A 1.5 trillion parameter model at $2 per million input tokens is the headline around Grok 4.5 in a July 2026 review of the release. The video reports the model sits slightly below GPT-5.6 and Fable 5 on benchmarks but undercuts United States frontier labs on price. Here is what the video actually shows, and where its evidence stops.
What Grok 4.5 Is, According to the July 2026 Review
Grok 4.5 is a frontier large language model whose July 2026 release was reviewed by the YouTube channel Caleb Writes Code in an eight-minute video. xAI is the company behind the Grok model family, and the review describes Grok 4.5 as the successor to the Grok 4 variants, built on a newly pre-trained base the video calls the V9 foundation model.
The single most important caveat for readers: this article reconstructs one video's claims. As of the October 10, 2026 publication date, the figures below are attributed to the reviewer or to vendor reporting cited in the video, not independently verified by us. Where a number comes from the model maker rather than the reviewer, the text says so.
The video's thesis is that Grok 4.5 occupies a pricing and intelligence wedge. It reportedly performs above open-weight models, slightly below the top closed models from OpenAI and Anthropic, and costs about as much as those labs' cheaper tier models. The rest of this article unpacks each part of that claim.
The V9 Base Model and the Reported Cursor Connection
The review's central technical claim is that Grok 4.5 is not an incremental tune-up. The video says it runs on a newly pre-trained V9 foundation model of 1.5 trillion parameters, roughly three times larger than the previous V8 base. A freshly pre-trained base matters because it is trained from scratch on a new data mix, which is where new capabilities usually come from.
The video then makes a claim we could not verify anywhere: that the model's maker acquired Anysphere, the parent company of Cursor, the AI-first code editor, in June, and that high-quality coding data from Cursor supplemented the V9 training mix. Treat this as a speaker assertion from the video. An acquisition of this size would normally be confirmed by both companies' official channels, and this article found no such confirmation at publication time.
If the reported Cursor data contribution is accurate, it would explain the model's positioning toward coding and knowledge work. If it is not, the capability story still rests on the larger base model and the vendor's own benchmark reporting, which is a weaker but different kind of evidence.
Where Grok 4.5 Sits on the Artificial Analysis Index
The video's benchmark reference is the Artificial Analysis model index, a third-party site that aggregates performance measurements across model providers. On that index, the reviewer reports Grok 4.5 ahead of open-weight models and just behind the leading closed models, named in the video as Fable 5 and GPT-5.6.
The reviewer also stresses how compressed the release window was. Per the video, Fable 5 arrived on June 9, 2026, GPT-5.6 was announced on June 26, 2026, and Grok 4.5 followed within the same thirty-day span, with Opus 4.8 the only recent model outside that window.
| Model (as named in the video) | Type | Reported standing | Reported release timing |
|---|---|---|---|
| Fable 5 | Closed | Slightly above Grok 4.5 | June 9, 2026 |
| GPT-5.6 / GPT-5.6 Sol | Closed | Slightly above Grok 4.5 | June 26, 2026 |
| Grok 4.5 | Closed | Top of open-model field | July 2026 |
| Chimera 2.6 | Open, 1T parameters | Below Grok 4.5 | Within 30-day window |
| DeepSeek V4 Pro | Open, 1.6T parameters | Below Grok 4.5 | Within 30-day window |
For historical context the video reaches back to September 2024, when OpenAI released o1, the first widely noted reasoning model, and says the industry needed about six months to respond, starting with the DeepSeek moment in January 2025. DeepSeek is a Chinese lab whose early-2025 release briefly reset cost expectations across the market. The reviewer frames the mid-2026 cluster of releases as a comparable leap, this time started by Anthropic Mythos and Fable line.
One limit the reviewer himself flags: the index snapshot he showed did not include GPT-5.6 Sol, so its placement is his estimate, not a measured point on the chart.
Pricing: Cheap Against US Frontier Labs, Expensive Against Open Models
The clearest number in the review is price: $2 per million input tokens and $6 per million output tokens for Grok 4.5. The reviewer's argument is that this puts API pricing at the level of lower-tier models from other labs, which he lists as Sonnet 5 from Anthropic, GPT-5.6 Lunar from OpenAI, and Gemini 3.5 Flash from Google's Gemini family.
The wedge cuts the other way against open models. The video names MiniMax M3, Kimi K2.6, DeepSeek V4 Pro, GLM-5.2 and Qwen 3.7 Max as noticeably cheaper options, with only the top Qwen entry approaching Grok 4.5's price. So the model is neither the budget choice nor the premium choice; it is priced to undercut closed rivals while charging more than open alternatives it claims to outperform.
A note on subscription plans: the reviewer observes that most users pay monthly subscriptions rather than raw API credits, and that subsidized plan pricing can obscure true per-token costs. When comparing models, always normalize to published per-million-token rates before drawing conclusions, because plan bundles are not comparable across providers.
| Comparison axis | Grok 4.5 position per the video |
|---|---|
| Versus US frontier labs | Priced like their cheaper tiers, intelligence claimed near their top tiers |
| Versus open-weight models | More expensive, reportedly more capable |
| Input tokens | $2 per million |
| Output tokens | $6 per million |
What the Reviewer Actually Tested
The release notes, as read in the video, signal three intended use cases: software development, general coding, and office knowledge work. The reviewer tested two of the three himself. These are first-hand experiences from one person on one day, so treat them as anecdotes with limited general reach.
Coding test in the vendor's own harness
Using the vendor's own coding tool, which the video calls Grok Built, the reviewer asked for a visualized site of major United States data centers. He reports the model produced a working interactive map with company filtering and zoom within a few minutes, and he calls the speed surprising. The vendor's own demonstration, a one-shot solar system simulator, is also shown in the video but is a curated example.
Office work test in Cursor
For knowledge work, the reviewer loaded Grok 4.5 inside Cursor, selected it from the model dropdown, and asked it to convert the data center site into a Word briefing document and a PowerPoint deck. He reports both files were generated in one to two minutes and matched his formatting instructions. Because Cursor integrates many models, this is also a way to compare Grok against rivals inside one workflow.
The reviewer also notes he ran out of credits while testing a more technical task, so the depth of his evaluation was bounded. The integration with Cursor, combined with what the video claims is large compute availability and a comparatively efficient model, forms the enterprise argument the reviewer sketches at the end.
Token Efficiency Versus Raw Speed
The throughput number in the video is an 85 tokens per second service-level figure for Grok 4.5, which the reviewer contrasts with a 750 tokens per second option announced alongside GPT-5.6. On raw speed, Grok 4.5 loses by that comparison.
The counterargument the video makes is efficiency: citing the vendor's report, the reviewer says Grok 4.5 solved SWE-bench Pro tasks with average output tokens up to 4.2 times fewer than Opus. SWE-bench Pro is a software engineering benchmark, and this figure is vendor-reported, not independently reproduced. If it holds, fewer tokens per task is a real cost saving at the $6 per million output rate, and it changes the total-cost picture even at slower generation speed.
Both numbers deserve the same treatment: raw speed is measured per token stream, while cost depends on tokens per task multiplied by price. A model that uses a quarter of the tokens can be cheaper end to end while looking slower on a speed chart.
Context Window and What to Verify Before You Rely on This Model
Two forward-looking details close the video. First, the reviewer says Grok 4.5 shipped with a 500,000-token context window. Second, he relays a statement attributed to Elon Musk on X that the context window would grow to 1 million tokens within days of release. A roadmap statement is a commitment, not a shipped capability; check the current model documentation before building against the larger window.
Before adopting the model based on this review, verify four things against primary sources: the current published pricing on the provider's own documentation, the current context window, independent entries on the Artificial Analysis index rather than a single video snapshot, and any confirmation of the reported Anysphere acquisition and Cursor data contribution. None of those were independently confirmed for this article as of October 10, 2026.
The honest summary is that the video makes a coherent case, grounded in one third-party index, one vendor efficiency report, and two quick hands-on tests, that Grok 4.5 competes on price-adjusted intelligence. Whether that case survives independent benchmarking is the question a careful buyer asks first.
Frequently Asked Questions About Grok 4.5
- How much does Grok 4.5 cost? Per the July 2026 review, the API price is $2 per million input tokens and $6 per million output tokens. The video frames this as comparable to cheaper-tier models from OpenAI, Anthropic Google, while costing more than most open-weight alternatives.
- Is Grok 4.5 better than GPT-5.6? The video reports Grok 4.5 sits slightly below GPT-5.6 and Fable 5 on the Artificial Analysis index. Its claimed advantage is price and token efficiency, not raw benchmark leadership. Both models are new, so independent comparisons remain thin.
- Can you use Grok 4.5 in Cursor? Yes, according to the review. The model appears in Cursor's model dropdown, and the reviewer used it there to generate a Word briefing and a PowerPoint deck. Availability of specific models in third-party editors can change, so check the editor's current model list.
- How big is the Grok 4.5 context window? The video says the model launched with a 500,000-token context window and quotes a statement that it would expand to 1 million tokens shortly after release. Treat the expansion as a claim to verify against current official documentation, not a confirmed shipped feature.
- Was Cursor's data really used to train Grok 4.5? The video claims the model's maker acquired Anysphere, Cursor's parent, in June and used Cursor coding data in the V9 training mix. This article found no independent confirmation of the acquisition or the data usage, so treat both as unverified claims from the video.
From Watched Video to Written Reference
The Grok 4.5 story in this article is a good example of why a single eight-minute video benefits from being turned into structured text: numbers, dates and caveats are much easier to check, quote and search when they live in an article rather than a playback timeline. If you have release analyses, benchmark walkthroughs or technical explainers sitting in your own YouTube videos, you can turn that spoken knowledge into a written reference with Skalablog: paste the video URL, get a transcription, and generate an article you can edit and publish. For a deeper TypeScript reading list, the resource at CrazyStack Typescript pairs well with model-evaluation workflows like the one described above. Thanks to Gustavo Dev Doido for the pointer that shaped this piece.
Fork this article
Start a new branch from the same video, shaped your way. You keep the credit; the original keeps the attribution.
A fork in another language is filed as a translation of this article, so the two pages point at each other. You can unlink it later from the editor.
0/240
You are creating
- Format
- For
- Language
- Source
- Your angle
You will be asked to sign in before it is generated.
Buy credits