Grok 4.5: Price–Speed–Agent Contender Co-Trained with Cursor
xAI launched Grok 4.5 at $2/$6 per MTok (~80 TPS), default in Grok Build and Cursor, co-trained with Cursor. AA ~54; Arena mid-table—efficiency and coding-agent slices over overall chat crown; EU deferred mid-July.
Read Grok 4.5 as a price–speed–agent contender that ties Sol on some coding-agent slices without owning the overall Arena or AA intelligence crown. Cursor interaction data is the structural moat for agentic engineering—not a claim of chat #1.
xAI launched Grok 4.5 on July 8, 2026—its smartest model for coding, agentic tasks, and knowledge work, co-trained with Cursor. Pricing is $2 / $6 per million input/output tokens, served at about 80 TPS, with claims of roughly 4.2× fewer output tokens than Opus 4.8 (max) on SWE-Bench Pro–style efficiency comparisons. Availability: Grok Build (default), Cursor (all plans), and the xAI API console; not yet in the EU (mid-July expected). Elon Musk called it an “Opus-class” model that is faster, more token-efficient, and lower cost.
What xAI announced
From the primary “Introducing Grok 4.5” post (x.ai/news/grok-4-5) and Cursor’s co-launch post:
- Positioning: Strongest xAI model yet; coding + agentic + knowledge/office work (not coding-only).
- Training: Co-trained with Cursor on large-scale coding and multi-step engineering RL; tens of thousands of NVIDIA GB300 GPUs; heavy data curation and long agentic rollouts.
- Speed / efficiency:
80 tokens per second; vendor charts show much lower average output tokens per SWE-Bench Pro task vs Opus 4.8 max (**4.2× fewer**). - Benchmarks (vendor / third-party harness figures cited by xAI): Competitive on Terminal Bench 2.1 and SWE Marathon; trails Fable 5 / GPT-5.5 on some DeepSWE and SWE-Bench Pro resolve rates—strength framed as cost and token efficiency, not always raw top score.
- Office: Default in Grok Build; Excel/PowerPoint/Word plugin workflows.
- Pricing: $2 / $1M input, $6 / $1M output.
- Access: Grok Build, Cursor (desktop/web/iOS/CLI/SDK; first-week usage boost noted by Cursor), API model id
grok-4.5; free limited Grok 4.5 usage offered in Grok Build/Cursor at launch. - EU: Explicitly not available yet in products or API console; mid-July target.
Context from contemporaneous TechCrunch/Axios/Forbes: first major model release after xAI’s public-company phase and Cursor acquisition path; joint Cursor–xAI training flywheel.
Claims vs checks
| Source | Snapshot | Standing |
|---|---|---|
| Artificial Analysis Intelligence Index | Launch-week / mid-July writeups | Early reports place Grok 4.5 ~54 — roughly #4 among labs with a model above 50 after the Jul 8–9 cluster (behind Fable 5, Sol, and often Opus 4.8 depending on snapshot) |
| AA Coding Agent Index | AA GPT‑5.6 note | Grok 4.5 in Grok Build ties Sol/Codex on SWE-Atlas-QnA; Sol still leads the full Coding Agent Index composite |
| Arena Text Overall | Board dated Jul 16, 2026 | grok-4.5 at 1465 ±8 (~5,493 votes) — roughly #34 overall; well below Fable/Muse Spark 1.1/Kimi K3/Sol top band |
| BenchLM / secondary boards | Verified ~Jul 16 | Text Overall Elo |
Note: OpenAI’s same-day SWE-Bench Pro audit (~30% broken tasks) means vendor Pro resolve/efficiency charts—including the 4.2× token claim—should be read with that noise floor in mind.
Limits
- EU unavailable at launch (mid-July target).
- Vendor SWE-Bench Pro efficiency charts sit beside OpenAI’s Pro audit noise finding.
- AA ~54 and Arena #34 are launch/mid-July snapshots.
- “Opus-class” is Musk framing, not an independent Elo match.
Sources
- xAI: “Introducing Grok 4.5” (x.ai/news/grok-4-5, July 8, 2026). Primary.
- Cursor: “Introducing Grok 4.5” co-launch blog (July 8, 2026).
- Artificial Analysis launch-week Intelligence Index / Coding Agent notes (incl. SWE-Atlas-QnA tie vs Sol).
- Arena.ai Text Overall leaderboard snapshot (July 16, 2026).
- TechCrunch, Axios, Forbes same-day coverage; Musk X framing (“Opus-class,” public availability).