Grok 4.5 is xAI’s (SpaceXAI) July 8, 2026 flagship for coding, agentic work, and knowledge workflows. The company calls it its strongest model yet, trained across science/engineering/math datasets alongside Cursor, and shipped as the default in Grok Build plus every Cursor plan.
The practical pitch is efficiency, not just peak score: xAI claims ~80 tokens/second serving and roughly 2× token efficiency versus comparable frontier models on the same tasks — fewer steps, lower wall-clock cost, at $2 / $6 per 1M input/output.
Specs at a glance
| Spec | Value |
|---|---|
| Model ID | grok-4.5 (aliases: grok-4.5-latest, grok-build-latest) |
| Context | 500K tokens |
| Modalities | Text + image in → text out |
| Knowledge cutoff | February 1, 2026 (docs); pretraining cutoff January 2026 (model card) |
| Reasoning effort | Low / medium / high (default) |
| Tools | Function calling, structured outputs, web search, X search, code execution |
| Throughput (vendor) | ~80 tok/s |
Full live limits for your org live on the model detail page.
Pricing
Base API rates (per 1M tokens), with a 200K prompt long-context surcharge — when the prompt reaches 200K, all tokens in that request bill at the higher tier:
| < 200K prompt | ≥ 200K prompt | |
|---|---|---|
| Input | $2.00 | $4.00 |
| Cached input | $0.50 | $1.00 |
| Output | $6.00 | $12.00 |
xAI strongly recommends setting prompt_cache_key (Responses API) or the x-grok-conv-id header (Chat Completions) so conversation turns hit the same cache-warm server. Without it, you often pay full input price on a cold node.
For a cheaper, longer-window sibling, keep Grok 4.3 ($1.25 / $2.50, 1M context) in the routing table.
Benchmarks (vendor-reported — read the caveats)
xAI’s launch post and model card lean on coding / agent harnesses. Figures below are vendor-reported (often with provider harnesses / AA or Cursor runs) unless noted. Modes in parentheses are the effort settings used in those reports.
| Benchmark | Grok 4.5 | Notable peers (same card / post) |
|---|---|---|
| DeepSWE 1.0 (pass@1) | 62.0% (high) | Fable 5 max 66.1%, GPT-5.5 xhigh 64.3%, Opus 4.8 max 55.8% |
| DeepSWE 1.1 | 53.0% (high) | Fable 5 max 70.0%, GPT-5.5 xhigh 67.0%, Opus 4.8 max 59.8% |
| SWE-Bench Pro | 64.7% (high) | Fable 5 max 80.3%, Opus 4.8 max 69.2%, GPT-5.5 xhigh 58.6% |
| SWE-Marathon | 29.0% (high) | Opus 4.8 max 26.0%, Fable 5 max 24.0% |
| Terminal-Bench 2.1 | 83.3% (high) | Fable 5 max 84.3%, GPT-5.5 xhigh 83.4%, Opus 4.8 max 78.9% |
| ApexSWE | 51.2% (high) | Fable 5 max 65.5%, Opus 4.8 max 47.3% |
Caveats that matter: harness choice moves scores; “high/max/xhigh” effort modes are not free; SWE-Marathon is ultra-long-horizon and still low absolute pass rates; DeepSWE 1.0 vs 1.1 show how fast the yardstick moves. Prefer independent Artificial Analysis / community replications before locking a production switch. For a landscape view, see the 2026 LLM guide.
Availability and access
As of the launch announcement:
- xAI API —
console.x.ai, modelgrok-4.5 - Grok Build — default coding-agent model (CLI + API); limited-time free usage was offered at launch
- Cursor — all plans
- Office add-ins — default in Word / PowerPoint / Excel plugins
- Gateways — OpenRouter, Vercel, Cloudflare, Snowflake, Databricks Mosaic
- Consumer Grok / X — planned later, not day-one
EU note: at launch, Grok 4.5 was unavailable in the EU across products and the API console, with availability “expected in mid-July.” Confirm live region support before shipping EU-facing features.
Minimal Responses API call:
curl -s https://api.x.ai/v1/responses \
-H "Authorization: Bearer $XAI_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "grok-4.5",
"input": "Find and fix the bug, then explain it: function median(a){a.sort();return a[a.length/2]}"
}'How it compares
- vs Grok 4.3: 4.5 is the coding/agent flagship at higher token prices and a 500K window; 4.3 stays the budget / 1M-context workhorse. Route hard agent loops to 4.5, long RAG dumps to 4.3 when cost dominates.
- vs Claude Opus 4.8: Opus still leads several SWE-Pro / DeepSWE splits in the same cards; Grok 4.5 counters on SWE-Marathon and price/token-efficiency positioning.
- vs GPT-5.5 / GPT-5.6: GPT remains strong on terminal suites; Grok 4.5 undercuts Sol-class pricing while chasing coding-agent share via Build + Cursor distribution.
- vs Gemini 3.1 Pro: Gemini remains the multimodal / value play; Grok 4.5 is narrower (text+image → text) and more coding-agent oriented.
Limitations and unknowns
- No public batch discount on 4.5 at launch (4.3 had 20%).
- Consumer chat surfaces were not day-one.
- Safety evals are documented in the model card (cyber, bio, jailbreaks, mental health) — read them if you ship in sensitive domains; this page does not restate every refusal number.
- Independent, vendor-neutral intelligence indexes for 4.5 were still settling as of mid-July 2026 — we will update when stable AA / community numbers land.
Related reading: Grok Build open-sourced (the harness that defaults to this model) and the 2026 LLM landscape.
Best for
- Agentic coding in Grok Build / Cursor
- Multi-step software engineering and terminal workflows
- Knowledge work in Word / PowerPoint / Excel add-ins
- Cost-conscious frontier coding where token efficiency matters
Pros & cons
Strengths
- Strong vendor-reported scores on long-horizon SWE (SWE-Marathon) and coding agents
- Competitive $2/$6 API pricing with ~2× token efficiency claims vs peers
- Default model in Grok Build and available on all Cursor plans
- Configurable reasoning effort (low / medium / high) plus web/X search and code execution tools
Limitations
- Context window (500K) is half of Grok 4.3's 1M
- Headline benchmarks are vendor-reported; independent suites still catching up
- EU availability was blocked at launch (mid-July target); verify current region status
- Long-context prompts (≥200K) double the token rates
Frequently asked questions
What is the API model ID for Grok 4.5?
Use `grok-4.5` (aliases include `grok-4.5-latest` and `grok-build-latest`). It works with the Responses API and Chat Completions at `https://api.x.ai/v1`.
How does Grok 4.5 pricing compare to Grok 4.3?
Grok 4.5 is $2 input / $6 output per 1M tokens (cached input $0.50), versus Grok 4.3's $1.25 / $2.50 (cached $0.20). 4.5 is positioned as the smarter coding flagship; 4.3 remains the cheaper, longer-context (1M) option. Prompts that reach 200K tokens bill at double rates on both models.
Is Grok 4.5 available in the EU?
At the July 8, 2026 launch, xAI said Grok 4.5 was not available in the EU in any SpaceXAI product or the API console, with availability "expected in mid-July." Check the live console / docs for your region before planning an EU rollout.
Where can I use Grok 4.5 today?
xAI API console, as the default model in Grok Build (CLI/API), on all Cursor plans, as the default in Word/PowerPoint/Excel add-ins, and via gateways such as OpenRouter, Vercel, Cloudflare, Snowflake, and Databricks Mosaic. Consumer Grok apps / X were listed as later.
Was Grok 4.5 trained with Cursor?
Yes. xAI's announcement and model card state the model was trained alongside Cursor, with supplemental training on anonymized Cursor workflow data to improve coding and agentic performance. Treat Cursor's public praise as vendor-adjacent, not a neutral third-party review.