Grok 4.5, Three Weeks In: Benchmarks, Pricing, and the Cursor Co-Training Bet That Redefines SpaceXAI

#ai #agents #grok45 #spacexai

TL;DR

Grok 4.5 launched July 8, 2026 — jointly trained by SpaceXAI and Cursor on trillions of real developer-agent interaction tokens
Architecture: Mixture-of-Experts, 500K context window, text+image input, text output, high-reasoning always-on
Pricing: $2/M input, $6/M output, $0.50/M cached input — roughly 4× cheaper than Claude Opus 4.8 on output
Benchmarks: Artificial Analysis ranks it #4 of 168 models on Intelligence Index (score 54) — behind Fable 5, GPT-5.5, and Opus 4.8
Cursor integration: Default model in Cursor IDE as of July 8. Also available via Grok Build CLI, SpaceXAI API, grok.com
Speed: ~80-87 tokens/sec throughput, but ~16.7s time-to-first-token due to always-on reasoning
Key differentiator: SpaceXAI claims 4× fewer output tokens per task than Claude Opus 4.8 on SWE-Bench Pro

The Model That Came With a $60B Acquisition

Grok 4.5 is not just another version bump. It is the first product of SpaceX's $60B acquisition of Anysphere, the company behind Cursor — and that acquisition defines everything about the model.

Where most frontier models are trained on web crawl data, textbooks, and synthetic datasets, Grok 4.5 was co-trained on trillions of tokens from real Cursor interactions: debugging traces, tool calls, multi-step agentic coding sessions, code reviews. The training data is structurally different from what any other lab has access to — not because the data is secret, but because the volume and shape of agent-to-codebase interaction data is something only Cursor has.

Lee Robinson, Cursor's ML lead, detailed the training flywheel at the AI Engineer stage in July 2026: a two-loop recursive improvement system using SpaceX's Colossus compute cluster, textual feedback RL for credit assignment, and agents that train models from Slack. (Source: explainx.ai — Recursive Model Improvement)

Benchmarks: What the Numbers Actually Say

Three weeks of independent testing give us a clearer picture than launch-day claims. Here's the data:

Benchmark	Grok 4.5	Claude Opus 4.8	GPT-5.6 Sol	Notes
SWE-Bench Pro	61.4%	69.2%	~65%	Opus leads on autonomous coding
Terminal-Bench 2.1	83.3%	79.1%	~91.9%	Grok strong, GPT leads
DeepSWE 1.0	56.0%	60.5%	59.2%	Using each provider's agent framework
DeepSWE 1.1	52.1%	48.9%	51.0%	Shared mini-swe-agent, Grok wins
CursorBench	[excluded]	—	—	Training data contamination
GDPval+ (Snorkel)	29%	21%	22%	Professional work tasks
AA Intelligence Index	54 (4th)	55 (3rd)	56 (2nd)	Out of 168 models

(Sources: DataCamp, explainx.ai, Codersera)

The pattern is consistent: Grok 4.5 is competitive but not dominant on raw coding benchmarks. It wins on some agentic tasks (DeepSWE 1.1 with shared setup, Terminal-Bench 2.1 vs Opus) but trails on SWE-Bench Pro. Its strongest performance is on GDPval+, where Snorkel AI's professional work evaluation scored it meaningfully higher than GPT-5.5 and Opus 4.8 — suggesting its agentic training data generalized beyond pure coding.

The Pricing Advantage

SpaceXAI's pricing is the most aggressive among frontier models:

Model	Input / 1M	Output / 1M	Cached input
Grok 4.5	$2.00	$6.00	$0.50
Claude Opus 4.8	$5.00	$25.00	—
GPT-5.5 / 5.6	$5.00	$30.00	—
Claude Fable 5	$10.00	$50.00	—

(Source: Codersera — Grok 4.5 Launch Guide)

But per-token pricing tells only half the story. SpaceXAI's real claim is token efficiency — that Grok 4.5 uses fewer output tokens per completed task. Its SWE-Bench Pro chart shows 15,954 output tokens per resolved task versus 67,020 for Claude Opus 4.8. If that ratio holds in production, the effective cost advantage is closer to 10-15× than the headline 4×.

What Changed in Three Weeks

Since the July 8 launch:

July 9: Public rollout on grok.com and X app
July 14: Privacy crisis — Grok Build wire tests captured full-repo uploads. SpaceXAI documented ZDR and /privacy, Elon Musk pledged total deletion of pre-July 14 data
July 17: Lee Robinson's AI Engineer talk revealed the training flywheel
July 16: SpaceXAI open-sourced the Grok Build harness (Apache 2.0) on GitHub
July 17: EU availability expected (still pending per launch communications)

The privacy incident is worth watching. Grok Build was uploading entire repositories to SpaceXAI servers during wire testing, a practice that was not clearly documented. SpaceXAI's response — retroactive deletion, a /privacy endpoint, and a public apology — was effective, but the incident highlights the tension between agentic coding tools and enterprise data governance.

The Competitive Position

Three weeks in, Grok 4.5 occupies a clear niche: near-frontier quality at mid-tier pricing. It is not the best model on any single benchmark, but it is the cheapest among models that can plausibly compete at this level.

For developers, the calculus is straightforward:

Best quality (no budget constraint): Claude Fable 5 or GPT-5.6 Sol
Best coding agent: Claude Opus 4.8 (SWE-Bench Pro leader) or Grok 4.5 (cheaper per task)
Best price/performance: Grok 4.5, especially for high-volume agentic workloads
Best ecosystem integration: Grok 4.5 if you already use Cursor; Claude Opus if you use Claude Code

The model's real test will come when SpaceXAI ships the next version. With a stated cadence of a new foundation model every month through end of 2026, Grok 4.5 may be the cheapest entry point into a rapidly improving family.

FAQ

How much does Grok 4.5 cost? — $2.00 per million input tokens, $6.00 per million output tokens, $0.50 cached input (75% discount). Roughly 4× cheaper than Claude Opus 4.8 on output tokens.

Is it better than Claude Opus 4.8? — On raw benchmarks, mostly no. It trails Opus 4.8 on SWE-Bench Pro (61.4% vs 69.2%). But it wins on cost and token efficiency, and Snorkel's GDPval+ evaluation suggests it outperforms on professional work tasks.

What's the context window? — 500,000 tokens, which is actually smaller than Grok 4.3's 1M-token window.

Can I use it in Cursor? — Yes, it became the default model in Cursor on July 8.

What happened with the privacy issue? — Grok Build was uploading full repos during wire testing. SpaceXAI implemented retroactive deletion, a /privacy endpoint, and pledged total deletion of pre-July 14 data.