Frontline Hotspot
Frontline Hotspot

Grok 4.5: SpaceXAI's First Coding+Agent Model, 1.5T and Co-Trained with Cursor

SpaceXAI released Grok 4.5 on July 8, 2026 - a 1.5T V9 model co-trained with Cursor, its first built specifically for coding and agents, priced $2/$6 with a 500K context and pitched internally as "comparable to Opus 4.7 but much faster"; no public benchmarks exist, so it positions as a cheap Opus-class substitute native to Cursor, differentiated from Opus 5 / GPT-5.6 Sol on ecosystem rather than leaderboard rank.

Published July 31, 20265 min read
<!-- grok-4-5-hotspot | hotspot | Grok 4.5: SpaceXAI's First Coding+Agent Model, 1.5T and Co-Trained with Cursor -->

On July 8, 2026, SpaceXAI publicly released Grok 4.5 - described in its own words as "our first model trained specifically for coding and agents," and the first product of the data merger that followed the Cursor acquisition. Two numbers grab the eye: 1.5 trillion parameters, and co-training with Cursor. Both need caveats. The 1.5T figure is a Musk disclosure reported by third parties; SpaceXAI has never officially confirmed it. "Co-trained with Cursor" doesn't just mean "fed Cursor's code" - it means Cursor's real-world development data was baked into a freshly pre-trained V9 foundation. To read this launch clearly, separate the facts, the specs, the capability ceiling, and where it actually stands.

Facts and Specs: 1.5T, V9, and a 500K Context Window

What's confirmed. Grok 4.5 went public on July 8, 2026, after roughly ten days of private beta inside SpaceX and Tesla (Musk first disclosed it on X on June 28). It shipped in three places: the Cursor editor, Grok Build (SpaceXAI's own coding-agent environment), and the SpaceXAI console - with Cursor offering double usage for the first week. EU access came later in July.

On specs, Grok 4.5 runs on a new V9 foundation. Multiple third-party reports cite Musk putting it at roughly 1.5 trillion parameters, nearly 3x the prior V8 - but note SpaceXAI has never officially confirmed this number; it is vendor-stated and media-relayed, not independently verified. What is confirmed is pricing and context: $2 per million input tokens, $6 per million output, with cache hits discounted 75% to $0.5/M; inputs longer than 200K tokens are billed at double. The context window is 500K - half of Grok 4.3's 1M - but vision input and configurable reasoning are retained; Musk said on X it would upgrade back to 1M. One detail easy to miss: xAI has merged into SpaceXAI, so the product entity and brand are both SpaceXAI, though the launch tweet still went out from the original xAI account - treat "xAI" and "SpaceXAI" as the same company when reading coverage.

Coding+Agent Capability: What Cursor Co-Training Actually Means

The "first model trained specifically for coding and agents" framing matters at the training stage, not inference. Cursor is the dominant AI coding editor; its parent Anysphere was acquired by SpaceX for about $60 billion, with the deal expected to close in Q3 2026 pending regulatory approval. Grok 4.5 folded Cursor's real development data - multi-file edits, debugging, long engineering chains - into a supplemental training stage on V9. SpaceXAI's claim is that this gives it "a structural advantage on coding-heavy tasks no other frontier model can currently match." Cursor itself called Grok 4.5 "our most powerful model yet" and stressed it was "the first we've built for more than software engineering."

Be honest, though: as of now there is no independent third-party benchmark (SWE-Bench, Terminal-Bench, etc.) with public Grok 4.5 scores. Musk's internal framing was "roughly comparable to Opus 4.7, but much faster" - note Opus 4.7, not Opus 5, which only launched on July 24. In other words, at launch Grok 4.5 was benchmarked against the prior Opus generation; its pitch is "Opus-class work at half the tokens and under half the price, and faster" - not "number one on the leaderboard." This is a different path from Anthropic and OpenAI's benchmark chasing; SpaceXAI chose the more utilitarian yardstick of "useful to SpaceX and Tesla engineers." So don't compare it to Opus 5 on SWE-Bench - it simply hasn't published a run.

One more thing often glossed over: the official line is "coding and agents," not coding alone. Cursor's own phrasing - "the first we've built for more than software engineering" - hints that its agentic capability (driving multi-step, multi-file changes inside the editor, running tests and looping back to fix) is the real pitch, not just code completion. But agentic capability likewise has no public Agentic Index or Terminal-Bench score to compare against, so here too you can only trust the vendor's word and test it yourself.

Competitive Position: vs Opus 5, GPT-5.6, and Claude Code

Place Grok 4.5 in the late-July landscape and its position is "a cheap Opus-class substitute with no public benchmarks." On price, $2/$6 is far cheaper than Anthropic's Opus 4.8 at $5/$25; against Opus 5 (also $5/$25, released July 24), input is 60% cheaper and output 76% cheaper. Versus GPT-5.6 Sol, input is level ($2) and output is slightly lower ($6 vs $8). But be clear: Opus 5 tops Artificial Analysis at 61 on the Intelligence Index and 55.3 on the Agentic Index; GPT-5.6 Sol sits at 59 on intelligence and, paired with Codex, leads the Coding Agent Index outright at 80 - Grok 4.5 has no comparable public score on any of these. In other words, on "per-million-token price" it is the cheapest; on "benchmark rank" it isn't on the board at all. And cheap per-token is not the same as cheap per-task - as the Opus 5 writeup noted, at max effort it can burn 8x the tokens on a single task. Grok 4.5 claims to "use half the tokens for the same work," which if true would lower per-task cost further, but that is a vendor claim, not third-party-verified; measure "total spend to finish one real task" yourself before concluding.

The real differentiator is the Cursor axis. Opus 5 and GPT-5.6 show their coding muscle paired with Claude Code and Codex respectively; Grok 4.5 is "native to Cursor" - it trained on Cursor's data and shipped inside the Cursor editor on day one. For teams already using Cursor as their primary IDE, switching to Grok 4.5 is nearly frictionless; for those on Claude Code, the migration cost is elsewhere. So this isn't "Grok 4.5 beats Opus 5" - it's "if you're already in the Cursor ecosystem, it's the smoothest path." SpaceXAI's bet is equally visible: use Cursor's data flywheel to lock in developers, and the next model (reportedly already training, around 2T parameters, with Cursor data wired in from pre-training) will dig that moat deeper.

What It Means for Developers

For developers, Grok 4.5 means "a cheap, editor-native option has joined the coding-agent field." If your work is high-volume multi-file editing and long debugging chains, $2/$6 plus the 75% cache discount compounds at batch scale; the 500K context, though shrunken, is enough for most single tasks, and long-context work can wait for the 1M upgrade. Three caveats: first, with no public benchmarks, don't let leaderboard scores drive your decision - measure "cost per task" and "error rate" on your own real workload; second, long inputs past 200K tokens bill at double, so cost out long-context runs first; third, EU and API channel rollout is still on the official schedule, and the most stable entry points today are Cursor and Grok Build. One more thing to keep in mind: binding your main workflow to Cursor plus Grok 4.5 means handing your data flywheel to SpaceXAI - the next model will be stronger and you'll be more locked in, which is the whole point of the Cursor acquisition. For small teams, use it, but don't feed all your engineering data into this one chain.

For everyone else, this launch is mostly "another entrant in the AI coding arms race." An extra model in Cursor is a net positive for subscribers; but if you don't write code, Grok 4.5's coding focus has little to do with your daily Q&A. The test is simple: do you normally work inside Cursor or a similar agent environment? If yes, it's worth a one-week trial; if no, wait for its capabilities to spill into general chat. "Fast and cheap" sounds tempting, but it comes down to one thing: can it, on your hardest coding task, make fewer mistakes and take fewer round-trips than the model you use now? If yes, the switch is worth it.


References

This article is AI-assisted and human-edited. Last updated: 2026-07-31

Related

Frontline Hotspot

Tops the Chart, Loses One Test: Is Grok Voice 3.0 Worth It

xAI released Grok Voice 3.0 on 2026-10-15, billed officially as the world's most intelligent speech-to-speech model, with end-to-end latency under one second and API pricing unchanged from the previous generation: the Voice Agent API at USD 0.08 per minute (USD 4.80 per hour) plus USD 0.004 per text input unit, with grok-voice-latest auto-upgrading so integrators change no configuration. All comparison baselines are quoted from xAI's own page citing Artificial Analysis (2026-10-13) and were not independently reproduced by this site. On the AA Speech-to-Speech Index, Grok Voice 3.0 tops the chart at 86.9, above Grok Voice 2.0 at 81.3, GPT-Live-1 (Astra, Medium) at 83.2 and Gemini 3.8 Live (Extended Thinking, High) at 82.6; on all four sub-scores it leads in that same order - Big Bench Audio 98.5/97.2/90.1/97.7 percent, Speech Agent Arena Elo 1070/1011/1048/990, Task Success Rate 94.8/94.6/87.4/89.1 percent, and Full Duplex Bench 97.0/95.1/94.9/91.9 percent. But tau-Voice, the customer-service task success rate under accents, background noise and interruptions, stands at 71.6 percent, below GPT-Live-1 at 74.5 percent (Gemini 3.8 Live 68.6, the previous generation 56.5) - first overall is not first everywhere, and the article says so plainly. What it buys is first-audio latency of 0.81s against GPT-Live-1's 1.34s, about 40 percent lower, while its own predecessor was faster at 0.70s: this generation lifted tau-Voice from 56.5 to 71.6 percent at the cost of 0.11s of time-to-first-audio, a trade-off rather than a pure gain. On cost, the AA Big Bench Audio figure is USD 0.025 per task against GPT-Live-1's 0.030 and GPT-Realtime-2.1 (High) at 0.055; on the enterprise side, EVA-Bench (from ServiceNow, over 200 real enterprise scenarios) is officially claimed as the highest at 74 percent, with the best joint accuracy and customer-experience score. The deployment numbers are all vendor statements without third-party audit: millions of in-car conversations a day at Tesla, roughly 20,000 multilingual support calls a day at Starlink with about 70 percent needing no human, and about one in four sales calls closing before hang-up. Division of labor with this site's October 7 GPT-Live-1 launch piece and its API SOP: this is a competitor's generational release, so it does not retell that launch or repeat its integration steps, only compares scoreboards and stated figures.

Oct 16, 20268 min read
Frontline Hotspot

AI Coding Agents in August 2026: Three Camps, Each Its Own Pole

By August 2026, AI coding agents have settled into three camps: browser turnkey online platforms (Replit/Bolt/Lovable), local IDEs deep-integrated with codebases (Cursor/Copilot/Trae), and autonomous terminal CLI agents (Claude Code/Codex/Cline). The camps are not tiers but different ranges; the rule is run it first, optimize later. Underneath all is the same context-execute-verify loop; the real barrier is task-decomposition skill.

Aug 5, 20266 min read