AI news · July 9, 2026

Grok 4.5 is the first fruit of the Cursor acquisition — and it's priced to win

SpaceXAI's $60B Cursor deal just produced a co-trained frontier model. Grok 4.5 is fast, cheap, and live in Cursor — but you don't have to use Grok to benefit.

Read time: ~7 minutes

A few months ago Grok felt like the punchline in a frontier-model conversation. Cursor was building its own Composer line. Elon Musk was reportedly renting compute to rival labs. SpaceXAI's place in the stack looked uncertain.

Then SpaceXAI closed its $60 billion acquisition of Cursor, and the two teams shipped Grok 4.5 — the first model they trained together. Benchmarks land in the same neighborhood as Claude Opus 4.8 and GPT-5.5. Serving speed hits roughly 80 tokens per second. Pricing is $2 per million input tokens and $6 per million output — about six times cheaper than Opus on a per-token basis and roughly three times cheaper than GPT-5.5 for comparable agent work, per SpaceXAI's launch messaging.

I'm writing this because I still haven't fully moved my daily workflow into Cursor — and that's embarrassing given I published a whole Replit → Cursor migration guide. Grok 4.5 is the push I needed. Even if you never select Grok in the model picker, Cursor is where the action is: Composer 2.5, Opus 4.8, Fable 5, and GPT-5.6 as OpenAI widens access.

Benchmarks: competitive, not dominant

SpaceXAI published a four-benchmark chart comparing Grok 4.5 against Opus 4.8, GPT-5.5, Composer 2.5, and Fable 5. The honest read: Fable 5 still tops several coding benches, especially SWE-Bench Pro. Grok 4.5 is closest on Terminal-Bench 2.1 and trades blows elsewhere — with a clear edge in token efficiency and cost, not raw peak scores.

On SWE-Bench Pro, SpaceXAI reports Grok resolving tasks with roughly 4.2× fewer output tokens than Opus 4.8 — about 16k tokens versus 67k on average. That's the business model: near-frontier intelligence at flash speed and a price that makes you stop flinching at every agent loop.

Pricing vs Opus and GPT-5.5

Grok 4.5 usage is temporarily included in Cursor and Grok Build plans, with doubled limits for the first week on Cursor individual and team tiers. API access is live via the SpaceXAI console. EU readers: Grok 4.5 was geo-blocked in the European Union at launch, with mid-July 2026 availability expected.

The Cursor codebase caveat

Cursor's own launch post notes that an early snapshot of the Cursor codebase was accidentally included in Grok 4.5 training data. That may inflate scores on Cursor-specific benchmarks. SpaceXAI says the data is removed for future models and Cursor is updating its internal bench. Worth knowing before you treat every number as gospel — the cross-vendor table above is the fairer read.

Cursor as the hub — Grok optional

The acquisition story isn't "you must use Grok." Cursor remains a multi-model IDE. Today's picker includes:

• Grok 4.5 — new default curiosity for agentic coding at low cost
• Composer 2.5 — Cursor's in-house model, climbing the coding ladder
• Claude Opus 4.8 and Fable 5 — still the peak for several SWE benches
• GPT-5.6 Sol, Terra, Luna — rolling out as OpenAI opens access beyond partner lists

Don't like Grok? Pick something else. The win is having frontier-class options in one editor with repo context, terminal access, and agents that actually touch your files.

Cursor on iOS — finally mobile agents

Easy to miss in the Grok noise: Cursor shipped an iOS app in this window. You get mobile access to agents, chat, and codebase-aware help from your phone — review a diff on the couch, kick off a refactor before you reach your desk, same account as desktop. For vibe coders who live in messages and notifications, that's a real workflow upgrade, not a marketing slide.

Quick hits from the same week

• ChatGPT Voice refresh — OpenAI rolled out GPT-Live-1 and Live-1-mini with a new architecture that lets the model respond while you're still speaking (no more hard cutoffs) and routes harder reasoning to GPT-5.5 in the background. Voice mode finally feels like a generation jump.
• Fable 5 on Claude subscriptions — Anthropic extended Fable 5 access for Claude subscribers until July 12, still capped at up to 50% of weekly limits; after that, prepaid usage credits apply. Useful if you're A/B testing Fable against Grok in Cursor.
• GPT-5.6 going wider — OpenAI said Sol, Terra, and Luna would reach general users shortly after the partner-only phase. Early tester chatter on X leans toward a split verdict: Sol may not feel "smarter" than Fable on every task, but more reliable on long agent runs — fewer loops, fewer silent failures. See our GPT-5.6 breakdown for the gated-launch context.

Also worth bookmarking

• Grok / Grok Build — SpaceXAI's agent surface; Grok 4.5 is default there
• Claude — Fable 5 until July 12 on subs
• ChatGPT — GPT-Live-1 voice + widening GPT-5.6 access
• OpenRouter — swap models without rewiring APIs

Full tool list with referral links: vibecoderslife.com/tools

Bottom line

Grok 4.5 clears SpaceXAI's name in the frontier tier: Opus-class territory on several benches, flash serving speed, and pricing that punishes expensive defaults. The Cursor acquisition already pays rent in both directions — Cursor gets a flagship partner model; SpaceXAI gets distribution inside the editor vibe coders actually use.

Peak benchmark crowns still sit on Fable 5 for some tests. Grok's bet is that good enough + fast + cheap wins the year everyone counts tokens. Me? I'm clicking the Cursor link, opening my repo, and catching up.