On July 8, 2026, xAI (SpaceXAI) launched Grok 4.5 — its new flagship for coding, agentic tasks, and knowledge work. The model is the default in Grok Build, available on all Cursor plans, and callable via the API as grok-4.5.
The commercial pitch is blunt: frontier-adjacent coding performance at $2 / $6 per 1M input/output tokens, served at a claimed ~80 tok/s, with roughly 2× token efficiency versus peers on the same tasks.
Key takeaways
- Positioning: coding + agents first — not a general chat refresh. Trained alongside Cursor, with anonymized Cursor workflow data in the model card’s supplemental training note.
- Specs: 500K context; text + image → text; reasoning effort low/medium/high (default high); tools include function calling, web/X search, and code execution.
- Price: $2 / $6 (cached input $0.50). Prompts that reach 200K tokens bill at $4 / $12 (cached $1.00) for the whole request.
- Distribution: API console, Grok Build (limited-time free usage at launch), Cursor, Office add-ins, plus OpenRouter / Vercel / Cloudflare / Snowflake / Databricks Mosaic.
- EU caveat: unavailable in the EU at launch; xAI said mid-July availability — verify live before EU rollouts.
Where the numbers come from
xAI’s announcement and model card lean on coding harnesses. A few vendor-reported highlights (effort modes in parentheses):
| Benchmark | Grok 4.5 | Context |
|---|---|---|
| SWE-Marathon | 29.0% (high) | Leads Opus 4.8 max (26%) and Fable 5 max (24%) on this ultra-long-horizon set |
| DeepSWE 1.0 | 62.0% (high) | Behind Fable 5 max / GPT-5.5 xhigh; ahead of Opus 4.8 max |
| SWE-Bench Pro | 64.7% (high) | Between Opus 4.8 max (69.2%) and GPT-5.5 xhigh (58.6%) |
| Terminal-Bench 2.1 | 83.3% (high) | Within a point of Fable 5 max / GPT-5.5 xhigh |
Treat these as vendor-reported until independent suites settle. Harness choice and “high/max” effort move scores.
Why it matters
- xAI is competing on the coding-agent stack, not just the chat leaderboard. Shipping the model as Grok Build’s default — and co-training with Cursor — is a distribution play as much as a benchmark play.
- Price pressure under Sol-class flags. At $2/$6, Grok 4.5 undercuts GPT-5.6 Sol ($5/$30) and sits in the same conversation as mid-tier frontier coding spend — if the efficiency claims hold on your workloads.
- Route carefully vs Grok 4.3. Grok 4.3 remains cheaper ($1.25/$2.50) with a 1M window. Use 4.5 for hard agent loops; keep 4.3 for long, cost-sensitive context.
Full wiki profile: Grok 4.5. For the open-sourced harness that defaults to this model, see Grok Build is now open source. Landscape context: 2026 LLM guide.