概要
Grok 4.5, released July 8, 2026 by SpaceXAI (the merged SpaceX/xAI entity), represents a significant leap forward for xAI's model lineup, jumping 16 points on the Artificial Analysis Intelligence Index from Grok 4.3's score of 38 to 54, placing it fourth overall behind only Claude Fable 5, Claude Opus 4.8, and GPT-5.5. The model was jointly trained with Cursor on trillions of tokens of real developer-agent interaction data, following SpaceX's $60 billion acquisition of Anysphere (Cursor's maker) in June 2026. It is a 1.5 trillion parameter mixture-of-experts model (per Musk's disclosure, not officially confirmed by SpaceXAI) trained on tens of thousands of NVIDIA GB300 GPUs.
Grok 4.5's positioning is deliberate: it does not claim to be the absolute smartest model, but rather the best intelligence-per-dollar option in the near-frontier tier. Its standout achievement is the #1 ranking on agentic tool use (τ³-Banking at 33%), and its token efficiency is remarkable — averaging just 1.9 million tokens per coding agent task compared to 6.2-7.2 million for competitors. At $2/$6 per 1M input/output tokens, it costs 60-75% less than Claude Opus 4.8 or GPT-5.5, while completing tasks at roughly $2.49 per Coding Agent Index task versus $5.07 for GPT-5.5 and $11.80 for Fable 5.
The model ships as the default in Grok Build (xAI's coding agent CLI), is available across all Cursor plans, and is integrated into Microsoft Office add-ins for Word, PowerPoint, and Excel. Access is also available through OpenRouter, Vercel, Cloudflare, Snowflake, and Databricks Mosaic. The trade-offs are real: it trails the top models on raw coding accuracy (64.7% on SWE-Bench Pro vs. Fable 5's 80.4%), its hallucination rate has increased, and it reduced the context window from 1M to 500K tokens. But for high-volume agentic workloads where cost and speed matter as much as peak accuracy, Grok 4.5 has established itself as the most compelling value proposition on the market.
ベンチマーク＆性能
## Independent Benchmark Results (Artificial Analysis, July 2026)
| Benchmark | Grok 4.5 | Claude Fable 5 (max) | Claude Opus 4.8 (max) | GPT-5.5 (xhigh) |
|---|---|---|---|---|
| Intelligence Index | 54 (#4) | 60 (#1) | 56 (#2) | 55 (#3) |
| τ³-Banking (Agentic Tool Use) | 33% (#1) | — | — | 31% |
| GPQA Diamond | ~93% | — | — | ~93% |
| Terminal-Bench v2.1 | 83.3% | 84.3% | 78.9% | 83.4% |
| GDPval-AA v2 Elo | 1543 (#4) | — | 1600 (#1) | — |
| AA-Omniscience Index | 26 | — | — | — |
| Speed (output tokens/sec) | ~80-86 | — | — | — |
## Coding Benchmarks (xAI Published, Datacurve/Harbor)
| Benchmark | Grok 4.5 | Claude Fable 5 (max) | Claude Opus 4.8 (max) | GPT-5.5 (xhigh) |
|---|---|---|---|---|
| DeepSWE 1.0 | 62.0% | 66.1% | 55.75% | 64.31% |
| DeepSWE 1.1 | 53% | 70% | 59% | 67% |
| Terminal-Bench 2.1 | 83.3% | 84.3% | 78.9% | 83.4% |
| SWE-Bench Pro | 64.7% | 80.4% | 69.2% | 58.6% |
## Cost Efficiency (Coding Agent Index)
| Model | Task Cost | Avg Tokens/Task | Score |
|---|---|---|---|
| Grok 4.5 (Grok Build) | $2.49 | 1.9M | 76 |
| GPT-5.5 (Codex) | $5.07 | 6.2M | 76 |
| Fable 5 (Claude Code) | $11.80 | 7.2M | 77 |
## Professional Work (Snorkel GDPval+, ~2000 tasks)
| Model | Mean Pass Rate |
|---|---|
| Grok 4.5 (Grok Build) | 29% |
| GPT 5.5 (Stirrup agent) | 22% |
| Claude Opus 4.8 (Stirrup agent) | 21% |
Grok 4.5 led in legal work (40% vs 27-28%), education (58% vs 35-42%), healthcare (35% vs 23-25%), and QA analysis (37% vs 19-27%).
## Hallucination Trade-off
Accuracy on AA-Omniscience Index improved from 35% to 52%, but hallucination rate rose from 25% to 54% — a common pattern where larger models know more but are also more confident when wrong.
詳細比較
## Grok 4.5 vs Claude Fable 5 (max)
Claude Fable 5 is the undisputed accuracy leader, scoring 60 on the Intelligence Index and 80.4% on SWE-Bench Pro versus Grok 4.5's 54 and 64.7%. However, Fable 5 costs $10/$50 per 1M input/output tokens — roughly 5x more expensive on output. At $11.80 per coding agent task versus $2.49 for Grok 4.5, the cost gap is nearly 5x. For the hardest multi-file refactors and debugging where a wrong edit is expensive, Fable 5 is the better pick. For high-volume agentic pipelines, Grok 4.5 delivers 95%+ of the quality at a fraction of the cost.
## Grok 4.5 vs Claude Opus 4.8 (max)
Elon Musk positioned Grok 4.5 as "roughly comparable to Opus 4.7" and the independent data largely supports this. On the Intelligence Index, Opus 4.8 scores 56 vs Grok 4.5's 54. Opus 4.8 leads on SWE-Bench Pro (69.2% vs 64.7%) and DeepSWE 1.1 (59% vs 53%), while Grok 4.5 wins on Terminal-Bench 2.1 (83.3% vs 78.9%) and DeepSWE 1.0 (62% vs 55.75%). Pricing is where they diverge sharply: Opus 4.8 is $5/$25 per 1M tokens, making Grok 4.5 about 2.5x cheaper on input and 4.2x cheaper on output. On the GDPval+ professional work benchmark, Grok 4.5 actually outperformed Opus 4.8 (29% vs 21% pass rate).
## Grok 4.5 vs GPT-5.5 (xhigh)
These two are the closest match. On the Intelligence Index, GPT-5.5 edges ahead 55 to 54. On the Coding Agent Index, they tied at 76. Grok 4.5 beats GPT-5.5 on SWE-Bench Pro (64.7% vs 58.6%) and τ³-Banking agentic tool use (33% vs 31%), while GPT-5.5 wins on DeepSWE 1.1 (67% vs 53%). GPT-5.5 costs $5/$30 per 1M tokens versus Grok 4.5's $2/$6, and on coding agent tasks costs $5.07 per task versus $2.49. Context windows vary by GPT-5.5 tier; Grok 4.5 offers a flat 500K. For teams already in the OpenAI ecosystem with Codex, GPT-5.5 offers deeper integration; for cost-sensitive agent work, Grok 4.5 is the stronger value.
## Summary Pricing Table
| Model | Input/1M | Output/1M | Cached Input/1M | Context Window |
|---|---|---|---|---|
| Grok 4.5 | $2.00 | $6.00 | $0.50 | 500K |
| Claude Opus 4.8 | $5.00 | $25.00 | — | Varies |
| GPT-5.5 | $5.00 | $30.00 | — | Varies |
| Claude Fable 5 | $10.00 | $50.00 | — | Varies |
コミュニティ評価
The community reaction to Grok 4.5 has been genuinely split, centering on two themes: cost-performance value and trust.
**Positive reception:** Cursor CEO Michael Truell called Grok 4.5 "an Opus-class model that's fast and low cost" and noted it had "become the daily driver for many on our team." Artificial Analysis highlighted it as sitting firmly on the performance/cost Pareto frontier. Developer Thomas Wiegold, in a hands-on review running his standard coding test suite, found Grok 4.5 one-shotted his Go poker simulation (a test only Claude Fable had previously passed cleanly) and caught a real bug on his own site that no other model — including Fable — had ever flagged. He concluded it was "a genuinely good coding model" and "the one I'll actually reach for day to day."
**Trust concerns:** The loudest discussion on Hacker News's launch thread was not about capability but about alleged political steering of outputs by xAI leadership. One commenter wrote: "I just don't think that I can ever trust an xAI model knowing that they are actively trying to shape its replies to fit a political narrative." Others pushed back, noting Grok appeared "very tame" in practice. This trust debate remains unresolved and is a genuine factor for enterprise adoption.
**Coding quality nuance:** In community head-to-head comparisons where developers had Grok 4.5, GPT-5.5, and Claude build the same apps, one commenter noted it was "so strange to write a whole post with Claude giving the best results and Grok consistently the worst, but awarding Grok the winner because at least it did the worst fastest." Another called it "pretty decent, comparable with some older Opus models, and fairly cheap per token." The consistent thread is: fast and cheap, with good-enough quality for most work, but not best-in-class output quality.
**Security concern:** In July 2026, an independent researcher found that Grok Build was uploading entire Git repositories (including full commit history and secrets) to xAI-controlled cloud storage. Elon Musk said SpaceXAI would delete previously uploaded data, and uploads were disabled server-side, but no formal security advisory or independent audit has been published as of mid-July 2026.
**Adoption patterns:** Early adoption is concentrated among cost-sensitive developer teams and agent builders. The integration into Cursor (all plans) and Office add-ins provides broad access. European developers face a delay due to EU AI Act compliance work.
ユースケース
### 1. High-Volume Agentic Coding Pipelines
Grok 4.5's #1 ranking on agentic tool use and its extraordinary token efficiency (1.9M tokens per task vs 6-7M for competitors) make it the optimal default model for CI/CD-integrated coding agents, automated code review bots, and large-scale refactoring pipelines. At $2.49 per coding agent task versus $11.80 for Fable 5, teams running thousands of agent tasks monthly see massive cost savings with only marginal quality differences on routine tasks. Use Grok 4.5 as the default tier, escalating only the hardest multi-file refactors to Claude Fable 5 or Opus 4.8.
### 2. Professional Knowledge Work and Document Generation
The Snorkel GDPval+ benchmark showed Grok 4.5 achieving a 29% mean pass rate on professional workplace tasks — outperforming GPT-5.5 (22%) and Opus 4.8 (21%) — with particularly strong results in legal work (40%), education (58%), and healthcare (35%). For enterprises building AI assistants that draft reports, analyze spreadsheets, prepare decision memos, or conduct research, Grok 4.5 offers near-frontier quality at dramatically lower cost. The 500K token context window accommodates substantial document collections in a single prompt.
### 3. Cost-Sensitive Startup and SMB Development
At $30/month for SuperGrok consumer access or $2/$6 per 1M API tokens, Grok 4.5 is the most affordable path to near-frontier AI capabilities. Startups and small teams can build production AI features — chatbots, code assistants, research tools — without the $25-$50 per 1M output token costs of premium Claude or GPT models. The ~80 tokens/sec output speed also means faster iteration cycles during development.
### 4. Terminal and DevOps Automation
Grok 4.5 scored 83.3% on Terminal-Bench 2.1, nearly matching GPT-5.5 (83.4%) and surpassing Opus 4.8 (78.9%). For teams building DevOps agents that manage infrastructure through command-line interfaces — deploying, debugging, monitoring, and responding to incidents — Grok 4.5 provides strong terminal-native performance at a fraction of the cost. The token efficiency advantage compounds in multi-step agent loops where each tool call adds to the bill.
### When to Choose Alternatives:
- **Choose Claude Fable 5** when task failure is expensive (critical production code, complex multi-file refactors) and accuracy must be maximized regardless of cost.
- **Choose Claude Opus 4.8** when you need the strongest balanced option across coding, reasoning, and long-context work.
- **Choose GPT-5.5** when deep OpenAI ecosystem integration (Codex, Assistants API, DALL-E) or multimodal capabilities are primary requirements.
最新ニュース
**July 8, 2026 — Grok 4.5 Launch:** SpaceXAI released Grok 4.5 as the default model in Grok Build, across all Cursor plans, and in Microsoft Office add-ins. The model was jointly trained with Cursor on trillions of tokens of developer interaction data, following SpaceX's $60 billion acquisition of Anysphere.
**July 8, 2026 — Pricing Announced:** $2.00/1M input, $6.00/1M output, $0.50/1M cached input. Server-side tools (web search, X search, code execution) billed at $5/1,000 calls. No batch discount at launch, unlike Grok 4.3's 20% discount.
**July 8, 2026 — Cursor Partnership:** Cursor offered a temporary 50% usage promotion (described as double usage) for the first week of Grok 4.5 availability.
**July 8, 2026 — Artificial Analysis Benchmark:** Grok 4.5 placed #4 on the Intelligence Index (score 54), took #1 on agentic tool use (τ³-Banking at 33%), and scored 76 on the Coding Agent Index tied with GPT-5.5.
**July 8, 2026 — Snorkel GDPval+ Results:** Grok 4.5 achieved 29% mean pass rate on ~2,000 professional work tasks, outperforming GPT-5.5 (22%) and Opus 4.8 (21%).
**Mid-July 2026 — Security Issue:** Independent researcher discovered Grok Build was uploading entire Git repositories to xAI servers. Musk stated data would be deleted; uploads disabled server-side. No formal security advisory published.
**Mid-July 2026 — EU Availability:** Grok 4.5 not yet available in the EU as of mid-July 2026 due to EU AI Act compliance work; expected later in July.
**July 14, 2026 — Speed Confirmation:** Artificial Analysis measured ~120 tokens/sec output speed (above SpaceXAI's claimed 80), with a 14.5-second wait before first token versus a 2.7-second median for the price tier.