Grok 4.5 launched on July 8, 2026, and the clean read is this: SpaceXAI now has a near-frontier coding and agentic-work model with unusually aggressive economics, but the benchmark story is mixed rather than triumphant. SpaceXAI calls Grok 4.5 its smartest model for coding, agentic tasks, and knowledge work, and lists it at $2 per 1 million input tokens and $6 per 1 million output tokens. Artificial Analysis independently places it at 54 on its Intelligence Index, ranked #4 overall. That is good enough to put SpaceXAI in the frontier conversation. It is not enough to turn this into a simple “Opus crushed” story.
The most useful way to understand Grok 4.5 is to separate three things that are being blurred together in same-day coverage: SpaceXAI’s official claims, independent evaluation, and Elon Musk’s own comparison language. The official launch page says Grok 4.5 was trained alongside Cursor, ships in Grok Build, Cursor, the xAI API, Microsoft Office add-ins, and several model gateways, and is not available in the European Union at launch. Independent data from Artificial Analysis says the model is efficient and competitive, but still sits behind Fable 5, GPT-5.5, and Opus 4.8 on its headline Intelligence Index.
That distinction matters because Grok 4.5 is arriving into a noisy release window. Axios reported that the launch is SpaceXAI’s first major model release since SpaceX went public and acquired Cursor, while also noting the tension that SpaceXAI trained this model on compute capacity it also leases to Anthropic and Google. TechCrunch reported Musk’s comparison to Anthropic’s Opus line. The article you want to publish should resist the bait. The news is not that Grok 4.5 is the new universal leader. The news is that it may be one of the strongest intelligence-per-dollar releases of the year.
The Benchmark Picture Is Mixed
SpaceXAI’s launch post publishes four coding and agentic-work benchmark comparisons. The first important caveat is that the DeepSWE results use two different harness stories. DeepSWE 1.0 was run inside each model provider’s own harness, which makes it less neutral. DeepSWE 1.1 uses a mini-swe-agent harness run by DataCurve, which is a cleaner comparison point. On the provider-harness result, Grok 4.5 looks stronger against Opus 4.8. On the neutral harness, Grok 4.5 falls behind Opus 4.8.
| Benchmark | Fable 5 (max) | GPT-5.5 (xhigh) | Grok 4.5 | Opus 4.8 (max) | GLM 5.2 |
|---|---|---|---|---|---|
| DeepSWE 1.0, provider harness | 66.1% | 64.31% | 62.0% | 55.75% | Not listed |
| DeepSWE 1.1, neutral mini-swe-agent harness | 70% | 67% | 53% | 59% | 44% |
| Terminal Bench 2.1 | 84.3% | 83.4% | 83.3% | 78.9% | Not listed |
| SWE Bench Pro resolve rate | 80.4% | 58.6% | 64.7% | 69.2% | 62.1% |
The honest read: Grok 4.5 leads Opus 4.8 on the provider-harness DeepSWE 1.0 score and on Terminal Bench 2.1, but trails Opus 4.8 on the neutral DeepSWE 1.1 run and on SWE Bench Pro. Fable 5 is the stronger model across all four official benchmark rows SpaceXAI chose to publish. If you only quote the best-looking Grok row, you miss the most reader-useful lesson: harness choice changes the story.
Artificial Analysis Puts Grok 4.5 Near The Frontier
Artificial Analysis gives the independent snapshot. Its July 8 analysis says Grok 4.5 scores 54 on the Artificial Analysis Intelligence Index, a 16-point jump over Grok 4.3 and good for #4 overall. The models ahead of it are Fable 5 at 60, Opus 4.8 at 56, and GPT-5.5 at 55. That is an elite result, but it is not the top result.
Where Grok 4.5 looks more distinctive is in agentic efficiency. Artificial Analysis says Grok 4.5 in Grok Build scores 76 on its Coding Agent Index, roughly on par with GPT-5.5 in Codex and below Fable 5 in Claude Code. It also reports a $0.31 cost per Intelligence Index task and says Grok 4.5 uses around 14,000 output tokens per Intelligence Index task, more than 60% fewer than Opus 4.8. That is the sharper competitive claim: not raw supremacy, but useful frontier-adjacent output with substantially less token burn.
On agentic knowledge work, Artificial Analysis places Grok 4.5 at #4 on GDPval-AA v2 with a 1543 Elo, between Opus 4.8 at 1600 and GLM-5.2 at 1513. On τ³-Banking, it reports Grok 4.5 at 33%, above GPT-5.5 xhigh at 31%. Those results support the idea that SpaceXAI has a credible agentic model, while still leaving room for real-world reliability testing.
Specs And Pricing: The Cost Gotcha Is Above 200K Tokens
The official docs are refreshingly specific on the API surface. The model string is grok-4.5, with aliases grok-4.5-latest and grok-build-latest. The context window is 500,000 tokens, smaller than the 1 million-token context listed for Grok 4.3. SpaceXAI also warns that requests above the 200,000-token context window are billed at different higher-context rates. For agent loops, that threshold matters more than the headline context size.
| Model string | grok-4.5 |
| Aliases | grok-4.5-latest, grok-build-latest |
| Context window | 500,000 tokens |
| Higher-context threshold | Above 200,000 tokens |
| Modalities | Text and image input; text output |
| Reasoning | Low, medium, or high; default high; cannot be disabled |
| Input price | $2.00 per 1 million tokens |
| Cached input price | $0.50 per 1 million tokens |
| Output price | $6.00 per 1 million tokens |
| Tools | Function calling, web search, X search, code execution |
| Rate limits | 150 requests per second; 50,000,000 tokens per minute |
| Regions | us-east-1 and us-west-2 |
| Knowledge cutoff | Not published for Grok 4.5 |
That pricing is the core commercial move. At $2.00 in, $0.50 cached in, and $6.00 out per million tokens, Grok 4.5 is priced for high-volume agent workloads. But the prompt-caching note in SpaceXAI’s own docs is easy to miss: if teams do not route conversations consistently, cache hits become less reliable and the real bill can drift upward.
Speed And Token Efficiency Are The Real Story
SpaceXAI says Grok 4.5 is served at 80 tokens per second. Artificial Analysis measured the API at 91.3 output tokens per second. The numbers are not identical because they come from different measurements, but they point in the same direction: Grok 4.5 is not trying to compete only by being cheaper. It is trying to be fast enough that lower token use becomes an experience advantage, not just a finance-team advantage.
SpaceXAI’s own token-efficiency example is SWE Bench Pro: it says Grok 4.5 uses 15,954 average output tokens per task versus 67,020 for Opus 4.8 max, or 4.2x fewer. Vendor token-efficiency claims deserve scrutiny, but Artificial Analysis independently supports the same direction with its roughly 14,000 output-token-per-task Intelligence Index figure. For builders, that may be the most actionable launch-day signal: Grok 4.5 could make long agent sessions cheaper, faster, and less verbose without dropping out of frontier range.
Training, Cursor, And The 1.5T Claim
Officially, SpaceXAI says Grok 4.5 was trained across tens of thousands of NVIDIA GB300 GPUs, with heavy data filtering, deduplication, quality scoring, and domain-focused selection. The official post also says reinforcement learning covered hundreds of thousands of tasks, centered on multi-step software engineering and technical work. That is what can be stated as official.
The 1.5 trillion-parameter and V9 language belongs in a different bucket. Musk has said Grok 4.5 is based on that V9 foundation-model framing, but the official launch post and official model docs do not publish a parameter count. Artificial Analysis’ model FAQ is explicit that SpaceXAI has not disclosed the model size or parameter count. So the clean wording is: Musk has self-reported the 1.5T/V9 framing; SpaceXAI has not confirmed it in the official launch materials.
The Cursor angle is real, but it also needs careful phrasing. SpaceXAI says Grok 4.5 was trained alongside Cursor, and the docs say it is available in Cursor on all plans. Axios frames the model as SpaceXAI’s first major release since the Cursor acquisition. That does not prove every claimed coding improvement came from Cursor data, nor does it settle how much of the training was pretraining versus supplemental work. It does tell us where SpaceXAI wants the model judged: not in chat demos, but in coding agents and office-style task execution.
Availability: Broad, But Not In The EU Yet
Grok 4.5 is available today in Grok Build, Cursor, and the SpaceXAI console. The docs also list it as the default model in Word, PowerPoint, and Excel add-ins, with gateway availability through OpenRouter, Vercel, Cloudflare, Snowflake, and Databricks Mosaic. SpaceXAI is offering free Grok 4.5 usage for a limited time in Grok Build and Cursor.
The geographic caveat is clear: Grok 4.5 is not available in the EU in SpaceXAI products or the API console at launch, with EU availability expected in mid-July. That matters for enterprise buyers because “available today” does not mean globally available today.
The Competitive Context
Musk’s framing is deliberately Opus-adjacent. The two useful quotes are: “It is an Opus-class model, but faster, more token-efficient and lower cost.” He also said: “Our internal assessment is that Grok 4.5 is roughly comparable to Opus 4.7, but much faster. The combination of capability, faster speed and lower cost is what makes it competitive.” TechCrunch cited both Musk posts, and the second quote is especially important because Musk benchmarked the internal comparison to Opus 4.7, not Opus 4.8.
That should shape the comparison. Grok 4.5’s best case is not that it magically erases Anthropic and OpenAI. It is that a model ranked #4 by Artificial Analysis, priced at $2/$6 per 1 million tokens, and producing far fewer output tokens on several agentic tasks may be the better default for workloads where speed and cost compound. If your workflow needs the absolute strongest neutral-harness coding result, the official benchmark table still points you back toward Fable 5 or Opus 4.8. If your workflow runs many tool-heavy coding or knowledge-work loops, Grok 4.5 deserves testing.
For Kingy readers comparing models, the practical next step is not “switch everything today.” It is workload-specific A/B testing against your current agent stack. Start with codebase repair, long-context research, spreadsheet-heavy office work, and multi-step tool use. Then measure completion rate, total tokens, latency, retry rate, and human review time. The launch-day evidence says Grok 4.5 may save a lot of money when it succeeds. It does not yet prove it succeeds often enough in every production environment.
Kingy Verdict
Grok 4.5 is not the new benchmark monarch. It is something more commercially interesting: a credible frontier-adjacent model whose strongest claim is intelligence per dollar and token efficiency. The official benchmarks are split, and the neutral-harness rows prevent an overconfident Opus narrative. But the Artificial Analysis data, the $2/$6 pricing, the 500K context window, and the token-efficiency profile make Grok 4.5 one of the most important agentic-work launches of 2026.
The unanswered question is reliability. Benchmarks can tell us whether a model is capable. They do not tell us whether it keeps its footing across messy repos, weird enterprise documents, inconsistent tool outputs, and long-running agent loops. If Grok 4.5’s efficiency holds up under those conditions, SpaceXAI has a serious wedge. If it saves tokens by cutting corners, the savings will vanish into review time. That is the real test.
FAQ
When was Grok 4.5 released?
Grok 4.5 was publicly launched on July 8, 2026. If you see a July 9 date, treat it as stale or extrapolated from pre-launch posts; the official SpaceXAI launch page is dated July 8.
How much does Grok 4.5 cost?
The official API price is $2.00 per 1 million input tokens, $0.50 per 1 million cached input tokens, and $6.00 per 1 million output tokens. Requests above 200,000 tokens can trigger higher-context pricing.
What is the Grok 4.5 context window?
Grok 4.5 has a 500,000-token context window. That is smaller than Grok 4.3’s 1 million-token context listing, so do not assume context grew with the new model.
Is Grok 4.5 available in the EU?
No. SpaceXAI says Grok 4.5 is not available in the EU in its products or API console at launch, with EU availability expected in mid-July.
How many parameters does Grok 4.5 have?
SpaceXAI has not published an official parameter count. Musk has self-reported a 1.5 trillion-parameter V9 foundation-model framing, but Artificial Analysis’ model page states SpaceXAI has not disclosed the model size or parameter count. Treat 1.5T as Musk-attributed, not independently confirmed.
Where can you access Grok 4.5?
You can access Grok 4.5 through Grok Build, Cursor, the xAI API console, Word, PowerPoint, Excel add-ins, OpenRouter, Vercel, Cloudflare, Snowflake, and Databricks Mosaic, subject to region availability.