Grok 4.5 Goes Public: SpaceXAI's First Ship Since the Rebrand

What changed

• xAI no longer exists as a standalone company. SpaceX acquired it in an all-stock deal on 2 February 2026, and the merged businesses now trade and operate as SpaceXAI, following an official rebrand and new logo in early July.
• Grok 4.5 is SpaceXAI's first ship under the new name — released 8 July 2026 and reaching general availability the next day, across the API, Grok Build, Cursor, and several model gateways.
• It is priced well below the frontier pack — $2 per million input tokens and $6 per million output tokens, against Claude Opus 4.8's $5/$25.
• The benchmarks are genuinely strong, not chart-topping. Grok 4.5 ranks fourth on the Artificial Analysis Intelligence Index, behind Fable 5, GPT-5.5 and Opus 4.8 — but well ahead of every open-weight model and Google's Gemini line.
• EU availability is delayed, pending EU AI Act General-Purpose AI compliance work, with SpaceXAI targeting mid-July without committing to a firm date.

From xAI to SpaceXAI: the fastest corporate rebrand of the year

The paperwork came first. On 2 February 2026, SpaceX and xAI announced an all-stock merger valuing the combined entity at roughly $1.25 trillion, structured as a straight share exchange rather than a cash deal. Musk framed the logic around infrastructure: xAI needed capital to keep pace with OpenAI and Anthropic on data-centre build-out, and SpaceX's satellite and launch business gave the combined group a route to what Musk has reportedly framed as orbital data centres. Tesla was reportedly not part of the deal.

What followed was months of quiet integration — xAI's Grok chatbot and the X social platform folded operationally into SpaceX — before the cosmetic rebrand landed in early July 2026. The new SpaceXAI logo blends the stylised "A" from SpaceX's wordmark into the original xAI mark, and the company has been explicit that this is not a joint venture or a licensing arrangement: xAI, as a distinct corporate entity, no longer exists.

The merger also explains a second deal that AI Tech Connect covered in June: SpaceX's $60 billion agreement to acquire Cursor. That deal — struck as an all-stock transaction and expected to close in Q3 2026 pending regulatory approval — built on a co-development partnership the two companies had already been running since April. Grok 4.5 is the first tangible output of that partnership: SpaceXAI's own materials describe it as the first model trained in collaboration with Cursor, using real developer session data rather than synthetic coding traces.

Grok 4.5, unpacked: what's actually new

Under the bonnet, Grok 4.5 runs on what SpaceXAI calls its V9 foundation architecture at roughly 1.5 trillion parameters, structured as a mixture-of-experts rather than a dense model, with a 500,000-token context window. It was trained across a large-scale GPU cluster on datasets weighted towards coding, science, engineering and mathematics — consistent with SpaceXAI's stated aim of building a model that is "smartest at coding, agentic tasks and knowledge work," with an explicit push into legal and financial-services use cases as well.

The independent numbers, via Artificial Analysis, are the most useful reference point:

On the Intelligence Index, Grok 4.5 now sits ahead of every open-weight model and all of Google's Gemini family, though still behind the three highest-scoring proprietary labs. On the separate Coding Agent Index, which measures how a model performs inside an actual harness rather than on isolated questions, Grok 4.5 running in Grok Build scores 76: it matches GPT-5.5 running in Codex, and trails Fable 5 running in Claude Code by a narrow margin. On GDPval-AA v2, a benchmark for sustained agentic knowledge work, Grok 4.5 posted an Elo of 1,543, sitting between Opus 4.8 and GLM-5.2.

The pricing math: why $2/$6 matters more than the leaderboard rank

The headline number that will actually move routing decisions is cost, not rank. At $2 input / $6 output per million tokens, Grok 4.5 undercuts Claude Opus 4.8's $5/$25 by a wide margin — roughly 75% cheaper on output tokens, which dominate cost in most agentic and long-form workloads.

Artificial Analysis puts a concrete number on what that means in practice: a single task on the Intelligence Index costs about $0.31 to run on Grok 4.5, reportedly around a fifth of the equivalent cost on Claude Sonnet 5 at its higher reasoning setting. On the Coding Agent Index, a full agentic coding task averages roughly $2.49 — not the cheapest option on the market, but meaningfully below Opus-tier pricing for a model that lands just one rung below it on agentic benchmarks.

Model-router decisions: what India and UK teams should actually do

For teams in Bengaluru, Pune, London or Manchester running production LLM traffic, the practical questions aren't about who's "winning" the leaderboard — they're about where a new, cheaper, near-frontier model earns a slot in an existing routing table. A few concrete starting points:

• Run your own coding-agent eval before switching. The Coding Agent Index result (tied with GPT-5.5) is encouraging, but it's a public benchmark, not your codebase. Indian fintech and UK legaltech teams in particular should re-run their existing SWE-style internal evals against Grok 4.5 before moving default traffic.
• Budget for the EU gap if you serve UK and EU customers together. A UK-based team serving both domestic and EU users cannot yet point EU-resident inference at Grok 4.5 through any access point. Keep a fallback model in the router for that segment until SpaceXAI confirms EU availability.
• Treat cache economics as a design decision, not an afterthought. The $0.50/M cached-input rate is central to Grok 4.5's cost story. Agent loops that don't warm a stable prompt cache will not see anything close to the $0.31-per-task figure quoted in public benchmarks.
• Watch the M&A overhang. SpaceXAI's pending $60 billion Cursor acquisition is still working through regulatory approval, expected to close in Q3 2026. Teams building on Cursor's API or SDK should keep an eye on how deeply Grok 4.5 gets wired into Cursor's own product roadmap once that deal closes.
• Don't over-index on the "Opus-class" line. It's a real quote from Musk, but it describes efficiency and price, not benchmark parity. Opus 4.8 still outranks Grok 4.5 on both the Intelligence Index and GDPval-AA.

This is exactly the kind of before/after benchmark comparison worth documenting if you're the person running the eval inside your team — a written record of which model won which workload, and why, is the sort of concrete, dated project detail that stands out on a Verified Builder profile far more than a generic "worked with LLMs" line.

What to watch next

Three things are worth tracking over the next few weeks. First, whether SpaceXAI hits its self-declared mid-July target for EU access, or whether — like several frontier labs before it — the delay stretches on while GPAI compliance documentation is finalised. Second, whether the pending Cursor acquisition closes on schedule in Q3 2026, and what that means for how tightly Grok 4.5 gets integrated into Cursor's own agent tooling versus staying a general-purpose API option. Third, SpaceXAI has signalled that its own inference stack — a C/C++ implementation designed to map directly onto GB300 hardware — hasn't shipped yet, and the company expects a further speed step-change once it does. None of that changes today's routing maths, but all of it is likely to move the numbers in this article within a quarter.

For now, Grok 4.5 is a legitimate, well-priced fourth option at the frontier — not a leaderboard-topper, and not yet usable for EU-resident production traffic, but a serious addition to any model router that already juggles Opus, GPT and Gemini traffic by cost and task type.

Primary sources: SpaceXAI's Grok 4.5 announcement, Cursor's launch post, and Artificial Analysis's benchmark write-up.