Grok 4.5 just launched. Here's what changed.
xAI shipped Grok 4.5 on July 8, 2026, and it's a different kind of release than the usual "our smartest model yet" announcement. Instead of chasing general chat quality, xAI trained this one on real Cursor developer session data and pointed it at coding and agentic work. If you're deciding whether it's worth your attention, or just want to try it without paying anything, here's the breakdown.
What's new in Grok 4.5
Grok 4.5 is a mixture-of-experts model trained on trillions of tokens, and the training data leans hard on real developer sessions from Cursor rather than generic web text. That shows up in the benchmarks: it ranks 4th out of 168 models on the Artificial Analysis Intelligence Index, and it posted the best agentic tool-use score on the entire board.
The efficiency numbers are the more interesting part. Grok 4.5 uses roughly 14,000 output tokens to complete an average Intelligence Index task, compared to 67,020 for Claude Opus 4.8. Since output tokens are what you get billed for, that gap shows up directly in your bill, not just on a benchmark chart.
Elon Musk called it "Opus-class, but faster, more token-efficient and lower cost," and put it roughly on par with Opus 4.7 in raw capability. Cursor's own team has reportedly made it their daily driver for coding work, which I'd weigh more than any leaderboard position.
Pricing: API and consumer plans
| Plan | Price | What you get |
|---|---|---|
| Grok free (grok.com / X) | $0 | Text chat only, roughly 10 messages per rolling 2-hour window |
| SuperGrok Lite | $10/mo | More usage, limited image/video |
| SuperGrok | $30/mo or $300/yr | Full Grok 4.5 access as it rolls out, image and video generation |
| Grok Heavy | $300/mo | Highest usage tier |
| API (per token) | $2/M input, $6/M output | Cached input drops to $0.50/M, a 75% discount |
The API also ships a 500K context window, generous for a model at this price point.
How to use it for free
- Go to grok.com or open Grok inside the X app. You don't need an X Premium subscription for basic access.
- Stick to text chat. Image and video generation moved behind a paywall in March 2026, so the free tier is text-only now.
- Watch your rolling window. The free cap isn't a daily reset, it's roughly 10 prompts per 2-hour rolling window. Send 10 messages at 3pm and you're clear again around 5pm, not at midnight.
- If you keep hitting the wall, SuperGrok Lite at $10/month is the cheapest way to get real headroom before jumping to the full $30 tier.
Is it worth using over Claude or GPT-5.6?
For coding, the token efficiency numbers are hard to argue with, and Cursor's team adopting it as a daily driver says more than a leaderboard ever could. If you're doing agentic dev work and care about cost per task, Grok 4.5 earns a real look next to [Claude](https://questloops.com/tools/claude) and OpenAI's [GPT-5.6 lineup](https://questloops.com/blog/openai-s-gpt-5-6-explained-sol-vs-terra-vs-luna-and-which-one-to-use).
Where it gets messier is trust. The loudest thread on Grok 4.5's Hacker News launch post wasn't about capability, it was about Musk allegedly nudging model outputs on political questions. One commenter pushed back, arguing Grok has actually been more restrained than GPT or Gemini in their own testing. Both readings sit in the same comment section, so this looks like an open argument rather than a settled one. If that concern matters to you, weigh it alongside the price.
For general research and assistant work rather than coding, it's worth comparing against options that don't carry that baggage, like our breakdown of [ChatGPT Plus vs Claude vs Perplexity](https://questloops.com/blog/chatgpt-plus-vs-claude-vs-perplexity).
The takeaway
Grok 4.5 is a strong, cheap coding model with real efficiency gains, not a benchmark bump dressed up as one. The free tier works fine for light text chat, but if you're doing serious agentic coding, SuperGrok Lite at $10/month is the realistic entry point. Go in aware of the trust debate around it too, since that's a factor no pricing table settles for you.
