Grok 4.5, Read Critically: What xAI's Launch Actually Shipped

xAI released Grok 4.5 on roughly July 8, 2026 (public availability July 9), with a launch page titled "Introducing Grok 4.5" and same-day coverage from TechCrunch and Axios. As of this writing the model is about a day old, which means most of the "Grok 4.5 destroys everything" takes you've seen are running on xAI's own numbers with no independent replication yet.

This piece is the series hub. The job here is not to hype the launch or dunk on it. It's to draw a hard line between the handful of things that have been independently verified and the larger pile of things that are vendor-claimed and still pending external validation — and then to translate the verified part into decisions a skill-builder or agent-builder can actually make this week.

If you ship agents on top of frontier models, a new low-cost frontier-class option is a real event. It changes routing math, cost ceilings, and which model you reach for on high-volume loops. But only if the claims hold. So let's sort the claims first.

Key Takeaways

• One number is independently anchored: Artificial Analysis puts Grok 4.5 at an Intelligence Index of 54, rank #4 of ~170 models — behind only Claude Fable 5, GPT-5.5, and Claude Opus 4.8. Lean on that; treat the rest as claims.
• The price is the story. API pricing is $2.00 / 1M input, $6.00 / 1M output with a 500K-token context — a headline >60% cheaper than Opus 4.8 and GPT-5.5 per Artificial Analysis.
• The first-party benchmark scorecard is mixed, not a sweep. Against Opus 4.8, xAI's own page shows Grok 4.5 winning 2 of 4 coding benchmarks. Claude Fable 5 leads all four.
• Elon Musk framed it as "Opus-class, but faster, more token-efficient, and lower cost" — attribute that as xAI's framing, not a measured fact.
• Grok 5 is not out. It's roadmap / in-training only. Anyone presenting it as available is wrong.
• The drop-in story is real: the API is OpenAI- and Anthropic-SDK compatible, so adding it to a multi-model stack is a base-URL-plus-model-string change, not a rewrite.

What's independently verified

The credible third-party anchor is Artificial Analysis's Intelligence Index, which measures models on a common battery rather than a vendor's chosen harness. Their read on Grok 4.5:

Two things stand out. First, #4 of ~170 is a strong, real result — this is a frontier-adjacent model, not a pretender. Second, it is explicitly behind three models, one of which (Opus 4.8) is the exact model Musk name-checked as the bar. So the honest framing is "just under the frontier, at a fraction of the price," not "new king."

The ~17s time-to-first-token is worth flagging for anyone building latency-sensitive UX. Heavy reasoning models think before they speak; if your skill shows a user a spinner, budget for it. The ~91 tokens/sec output rate is healthy once generation starts.

What's only xAI-claimed (for now)

Everything on the launch page is first-party. Treat it as a hypothesis until someone reproduces it. The headline claims:

• Coding-agent parity at lower cost. xAI says Grok 4.5 scores on par with GPT-5.5/Codex on a Coding-Agent Index, measured inside xAI's own "Grok Build" harness, at much lower cost. On-par-at-lower-cost is plausible and consistent with the independent price data — but "measured in our own harness" is exactly the phrase that should slow you down.
• The head-to-head vs Opus 4.8. On xAI's page, Grok 4.5 wins DeepSWE 1.0 (62.0% vs 55.75%) and Terminal-Bench 2.1 (83.3% vs 78.9%), and loses DeepSWE 1.1 (53% vs 59%) and SWE-Bench Pro (64.7% vs 69.2%). That's a 2-of-4 split, and Claude Fable 5 leads on all four. I unpack that scorecard in Grok 4.5 vs Claude Opus 4.8: Reading Vendor Benchmarks Without the Hype.
• The token-efficiency claim. xAI reports Grok 4.5 averaging ~15,954 output tokens vs Opus 4.8's ~67,020 on SWE-Bench Pro — about 4.2× fewer. If it holds, that's a bigger deal for your bill than the raw benchmark score. The economics get their own piece: The Token-Efficiency Play.

None of these are disqualifying. First-party benchmarks are marketing by construction — the question is whether the vendor picked a fair fight and reported it straight. For a repeatable way to interrogate any launch scorecard, see First-Party Benchmarks Are Marketing: A Skeptic's Checklist.

The price is the actual news

Strip away the leaderboard drama and the durable fact is the price. At $2 in / $6 out per million tokens, Grok 4.5 undercuts Claude Opus 4.7's $5 in / $25 out dramatically — the output side alone is roughly 4× cheaper — and Artificial Analysis pegs it >60% below Opus 4.8 and GPT-5.5. Combine that with a top-5 Intelligence Index score and you get a strong price-per-intelligence ratio, which is the metric that actually governs a high-volume agent's monthly bill.

That's the ratio worth optimizing across a whole stack. If a task needs frontier-grade reasoning you keep Opus 4.8 or Fable 5; if it needs 80% of that at a third of the cost, Grok 4.5 becomes a candidate. Turning that into an explicit routing policy is the subject of A Model-Router Skill: Grok, Opus, Fable, and GPT by Price-per-Intelligence and Grok 4.5's Price-per-Intelligence for the Agent Budget.

What it ships beyond the base model

A few capabilities matter for agent design specifically:

• DeepSearch turns Grok into an iterative RAG research agent with live X integration — it reads public posts for current events and uses native tools in a retrieval loop. For research skills that need fresh data, that's a real primitive; I wire it into a research agent in DeepSearch as an Agent Primitive.
• A code-focused model, grok-build (~256K context), sits alongside the general model for coding-agent work. How it stacks up against Codex-style agents is covered in Grok Build vs Codex.
• The 500K-token window shapes what you can stuff into a single call. The full spec sheet and what to do with that context lives in Grok 4.5's Spec Sheet: What 500K Context Actually Buys You.
• Trained alongside Cursor — corroborated by Cursor's own blog — which signals the coding-agent use case was a first-class target, not an afterthought.

One clarification, because the timeline blurs it: a separate voice product layer shipped around the same time (multilingual voices, voice cloning, a TTS API, a Voice Agent Builder). That is not part of the Grok 4.5 LLM. Don't conflate them, and don't assume native voice input to grok-4.5 — that specific claim is unconfirmed.

How a builder should act on this

Here's the practical filter for a skill or agent maintainer this week:

1. Add it as an option, don't rip anything out. The SDK-compatible API means Grok 4.5 can go into a router behind a feature flag in minutes. See Drop Grok 4.5 Into Your Multi-Model Skill Stack in Five Minutes.
2. Route on economics, not vibes. Point high-volume, output-heavy loops at the cheaper model and measure. If the token-efficiency claim holds on your workload, the savings compound.
3. Verify the claims on your own evals. Vendor benchmarks tell you where to look, not what you'll get. Run your real tasks through it before you trust the leaderboard.
4. Keep the frontier for the hard 20%. Top-of-index work still belongs on Fable 5 or Opus 4.8. Grok 4.5 is a cost lever, not a replacement for the ceiling.

If you're browsing for the pieces that make a multi-model setup work — routers, research agents, cost monitors — the skills catalog, agents, workflows, and automation loops are where those live.

The takeaway

Grok 4.5's launch shipped one thing you can bank on today and a pile of things you should verify. The bankable part: a top-5 independent Intelligence Index score at a >60% lower price, with a 500K context and a drop-in-compatible API. The verify-later part: every first-party benchmark, the token-efficiency multiplier, and the "Opus-class" framing.

That's actually the ideal shape for a builder. You don't need the vendor's superlatives to be true — you need the price to be real and the integration to be cheap, and both of those check out. Everything else you can test yourself. Start at the Grok series hub and read the pieces that match what you're building.