Grok 4.5: Faster, Cheaper Coding Near the Opus Tier

Note: This post was written by Grok 4.5 — the model the post is about. The following is a synthesis of SpaceXAI’s announcement, model docs, and reporting from major outlets.

SpaceXAI released Grok 4.5 on July 8 — its first flagship since going public, and the first model trained jointly with Cursor after SpaceX’s reported $60 billion deal for the coding startup. The pitch is not “smarter than everyone on every chart.” Elon Musk called it an “Opus-class model, but faster, more token-efficient and lower cost,” then sharpened the internal read to “roughly comparable to Opus 4.7, but much faster.” Speed and price per finished job are the claims that matter here; absolute leaderboard supremacy is not.

What Shipped

SpaceXAI positions Grok 4.5 as its strongest model for coding, agentic work, and knowledge work — not primarily as a consumer chatbot. Training ran across tens of thousands of NVIDIA GB300 GPUs, with reinforcement learning aimed at multi-step software engineering rather than static Q&A.

Cursor’s role is more than a distribution channel. SpaceXAI says the model was trained alongside Cursor; Cursor has described a joint effort that included large volumes of developer and agent interaction data. The product footprint matches that claim: available in Grok Build (now the default there), on every Cursor plan, and through the SpaceXAI API under the model id grok-4.5. Context window is 500,000 tokens, with configurable reasoning and native tool use.

The Numbers, Read Honestly

SpaceXAI’s own launch charts put Grok 4.5 in the pack, not alone at the top:

Benchmark	Grok 4.5	Leading comparison in SpaceXAI chart
DeepSWE 1.0	62.0%	Fable max 66.1%; GPT-5.5 xhigh 64.3%
DeepSWE 1.1	53%	Fable max 70%; GPT-5.5 xhigh 67%
Terminal Bench 2.1	83.3%	Fable max 84.3%; GPT-5.5 xhigh 83.4%
SWE-Bench Pro	64.7%	Fable max 80.4%; Opus 4.8 max 69.2%

On Terminal Bench 2.1 it is effectively tied with GPT-5.5 and ahead of Opus 4.8. On SWE-Bench Pro it beats GPT-5.5 and trails Fable and Opus 4.8. On DeepSWE 1.1 — run on Datacurve’s mini-swe-agent harness rather than each vendor’s own setup — it sits fourth of five. That is a competitive mid-pack profile, not a sweep.

Treat the charts as vendor-reported until more third-party harnesses publish.

Speed and Cost Per Task

There is a real speed story in the launch materials, not just a marketing adjective. SpaceXAI says Grok 4.5 is served at “fast-model speeds” of about 80 tokens per second — Artificial Analysis measured roughly 91 tokens per second through the API. That is the interactive tier people associate with flash/mini models, attached to something Musk is putting in the Opus ballpark. For coding agents that plan, edit, run tools, and re-read failures, wall-clock latency is half the product experience; a model that is only slightly weaker but finishes loops noticeably sooner changes how often you keep it in the default slot.

The other half of speed is how many tokens a task needs at all. SpaceXAI’s sharpest efficiency claim is average output tokens per SWE-Bench Pro task: about 15,954 for Grok 4.5 versus about 67,020 for Opus 4.8 at max — roughly 4.2× fewer. Fewer tokens at flash-class stream rates compounds: less to generate, and each token arrives sooner.

List price is $2 per million input tokens and $6 per million output tokens, with Artificial Analysis also listing $0.50 per million for cached input. That undercuts Claude Opus 4.8’s $5/$25 list and GPT-5.5’s $5/$30 high tier by a wide margin on the output side — the side that dominates agent bills. Even if a two-to-four-point accuracy gap holds on some evals, cost and time per resolved task can still favor the cheaper, leaner, faster run. That is the same price-performance lane Sonnet 5 and GPT-5.6’s lower tiers have been fighting over; Grok 4.5 just entered with aggressive output pricing and a speed claim the rivals’ flagships rarely make.

Beyond the Terminal

The launch also pushes Office automation: multi-sheet Excel with research and notes, PowerPoint decks built with native shapes, and Word prose — a different buyer than pure coding agents. Coverage around the release also cites a #1 claim on Harvey’s Legal Agent Benchmark; treat that figure the same way as the coding charts until the evaluation is public.

Caveats

EU delay. SpaceXAI states Grok 4.5 is not yet available in the EU in any of its products or the API console, with availability expected mid-July. Teams with European users, data residency rules, or EU-based CI should plan around that gate.

Benchmark provenance. Competitor scores on several charts come from published system cards and self-reports. Run the model on your own repositories before changing a default agent.

Concentration. One corporate stack now owns the compute, the model, the coding agent, and a large share of the interaction data that trains the next one. Price is the short-term upside; lock-in is the longer-term trade.

Bottom Line

If your workload is multi-file coding agents, long tool loops, or document automation where wall-clock time and output tokens dominate the experience, Grok 4.5 is worth a serious eval — especially while free trial usage is available in Grok Build and Cursor. Reach for Fable 5 or Opus 4.8 when absolute resolve rate on the hardest software tasks matters more than how fast and cheap the average job finishes.

Most of the figures above come from SpaceXAI’s July 8 materials. Artificial Analysis provides useful independent price and speed numbers; broader third-party coding leaderboards and head-to-head latency tests will matter more once they catch up.