Update — July 22, 2026: Cursor restated 2× included usage for Grok/Composer on individual and Teams plans — permanent first-party pool, not a second double. Grok 4.5 50% launch discount ended July 21: usage limits clarified.
Update — July 17, 2026: Lee Robinson's AI Engineer talk explains the training flywheel behind Grok 4.5 — two-loop recursive improvement, Colossus compute, textual feedback RL — teased on stage days before this launch.
Update (July 9, 2026 — Cursor official blog): Cursor and SpaceXAI released Grok 4.5 in the IDE on July 8 — not just a Musk tweet for July 9. Mixture-of-experts model, trillions of Cursor interaction tokens in training, cyber safeguards, and published agentic benchmarks. GPT-Live and SWE-1.7 shipped the same night.
Update — July 16, 2026: SpaceXAI open-sourced the Grok Build harness (Apache 2.0), reset usage limits, and advertised local-first inference — full guide.
Security update — July 14: Grok Build privacy crisis — wire tests captured full-repo uploads; SpaceXAI documented ZDR and /privacy; Andrew Milich (ex-Skiff) confirmed retroactive delete for all users; Elon Musk pledged total deletion of all data uploaded before July 14 ("zero anything whatsoever will remain"). Full evidence and limits: Grok Build upload report.
On July 8, 2026, Elon Musk posted that SpaceXAI will release Grok 4.5 to the public tomorrow — July 9 — after strong positive feedback from beta customers. His pitch in one line: an Opus-class model, but faster, more token-efficient, and lower cost.
Hours earlier, Cursor's engineering blog confirmed Grok 4.5 is already live inside Cursor — the first jointly trained SpaceXAI + Cursor model aimed at more than software engineering alone.
That lands in the same week Anthropic is fighting a trademark lawsuit from Abnormal AI, OpenAI is pushing GPT-5.6 Sol toward general availability, and frontier-model releases are stacking faster than most teams can benchmark them. Grok 4.5 is not a surprise drop — it graduates from the private beta we covered on June 28. The news is the public date and Musk's efficiency framing.
TL;DR: What People Are Asking
| Question | Answer |
|---|---|
| In Cursor now? | Yes — July 8, 2026 desktop, web, iOS, CLI, SDK |
| xAI public? | July 9, 2026 per Musk / SpaceXAI tweet |
| Architecture? | MoE — joint Cursor + SpaceXAI training |
| Terminal-Bench 2.1? | 83.3% (GPT-5.5: 83.4%, Fable 5: 84.3%) |
| SWE-Bench Pro? | 64.7% high (Fable 5: 80.3% max) |
| API price (Cursor)? | $2/M in, $6/M out — fast: $4/$18 |
| Composer 2.5? | Stays offered — different weight class |
| CursorBench? | Excluded — accidental Cursor repo in training |
Cursor Benchmarks (July 8, 2026)
| Benchmark | Grok 4.5 | Opus 4.8 | GPT-5.5 | Composer 2.5 | Fable 5 |
|---|---|---|---|---|---|
| Terminal-Bench 2.1 | 83.3% | 78.9% | 83.4% | 73.0% | 84.3% |
| SWE-Bench Multilingual | 78.0% | 84.4% | 77.8% | 71.6% | — |
| DeepSWE 1.0 | 62.0% (high) | 55.8% (max) | 64.3% (xhigh) | 18.0% | 66.1% (max) |
| SWE-Bench Pro | 64.7% (high) | 69.2% (max) | 58.6% (xhigh) | 54.0% | 80.3% (max) |
Third-party scores on SWE-Bench Pro and Terminal-Bench are self-reported per Cursor footnotes. GPT-5.5 multilingual from Cursor internal run.
Honest caveat: CursorBench contamination
Cursor disclosed that an earlier Cursor codebase snapshot was accidentally included in training. Grok 4.5 may have had an advantage on CursorBench; exact impact unclear. That data is removed for future models, and CursorBench is getting a larger update — hence its exclusion from published charts. Treat in-IDE performance as optimistic until independent repro.
Cursor Blog: Training Beyond Coding
Per Cursor's July 8 post:
| Topic | Detail |
|---|---|
| Architecture | Mixture-of-experts, jointly trained with SpaceXAI |
| Data | Trillions of tokens of Cursor user interactions with codebases and tools |
| vs Composer 2.5 | Composer = coding specialist; Grok 4.5 = broader STEM, research, knowledge work |
| RL | Difficult problems in realistic environments; distributed agent system built training envs at scale |
| Cyber | New safeguards reflecting cybersecurity capabilities |
| Pricing | Base $2/M input, $6/M output; fast $4/$18 |
| Usage | Significant pool in Cursor plans; double usage first week |
| Composer 2.5 | Remains available — two weight classes, both supported going forward |
What Musk Announced (July 8, 2026)
The announcement came from Elon Musk on X, routing through SpaceXAI — the merged SpaceX/xAI entity that has been shipping models on an aggressive cadence since the xAI acquisition closed in February 2026:
"Based on strong positive feedback from customers in our beta test program, @SpaceXAI will make Grok 4.5 available to the public tomorrow.
It is an Opus-class model, but faster, more token-efficient and lower cost."
Three claims to track separately:
- Capability tier — "Opus-class" positions against Claude Opus 4.8, Anthropic's top reasoning tier before Fable 5's export-control suspension
- Speed — latency and throughput vs. Opus, not just benchmark accuracy
- Economics — token efficiency and price — the pitch Musk uses when competing with both Anthropic and OpenAI on developer wallets
Musk did not publish benchmark numbers, API pricing, or context-window specs in the tweet. Those details typically follow on launch day.
The Build: 1.5T V9 + Cursor Data
This is not a minor Grok 4.x point release. The architecture was described in the June 28 beta announcement:
| Component | Detail |
|---|---|
| Base model | 1.5T parameter V9 foundation model |
| Supplemental training | Cursor IDE interaction data |
| Beta environments | SpaceX and Tesla internal engineering |
| Training infra | Colossus cluster (220K+ NVIDIA GPUs in Memphis) |
| Ongoing improvement | RL via Grok Build harness |
Cursor data matters because it captures real agentic coding sessions — multi-turn edits, large-repo context, error recovery — not static code corpora. That is the same signal distribution coding benchmarks and agent harness evals target.
Reporting from The Information, amplified by Walter Bloomberg and Andrew Curran on X, tied the July launch to the first jointly developed model from SpaceX and Cursor — expected around 1.5 trillion parameters and competitive with Opus 4.8 and GPT-5.5 in some dimensions. A delay earlier in the week was described as polishing efficiency before release.
The Cursor angle connects to the SpaceX → Anysphere (Cursor) merger disclosed in an SEC 8-K on June 16, 2026 — $60B implied equity, all-stock, expected close Q3 2026. Grok 4.5 is the first visible output of that integration strategy: model lab + IDE interaction data + massive compute under one roof.
July 9: The Busiest Frontier Day of 2026 So Far
Grok 4.5 is not launching into a quiet market.
| Release | Vendor | Timing | Positioning |
|---|---|---|---|
| Grok 4.5 | SpaceXAI / xAI | July 9, 2026 | Opus-class, faster, cheaper |
| GPT-5.6 Sol (+ Terra, Luna) | OpenAI | July 9, 2026 public launch | Flagship GPT-5.6 tier; Terminal-Bench ~91.9% on preview |
| Fable 5 | Anthropic | Relaunched July 1 for eligible users | Top-tier coding/reasoning — when available |
| GLM-5.2, Kimi K2.7, etc. | Open-weight labs | Ongoing | Fable alternatives for teams blocked on Anthropic |
One X reply captured the developer mood: "tomorrow gonna be a busy day for competitors — Fable 5, Sol, now Grok 4.5. We truly are spoilt for choice."
Another thread pointed at Anthropic specifically: "this is clear now everybody is coming for Anthropic" — a sentiment amplified by the same week's Abnormal trademark fight and ongoing Claude Max litigation. Whether Grok 4.5 actually dents Anthropic's developer share depends on benchmarks and pricing, not X bravado.
What X Is Asking: "Opus-Class — But Which Opus?"
Musk's superlative landed. The specificity did not.
Top reply themes from the July 8 thread:
| Reply theme | Why it matters |
|---|---|
| "Which Opus exactly? Hoping at least 4.7 levels" | Opus spans 4.5 → 4.8 with measurable benchmark jumps; "Opus-class" without a version is meaningless for procurement |
| "Why compare to Opus — this is Grok" | Brand positioning choice: Anthropic Opus is still the developer mental benchmark for top-tier reasoning |
| "Everybody is coming for Anthropic" | Competitive narrative stacking — legal, pricing, and model launches hitting the same week |
| Token efficiency / cost | Developers care about $/task more than parameter count — Musk lead with economics, not SWE-bench |
This is the same verification gap we flagged in the June beta coverage: internal evals at SpaceX and Tesla are not public benchmarks. Until xAI publishes scores on SWE-bench, Terminal-Bench, GPQA, or similar suites, "Opus-class" is a claim to test, not a spec to buy against.
For reference, Claude Opus 4.8 hit 69.2% on SWE-bench Pro and reduced code-flaw pass-through rates vs. 4.7. GPT-5.6 Sol preview builds reported ~91.9% on Terminal-Bench 2.1. Grok 4.5 needs numbers in that neighborhood to match the marketing.
What to Watch on Launch Day (July 9, 2026)
- Public benchmarks — Does SpaceXAI publish SWE-bench, HumanEval, or agentic scores alongside the release?
- API pricing and context window — Musk emphasized cost; developers need $/M tokens and max context
- Grok on X vs API — Which tiers get Grok 4.5? SuperGrok subscribers first, or immediate API access?
- Cursor integration — Will Grok 4.5 appear as a selectable model in Cursor, or stay xAI-native?
- Head-to-head with Sol — Same-day launch invites direct comparison on coding tasks, latency, and price
- Monthly model cadence — Musk previously said SpaceXAI plans new from-scratch models monthly through 2026; Grok 4.5 is the first public proof point
Honest Limitations
This post is based on Musk's July 8 announcement, June beta details, and press reporting — not hands-on access to Grok 4.5 on launch day.
What we cannot confirm yet:
- Exact Opus equivalence (4.7 vs 4.8 vs Fable-adjacent)
- Public API pricing and rate limits
- Whether the reported launch delay improved efficiency measurably or was schedule slippage
- Independent benchmark reproduction
The $60B Cursor acquisition is disclosed but not yet closed (expected Q3 2026). Grok 4.5's Cursor training data predates full merger integration — the legal and product relationship will evolve after close.
The Bottom Line
Grok 4.5 going public on July 9, 2026 is the graduation of a serious beta — 1.5T parameters, Cursor coding data, Colossus compute, production testing at SpaceX and Tesla. Musk's pitch is deliberately economic: Opus-tier capability at better speed and lower cost.
The same day, OpenAI ships GPT-5.6 Sol publicly. Anthropic's Fable 5 is back for eligible users but entangled in export-control and legal noise. Developers are not choosing one frontier model anymore — they are building routing stacks across providers.
Verify before you switch. "Opus-class" is not a benchmark score. Tomorrow's releases will give you the numbers — or the same vague superlatives with extra polish.
Related on explainx.ai
- Grok 4.6 and 4.7: Musk's Announced Timeline, Explained — July 25 roadmap claim and the Opus 5 Pareto-frontier framing
- Grok 4.5 VulcanBench 91.3% + Graffiti Conjecture 284 (July 24) — consumer wide release, cost bench, math claim
- Grok Build open source: Apache 2.0 harness, install, local-first
- Grok Build repository-upload report: evidence, limits, and mitigation
- Musk vs Altman scammer feud — July 2026 — space datacenter insults same week as Grok 4.5 push
- Apple sues OpenAI over AI hardware trade secrets — legal backdrop to Musk–OpenAI rivalry
- Grok 4.5 vs Opus 4.7 and 4.8 — full comparison (July 9) — SWE-Bench, GDPval+, token efficiency
- Grok 4.5 Private Beta at SpaceX and Tesla (June 28) — the beta announcement this launch follows
- SpaceX Acquires Cursor for $60 Billion — SEC filing and merger context
- GPT-5.6 Sol, Terra, Luna Preview — OpenAI's July 9 launch
- Claude Opus 4.8 Launch — the Opus benchmark Grok claims to match
- GPT-5.6 vs Claude Fable 5 Comparison — frontier model positioning
- Cursor Big Day — iOS Launch (June 29) — Cursor product timeline
- Cursor 2× usage limits clarified (July 21) — first-party pool vs Grok promo end
- Anthropic × SpaceX Colossus-1 Partnership — shared compute context
- AI Benchmarks Complete Guide 2026 — how to evaluate launch claims
Sources: Elon Musk on X (July 8, 2026); June 28 Grok 4.5 beta announcement; The Information reporting via Walter Bloomberg and Andrew Curran on X.
Launch timing, performance claims, and pricing reflect public announcements as of July 8, 2026. Verify against SpaceXAI and xAI release notes on July 9.