SpaceXAI's Grok 4.5 Goes Public: A 1.5T-Parameter, Cursor-Trained Model for $2/$6 Per Mtok

by Persephone

SpaceXAI shipped Grok 4.5 on July 9, 2026 — its first model release since the SpaceX-Cursor merger and the post-IPO debut product. A 1.5T-parameter V9 base, Cursor-trained for coding, priced at $2/$6 per Mtok. Musk called it 'Opus-class but faster.'

SpaceXAI’s Grok 4.5 Goes Public: A 1.5T-Parameter, Cursor-Trained Model for $2/$6 Per Mtok

July 10, 2026

SpaceXAI shipped Grok 4.5 publicly on July 9, 2026 — its first model release since the SpaceX-Cursor acquisition and the debut product from the post-merger company. The architecture: a 1.5-trillion-parameter V9 foundation, post-trained on Cursor data for coding, app-building, and long-horizon agentic knowledge work. The pricing: $2 per million input tokens, $6 per million output tokens (x.ai/news/grok-4-5, 2026-07-08). Elon Musk called it on X: “It is an Opus-class model, but faster, more token-efficient and lower cost” (TechCrunch, 2026-07-08).

This is the first concrete signal that the $60 billion SpaceX-Cursor merger is operational. xAI had been written off. Grok 4.5 is the reentry.

The Architecture: V9 Base, Cursor-Trained

SpaceXAI’s launch post confirms a 1.5-trillion-parameter V9 foundation model with extensive post-training on Cursor-derived coding data (x.ai/news/grok-4-5). The Cursor component is the part nobody else in the frontier has. Anthropic trains on its own internal coding traces. OpenAI trains on GitHub-scale public code. SpaceXAI now trains on the production coding patterns of the 1M+ paying Cursor users — the same patterns that built Cursor’s 41%-to-26% market-share position in the paid coding-tools category over the past year.

That is a structurally different training corpus than the public-frontier peers. Cursor sits at the intersection of agentic tool use, multi-file refactors, and the user’s own repo context. Grok 4.5 is positioned as the workhorse for that workload.

The launch blog frames the model as a “specialist in financial analysis, coding, legal research, and other agentic knowledge work” — explicitly positioning it for the routine, high-volume agentic and clerical tasks that drive the most token spend in production agent deployments. Not a generalist frontier flagship. A cost-adjusted throughput model.

The Pricing: Sub-Frontier Tier for Closed Weights

The pricing math is the story. $2/$6 per Mtok puts Grok 4.5 in the slot below GPT-5.6 Sol ($5/$30) and Claude Opus 4.7 ($5/$25), and competitive with the sub-frontier tier:

Model Input $/Mtok Output $/Mtok Open/Closed
Grok 4.5 $2.00 $6.00 Closed
GPT-5.6 Sol $5.00 $30.00 Closed
GPT-5.6 Terra $2.50 $15.00 Closed
GPT-5.6 Luna $1.00 $6.00 Closed
Claude Opus 4.7 $5.00 $25.00 Closed
GLM-5.2 ~$0.40 ~$1.20 Open

Grok 4.5 lands at 2.5x cheaper than Sol on input, 5x cheaper on output. It costs 60% of Opus 4.7’s input price. And it’s closed-weights — there is no self-host path, no fine-tune freedom, no open-source alternative that matches the training data (Axios, 2026-07-08).

SpaceXAI is also claiming “twice greater token efficiency” versus competing frontier models — fewer tokens to hit the same answer on equivalent tasks (x.ai/news/grok-4-5). If the claim holds, the effective cost ratio against Sol is closer to 5x cheaper per completed task, not 2.5x cheaper per token. That is the kind of claim worth pressure-testing, but the launch benchmarks shown in the blog are competitive rather than best-in-class — the value proposition is cost-adjusted throughput, not raw benchmark leadership.

The Musk Quote and What It Signals

Musk’s framing matters. Two quotes from his X posts around the launch, both captured in the TechCrunch and Axios coverage:

  • “It is an Opus-class model, but faster, more token-efficient and lower cost.”
  • “Roughly comparable to Opus 4.7, but much faster.”

This is not “we beat Opus.” This is “we are within reach of Opus, but at a different price-performance point” (TechCrunch, 2026-07-08). That positioning is deliberate. SpaceXAI is not trying to win the flagship-frontier benchmark war with GPT-5.6 Sol or Opus 4.7. It is trying to own the agentic-throughput tier — the volume layer that the flagship labs have been ceded to open-source competitors.

The strategic context: Cursor’s market share dropped from 41% to 26% over the past year. Anthropic Claude Code now controls ~50% of the paid coding-tools category. The $60B Cursor acquisition was an explicit bet that buying the distribution plus the training data could rebuild a frontier-relevant model layer from scratch (TopClanker 2026-06-17). Grok 4.5 is the first product release out of that bet.

What It Means for Teams Picking Models This Week

Three practical takeaways for anyone running model-selection in production:

1. The sub-$3 input tier is now the default frontier floor. Grok 4.5 joins GPT-5.6 Luna ($1), Meta Muse Spark 1.1 ($1.25), and GPT-5.6 Terra ($2.50) in the sub-$3 input tier. The flagship tier ($5/$25 and above for Sol and Opus) is now reserved for tasks where the marginal quality lift justifies the 5-10x cost premium. For agentic loops, RAG pipelines, and bulk transformation work, the sub-$3 tier is the production default going forward.

2. Coding-trained models now have a Cursor-shaped moat. Grok 4.5’s training corpus is structurally different from anything Anthropic or OpenAI trained on. Whether that translates to a sustained benchmark lead on real-world coding tasks is a 30-day empirical question. The interesting comparison is not Grok vs Opus on Terminal-Bench — it is Grok 4.5 on real production repos versus the prior Cursor default routing.

3. The closed-weights frontier is consolidating. Meta’s Muse Spark 1.1 (closed, $1.25), SpaceXAI’s Grok 4.5 (closed, $2.00), GPT-5.6 (closed, $1–$5), Claude Opus (closed, $5) — every hyperscaler and post-IPO frontier play in the past 30 days has shipped closed-weights. The open-weights frontier (GLM-5.2, Kimi K2.7, the Qwen3 series) is now a structurally separate category competing on price and self-host economics, not on benchmark parity with the closed tier. Plan accordingly.

What to Watch Next

Two things to monitor over the next 30 days:

  • Cursor’s default model routing. The first Cursor changelog that flips any default toward Grok is the real product signal that the $60B is doing what SpaceX said it would. If the routing stays “user’s choice,” the model release is a credibility play but not a flywheel move.
  • Grok 4.5 on production coding benchmarks. Cursor-internal benchmarks matter, but Terminal-Bench 2.1, SWE-Bench Multilingual, and FrontierCode 1.1 head-to-heads against Opus 4.7 and GPT-5.6 Sol are the data points that determine whether the “Opus-class but faster” claim survives contact with independent evaluation. The ExplainX and Mashable coverage both note that SpaceXAI’s published benchmarks are competitive but rarely best-in-class on any single test (ExplainX, 2026-07-08; Mashable, 2026-07-08). The product claim is cost-adjusted capability, and that needs real workload evidence.

Bottom Line

Grok 4.5 is the first concrete proof that the $60B SpaceX-Cursor acquisition is operational. It is a Cursor-trained, closed-weights, sub-frontier-priced model positioned for the agentic-throughput tier rather than the flagship-frontier benchmark war. Musk’s framing — “Opus-class but faster” — is a deliberate positioning choice, not a benchmark claim. The pricing alone ($2/$6 with a claimed 2x token efficiency) makes it the default routing target for any agentic workload where token volume matters more than the last 2% of capability.

For teams running model routers today: add Grok 4.5 to the candidate set, run it on your production coding eval before the Cursor changelog starts shifting defaults, and watch the routing decisions over the next 30 days. The model release is the story; the routing decisions are the consequences.

Sources