Grok 4.5 (xAI)

Source: raw/How_the_4_New_Models_Released_This_Week_Will_Change_How_You_Work.md (The AI Daily Brief, NLW — weekly roundup covering four model releases in one week)

Grok 4.5 is xAI’s first model built from its new collaboration with Cursor, following SpaceX/xAI’s acquisition of Cursor — explicitly positioned as “built for real world engineering” rather than a general-purpose chat release. It ranks 4th on the Artificial Analysis Index, behind Fable 5, Opus 4.8, and GPT-5.5, but is by far the most cost-efficient near-frontier model tested and the outright leader on Artificial Analysis’s AutomationBench. Early-adopter framing treats it as a cheap, fast implementation-agent model to run underneath a Fable 5 or GPT-5.6 orchestrator — not as a like-for-like replacement for either.

Key Takeaways

• Origin: first model from the SpaceX/xAI + Cursor collaboration. Grok 4.5 follows SpaceX/xAI’s acquisition of Cursor. xAI’s own framing: “built for real world engineering… excels in large code bases and handles long-running tasks that span multiple repositories, hundreds of skills, and a variety of tools.” Cursor’s own tweet: “It’s our most powerful model yet, and the first we’ve built for more than software engineering.” Cursor’s existing Composer line (built on a Kimi base) stays a separate, smaller weight class — Composer 2.5 remains offered, with more Composer-size models planned.
• Artificial Analysis Index: 4th place, but the cost-efficiency leader. Ranks 4th overall behind Fable 5, Opus 4.8, and GPT-5.5^[this ranking is dated to the source’s capture window, before GPT-5.6’s 2026-07-09 GA — see gpt-5-6-series for GPT-5.6’s fuller benchmark picture]; particularly strong in agentic knowledge work and coding, and also 4th on GDPval. On cost, though, it leads outright: **$0.31/task\ast \ast ontheoverallindex—aboutathirdofGPT-5.5^{\prime }scost,afifthofOpus4.8^{\prime }s($1.80/task), and (the source’s fraction word is garbled in transcription, but the math checks out) roughly a ninth of Fable 5’s ($2.75/task)^[inferred ratio — computed from the two dollar-per-task figures stated in the same source]. Slightly cheaper than GLM 5.2 and Kimi K2.6, slightly pricier than Haiku 4.5.
• Roughly matches Opus 4.8 and GPT-5.5 on core coding benchmarks. Described as “more or less a match” for both across Terminal-Bench 2.1, SWE-bench Pro, and “Deep SWE-bench 1.0”^[benchmark name normalized from the source’s spoken “Deep Suite 1.0”; not independently verified against a primary listing].
• State-of-the-art on AutomationBench (agentic tool-use across simulated real-world SaaS workflows — Excel, Gmail, Slack): top score 51.4% vs Fable 5’s 48.6%, at **$0.34/task\ast \ast vsFabl{e}^{\prime }s$1.35/task — using under half of Opus 4.8’s tokens and fewer turns than other frontier models.
• 3rd-best harness/model pairing on Artificial Analysis’s Coding Agent Index: Cursor + Grok 4.5^[inferred — audio transcribed as “Grok build”; read here as “Cursor + Grok 4.5” given the described SpaceX/Cursor collaboration, not independently confirmed] ranks 3rd, ahead of Opus 4.8 in Claude Code, just behind GPT-5.5 + Codex and Fable 5 + Claude Code (relative order between those top two not stated in the source).
• Early-adopter framing: an implementation agent, not a flagship replacement. Elon Musk, quoted directly: “Fable is definitely better than Grock 4.5 but most tasks don’t require Fable level capability.” The pattern described among early adopters is running Grok 4.5 as one of several implementation agents while Fable 5 or GPT-5.6 acts as orchestrator — not a drop-in swap for either.
• Positive hands-on creator review (Theo). Direct quotes: “I had no ergonomic issues with 4.5 at all. It was actually really pleasant to work with… Was it able to go back and forth with me on really big, heavy tasks and be pleasant to use while also being very fast and cheap? Yeah.” Theo’s verdict: a good alternative to something like Opus 4.8 that “kind of outclasses GLM 5.2” on capability (though GLM keeps the open-weights advantage) — “a weirdly good default code model.” Elon Musk retweeted Theo’s coverage approvingly.
• Positioned against Chinese open-weight models, not just Claude/GPT. Community commentary quoted in the source (attributed to “Ston Daddy” on X) frames Grok 4.5 as delivering “better than Chinese open source performance at near Chinese open source cost” without the data-sovereignty, compliance, and geopolitical hesitation that keeps some enterprises off GLM 5.2 or Kimi K — western-built, and already riding Cursor’s existing enterprise footprint as a distribution channel.

Try It

• If evaluating a multi-model agentic setup, treat Grok 4.5 as an implementation-agent candidate (cheap, fast, “more or less” frontier-competitive on coding) with Fable 5 or GPT-5.6 Sol as orchestrator — not as a like-for-like Fable/Sol replacement.
• If already inside Cursor, Grok 4.5 is reachable without adding a new vendor relationship — the lowest-friction way to test the cost-efficiency claims above.
• Before switching a workflow off GLM-5.2 or another open-weight model purely for data-sovereignty reasons, weigh Grok 4.5 as the western-built alternative the community discussion above frames it as.

Open Questions

• Single-source article. Every claim here traces to one podcast source (The AI Daily Brief, ingested 2026-07-10), which itself aggregates creator and community reactions (Theo, Elon Musk tweets, “Ston Daddy” commentary) rather than primary xAI or Artificial Analysis documentation. Treat quantitative claims as secondary-sourced pending an official primary citation.
• Exact API pricing not confirmed here. This article does not independently verify Grok 4.5’s per-token API pricing; scope was kept to the single assigned source rather than pulling in figures from other raw sources ingested the same day.
• “Cursor + Grok 4.5” harness pairing. The Coding Agent Index pairing is transcribed ambiguously in the source audio (“Grok build”) — read here as “Cursor + Grok 4.5” but not verified against a primary Artificial Analysis listing.
• “Deep SWE-bench 1.0” name. Best-guess normalization of the source’s spoken “Deep Suite 1.0” — not confirmed against an official benchmark listing.

Related

• gpt-5-6-series — the direct sibling competitor-model article; GPT-5.6 GA’d the same week and covers the fuller cross-model benchmark picture (Agent’s Last Exam, Buybench, AA Coding Index), including Grok 4.5 cameo appearances (SWE-Marathon lead, Terminal-Bench/Deep-SWE-bench figures).
• claude-fable-5-mythos-5 — the frontier anchor Grok 4.5 is benchmarked against and explicitly not positioned to replace.
• glm-5-series-zai — the open-weight competitor Grok 4.5 is positioned against on the “avoid Chinese open-weight models” enterprise axis.
• claude-opus-4-8 — the capability tier creator Theo places Grok 4.5 closest to.
• ai-competition-shifts-beyond-model-quality — the wider thesis (Nate B Jones) that competition is shifting to infrastructure/distribution/cost, which Grok 4.5’s cost-efficiency story fits directly.
• _index — competitor-model / benchmark tracking topic index.