly with the Cursor team, and the positioning reflects that lineage directly. This is not being marketed as a general chatbot upgrade. It is aimed squarely at coding, agentic workflows, and knowledge work, the same territory Anthropic's Opus and Claude Code, and OpenAI's newly launched GPT-5.6 family, are competing in.

The launch also landed on an unusually crowded day for frontier AI. OpenAI rolled out GPT-5.6 publicly the same week, after a staggered release tied to government review. For the first time in months, all three major US AI labs, OpenAI, Anthropic, and SpaceXAI, have current flagship-tier models publicly available at roughly the same moment.

📜 The Announcement 📜

Musk introduced Grok 4.5 directly on X: "It is an Opus-class model, but faster, more token-efficient and lower cost," later clarifying, "Our internal assessment is that Grok 4.5 is roughly comparable to Opus 4.7, but much faster." That comparison is notable for what it does and does not claim: Opus 4.7 is a prior Claude generation, not Anthropic's current frontier model. SpaceXAI's own published benchmark charts show Grok 4.5 beating Opus 4.8 on two of four coding benchmarks and losing on the other two, with Anthropic's Fable 5 leading most of those charts outright. The company has been fairly candid about this positioning: it says Grok 4.5 beats some rival models on speed, price, and performance, but not the largest, latest models from those same competitors.

Independent verification from Artificial Analysis places Grok 4.5 fourth on its GDPval index for real-world agentic knowledge work, behind the newest Claude releases, with an Elo rating of 1543. But the more interesting number is token consumption. On one software engineering benchmark, Grok 4.5 completed tasks using an average of roughly 15,900 output tokens, compared to about 67,000 for Opus 4.8, a gap of more than four times. Artificial Analysis calculated the effective cost at roughly $0.49 per completed task, and described the model as sitting clearly on the frontier for performance relative to cost, even though it is not the most capable model available on raw benchmarks alone.

Grok 4.5 is live now in Grok Build, inside Cursor on all plans, and through the SpaceXAI console and API. Public rollout in the Grok app began July 9. It is not yet available in the EU, with availability there expected by mid-July.

⚙️ How It Works ⚙️

Trained with Cursor - Grok 4.5 was developed jointly with the Cursor team following SpaceXAI's acquisition of the company, with Cursor's real-world engineering interaction data feeding directly into how the model handles code writing, review, and debugging.

Token efficiency as the core pitch - Rather than leading with raw capability, SpaceXAI's central argument is that Grok 4.5 completes comparable tasks using significantly fewer output tokens than rivals like Opus 4.8, which compounds with its lower per-token price to produce a much lower total cost per completed task.

Pricing - $2 per million input tokens and $6 per million output tokens, undercutting Opus 4.8's $5 and $25, and landing close to OpenAI's cheapest GPT-5.6 tier, Luna, at $1 input and $6 output.

Availability across surfaces - Accessible today through the Grok Build agent, inside Cursor across all subscription plans, and via the SpaceXAI developer console and API. The consumer Grok app began its public rollout July 9.

Shared compute with competitors - Grok 4.5 was trained using compute capacity that SpaceXAI also leases to competitors including Anthropic and Google, an unusual dynamic that may force future capacity trade-offs between running its own models and continuing to lease compute as a revenue source.

💡 Why This Matters 💡

This is a genuine cost-efficiency story, not a capability-leadership story, and SpaceXAI is not hiding that - The company's own benchmark charts openly show Grok 4.5 trailing the newest models from Anthropic and OpenAI. The pitch is specifically about cost per completed task, and the token-efficiency data from independent testing supports that framing more than it supports any claim of outright superiority.

Token efficiency matters as much as sticker price for anyone running agentic workflows at volume - A model that costs less per token and uses meaningfully fewer tokens per task compounds savings in a way that a simple price comparison misses. This is a genuinely useful number to understand if you are running high-volume coding or agent tasks anywhere.

The Cursor integration is the most immediately relevant detail for developers - Grok 4.5 being available inside Cursor on all plans gives anyone already using that tool a direct, no-friction way to test it against whatever model they are currently using, without switching platforms.

This is the clearest three-way frontier competition moment in a while - With GPT-5.6, Fable 5's ongoing saga, and Grok 4.5 all live within the same window, this is a useful moment to actually compare real cost and performance across providers rather than relying on any single company's launch messaging.

🏢 What This Means for Businesses 🏢

If you already use Cursor, this is a low-friction test - Switching the model in Cursor's settings costs nothing but a few minutes. Running a real coding or agent task through Grok 4.5 alongside whatever you currently use is the fastest way to form your own opinion rather than relying on any vendor's claims.

Evaluate cost per completed task, not just cost per token - The token-efficiency data suggests Grok 4.5's real advantage may be in how few tokens it needs to finish a job, not just its listed price. If you are comparing models for a high-volume workflow, measure total cost to complete your actual task, not just the headline per-token rate.

Do not expect frontier-leading capability - Independent benchmarks place Grok 4.5 behind the newest Claude models on real-world agentic work. If your task genuinely needs top-tier reasoning or the largest context window, this is not positioned to be that model. If your task is more routine and cost-sensitive, it is worth testing.

Not yet available in the EU - If your business operates there, factor in the mid-July timeline before planning around this model specifically.

No model card has been published yet - Independent observers, including Ethan Mollick, flagged the absence of a model card at launch as a transparency gap worth noting. Treat capability claims with appropriate caution until more independent, structured evaluation is available.

🔚 The Bottom Line 🔚

Grok 4.5 is a legitimate, well-positioned release, but it is worth being precise about what it actually claims. This is not a frontier-capability launch. It is a cost-and-efficiency play, aimed squarely at developers and businesses running coding and agentic workloads at volume, where token efficiency and price compound significantly over time. The company's own data is candid that it is not beating the newest models from Anthropic or OpenAI outright, and independent testing backs that framing.

For our audience, the most realistic takeaway is this: if you are already in Cursor, Grok 4.5 is a free, low-effort model to test against what you currently use, particularly for cost-sensitive, high-volume tasks. If you need the most capable model available for a genuinely hard problem, this is not positioned to be that model, and the available evidence agrees with SpaceXAI's own framing on that point.

💬 Your Take 💬

If a model costs a fraction of the price and uses far fewer tokens per task, but trails the top models on raw capability, is that trade-off worth it for your workflow, or do you stick with the most capable option regardless of cost? 🤔

1 comment

📰 AI News: SpaceXAI Launches Grok 4.5, Its First Model Since Going Public and Buying Cursor 📰