Grok 4.5: What's Actually Confirmed So Far

Grok 4.5 is now live. As of July 9, 2026 it's available on the xAI API, in Cursor (all plans), and in Grok Build — model ID grok-4.5, a 500K-token context window, $2 per million input tokens and $6 per million output ($0.50/M cached), configurable reasoning, and roughly 80 tokens/sec output. xAI calls it its "most intelligent and fastest model," positioned for code and agentic tool use. The headline caveat: those are xAI's own launch numbers, and its benchmark claims (64.7% SWE-Bench Pro, 4.2× fewer output tokens than Opus 4.8) are xAI-reported, not independently verified. The credibility-ranked timeline below shows how we got from June 28 private beta to a live model in under two weeks — including xAI's own @grok account denying a "tomorrow" release the day before Musk announced it.

The short version:

Where Grok 4.5 fits in the lineage

(One more thing: the "GROK 4.5" crypto token circulating online has no connection to xAI's model. The only official channels are x.ai, docs.x.ai, and xAI's own X account.)

Architecture: Musk confirmed 1.5T V9 + Cursor data, not much else

Musk's own post confirms the core: Grok 4.5 runs on a foundation model codenamed V9, roughly 1.5 trillion parameters, with Cursor coding data mixed into supplemental training. That's the most reliable piece of this whole story. Everything past that is unconfirmed — the community estimates 1.5T at roughly three times the earlier V8-small architecture, but whether the Cursor data means IDE usage traces, agentic coding logs, or something else — along with licensing and volume — hasn't been disclosed. The theory holds together: Cursor is one of the busiest agentic coding surfaces around, and training directly on that data could plausibly help with long-horizon coding tasks. But xAI has only published API specs (context, pricing, rate limits) — no training-data disclosure or detailed model card, so the Cursor-data specifics remain undisclosed.

How credible is "close to or beyond Opus"?

This was the most-quoted line before launch, and it's worth taking apart. Pre-launch there was no benchmark — just a relayed quote. The July 9 launch changed that, partly: xAI's release cites 64.7% on SWE-Bench Pro and 83.3% on Terminal Bench, and claims 4.2× fewer output tokens than Claude Opus 4.8 on SWE-Bench Pro (launch figures shared on X). Those are xAI-reported numbers, not independent evals (no third-party confirmation yet), and they're coding-focused — they say little about general reasoning or writing. The comparison target also stays fuzzy: Anthropic's current flagship is Claude Opus 4.8. So "Opus-level, faster, cheaper" is now backed by xAI's own coding benchmarks, but treat the broader claim as provisional until independent results land.

Where Grok 5 fits in

The Grok 5 rumor (6T/10T, timeline above) comes from a separate Kretschmann post a week earlier, also unconfirmed. Two details the table doesn't cover: it's reportedly a Mixture-of-Experts design with a context window around 1.5M tokens. Read together, the 1.5T Grok 4.5 is the current flagship (Opus-class on xAI's coding benchmarks), with an even larger generational leap — 6T/10T — rumored for Grok 5.

Grok 4.5 is live — the confirmed specs (and where 4.3 still wins)

Grok 4.5 is now the flagship on the xAI API. Checked against docs.x.ai on July 9, 2026: model ID grok-4.5, a 500K-token context window, $2 per million input tokens and $6 per million output ($0.50/M cached), and configurable reasoning. xAI's @grok account adds that rate limits reach up to 150 requests/second and 50M tokens/minute (tier-dependent) with roughly 80 tokens/sec output.

Here's the counterintuitive part: Grok 4.5 costs more and has half the context of Grok 4.3. Grok 4.3 still lists a 1M-token window at $1.25/$2.50 per million input/output — cheaper per token and double the context. So the split is real: pick Grok 4.5 for hard coding and agentic reasoning where its launch benchmarks (SWE-Bench Pro 64.7%) matter; pick Grok 4.3 for long documents, large context, or cost-sensitive volume.

Grok 4.5 in Cursor: how to turn it on, and what it costs

Grok 4.5 shipped inside Cursor on launch day via xAI's partnership, and it's the use case most people are searching for. It's on every Cursor plan, including the free Hobby tier — update Cursor to the latest version, then Settings → Models → Grok 4.5 (per xAI's @grok account).

What you get depends on the plan: the free Hobby plan includes Grok 4.5 with a shared usage pool (xAI says limits are doubled during launch week), while Pro and above (~$20/mo, with launch promos) get bigger weekly pools and full agent/Auto-mode access. The catch: those limits draw from Cursor's partnership pool, not a dedicated xAI quota, so heavy free-tier users hit the ceiling fast.

Don't want to pay Cursor twice? Bring your own xAI API key — grab one at console.x.ai, paste it into Cursor's Settings → Models → API Keys (OpenAI-compatible field), point the Base URL at xAI's endpoint, and add grok-4.5 manually (walkthrough). You then pay xAI's $2/$6 per-million rate directly, which makes sense if you're already on the API or want to skip the subscription.

What to actually do right now

Two things now that 4.5 is live: first, the model ID is grok-4.5 — it's a drop-in swap on the same api.x.ai endpoint if you're already on Grok 4.3, but budget for the higher price ($2/$6 vs $1.25/$2.50) and the smaller 500K context before you migrate. Second, keep the model ID in config, not hardcoded, so you can route long or cost-sensitive work to 4.3 and hard coding to 4.5 without a code change.

Note: if what you actually need is image or video generation rather than text, Grok Imagine 1.5 is already public, and AIReiter has it integrated — but it's a completely different product line from Grok 4.5 and won't substitute for text reasoning or code generation.

FAQ

Is Grok 4.5 available now? Is there an API or waitlist?

Yes — as of July 9, 2026 it's live on the xAI API (model grok-4.5), in Cursor (all plans), and in Grok Build. There's no waitlist for the API; you call it the same way as Grok 4.3 on api.x.ai. (For context: xAI's own @grok account denied a "tomorrow" release on July 7, then Musk announced it July 8 and it shipped July 9.)

Is Grok 4.5 free? What will it cost?

$2 per million input tokens, $6 per million output, $0.50/M cached — per docs.x.ai and the @grok account. That's pricier than Grok 4.3 ($1.25/$2.50) and roughly competitive with Claude Opus 4.8's tier; Musk's "cheaper than Opus" claim holds at the output-token level.

Is Grok 4.5 free in Cursor?

Yes — it's on Cursor's free Hobby plan with a shared usage pool (limits doubled during launch week). Pro and higher (~$20/mo) unlock bigger pools and full agent features, or skip the subscription entirely with your own xAI API key.

Is Grok 4.5 actually better than Claude Opus?

On coding, xAI says yes — 64.7% on SWE-Bench Pro and 4.2× fewer output tokens than Opus 4.8. Those are xAI-reported, not independently verified, and coding-only, so the broader "beats Opus" claim is provisional. Wait for third-party evals before calling it settled.

How does Grok 4.5 compare to GPT-5.5 or Gemini 3.1?

Still mostly unanswerable. xAI published a coding benchmark (SWE-Bench Pro 64.7%) but no broad head-to-head against GPT-5.5 or Gemini 3.1, and nothing independent. Treat any three-way comparison as provisional until third-party evals land. For the Claude matchup specifically, see our Grok 4.5 vs Claude Sonnet 5 comparison.

Is Grok 5 the same thing as Grok 4.5?

No. Grok 4.5 is the now-live flagship; Grok 5 (rumored 6T/10T) is a much larger follow-up with no official date — community estimates put it late 2026 or Q1 2027.

Is the "GROK 4.5" token I saw online official?

No. It's a cryptocurrency token unrelated to xAI's model. Official information only comes from x.ai, docs.x.ai, and xAI's own X account.

Can I use Grok 4.5 through AIReiter?

No — AIReiter doesn't currently offer Grok 4.5 (text). What AIReiter does carry from xAI is Grok Imagine 1.5, the image/video product line, which is separate from the Grok 4.5 text model.