Is Grok 4.5 Really Opus-Class?

A reviewer paid roughly eight dollars to render a detailed 3D city of fifty blocks using Grok 4.5 under a specification written by a bigger model. That demo, from Wes Roth's July 2026 video, is the clearest picture of what xAI's new release is sold for: cheap, fast, instruction-following construction work.

What Is Grok 4.5?

Grok 4.5 is a large language model that xAI, Elon Musk's AI company, released in July 2026, and the video's central claim is that it finally reaches Opus-class coding ability at lower cost. That framing belongs to reviewer Wes Roth, who tested the model within two hours of launch. Nothing in this article independently confirms his results; it reports them as first-hand reviewer experience.

According to the video, Grok 4.5 is the first model trained after xAI's reported acquisition of Cursor, the AI-first code editor built on VS Code. Roth argues the combination matters because Cursor sees real end-to-end developer sessions, which he describes as valuable training signal. Whether that acquisition and data advantage produced the benchmark gains is his interpretation, not documented fact.

The model is also, per Roth, the default model in Grok Build, a terminal-based coding tool he positions as xAI's answer to Claude Code, Anthropic agentic terminal coding assistant. He ran most of his demos through it in auto mode.

Which Benchmark Numbers Does The Video Report?

The benchmarks Roth cites are figures he read on screen during the video. He names Deep SWE 1.0 as a benchmark he considers realistic and applicable, and places Grok 4.5 between OpenAI's GPT-5.5 extra high and Anthropic Opus 4.8 Max on it. He describes Deep SWE 1.1, terminal bench, and Sweet Bench Pro results as close to GPT-5.5. None of these scores can be verified here against a primary leaderboard, so treat the table below as reported, not confirmed.

Roth also repeats an unverified community estimate that Grok 4.5 has about 1.5 trillion parameters. He offers no source, and parameter counts for closed models are rarely published, so this should be read as rumor he passed along.

Benchmark (as named in the video)	Reported Grok 4.5 placement	Reported comparison
Deep SWE 1.0	Between GPT-5.5 extra high and Opus 4.8 Max	Slightly below both leaders
Deep SWE 1.1 / terminal bench	Neck and neck	On par with GPT-5.5, above Opus
Sweet Bench Pro	Very close	Right at GPT-5.5 level

How Did The 3D Sailing Game Test Go?

The sailing game is Roth's signature test, previously used on the top model he calls Fable 5. He asked Grok 4.5 via Grok Build to build a 3D first-person and third-person sailing game with wind physics, sail trimming, wave Notion, and audio generated through the ElevenLabs API, the voice and sound platform he supplied an API key for. He judged the result a win for Grok 4.5 over Fable 5 on this specific task.

He highlights three successes. The ship rolls with the waves without locking the camera orientation, a failure he saw in earlier models. Sail mechanics work: you cannot sail directly into the wind, and sails visibly fill when trimmed correctly. Weather tiers from calm through storm to hurricane work, including capsizing if sails stay out, and a health-and-warning system Roth requested in follow-up prompts was implemented as asked.

He also lists defects: minor object clipping, a weak lightning sound he blames on ElevenLabs rather than the model, and an underwater sinking sequence that needs polish. His verdict is that visual generation, historically the weakest skill for coding models, held up better than expected.

What About The RPG And 3D Scene Tests?

The second major test was an Elder Scrolls-style village RPG with voiced dialogue, again using ElevenLabs. Roth reports strong branching storytelling: blackmailing a merchant changed what information he received on a replay, and a peaceful resolution of the crypt boss was reachable through clues. He found one mouse-focus bug and says Grok fixed it after being told. Voice selection was random, because the model cannot hear the audio it wires up.

Quick 3D and SVG tests showed a consistent pattern. An ancient Rome scene in HTML rendered well except for upside-down roofs, fixed with one correction prompt. A Mars base looked good; an old-town RPG scene had broken lighting and needed regeneration. Handcoded SVGs, including an animated cyberpunk alley, came out well. He notes that Anthropic Opus 4.6 and OpenAI's GPT-5.4 remain strong at SVG work in his experience.

The Architect And Construction Crew Workflow

The most useful demo in the video is a two-model workflow Roth calls the mountain of evidence for model routing. He asked Fable 5, the large expensive top-tier model, only to write a specification: precise directions per city district plus acceptance checks that reject outputs failing spec. Grok 4.5 then executed the build district by district, never seeing the whole city.

The result, per Roth: roughly fifty unique city blocks with a botanical dome, docks with animated cranes, a stadium, and a ferris wheel, built from 1.35 million tokens of JavaScript for about eight dollars through the API. He estimates the same build on Fable 5 would have cost seventy to eighty dollars. Those are his figures from his own run, not a published study, but the workflow pattern is reproducible by anyone with API access to both models.

1. Ask the strong model for a specification with explicit acceptance criteria.
2. Hand the spec to the cheaper model and have it build in small, independently checkable units.
3. Return any unit that fails the checks for revision instead of regenerating the whole project.

This pattern explains his economic argument. Most engineering work does not need the most capable model; it needs a capable model that follows a good specification cheaply and fast.

Why Speed And Price Are The Real Story

Roth's core argument is about position on the intelligence-per-dollar curve, not raw capability. He reports that Grok 4.5 is fast and markedly cheaper than the models it matches on coding quality. If that holds in ordinary use, work currently delegated to mid-tier models becomes cheaper, and heavy users of metered APIs see the difference first.

He frames a second consequence for Cursor, the code editor now reportedly owned by xAI. Before this, Cursor paid API prices for frontier models that those labs subsidized for their own first-party tools. Owning a competitive model removes that dependency. This is Roth's strategic analysis, and the acquisition itself is reported here only as he reports it; no transaction document was reviewed for this article.

The practical takeaway for readers: route expensive models to planning and review, route Grok 4.5-class models to construction, and measure the savings on your own workload rather than trusting any launch-day demo.

What Remains Unverified

Every capability claim in this article traces to one reviewer working within hours of release, plus on-screen numbers whose original sources the video does not fully show. Roth himself calls his Fable 5 comparison unfair and his parameter estimate unconfirmed. xAI's own announcements and published evaluations are the right place to confirm benchmark scores, pricing, and the Cursor acquisition before making procurement decisions.

The demos do establish one thing a skeptic should weigh: a model that produces a playable sailing game, a branching RPG, and a million-token 3D city from prompts, within hours of launch, is genuinely capable even if every exact number is wrong. Watch for independent replication over the following weeks rather than treating day-one enthusiasm as settled.

FAQ

• What is Grok 4.5? It is xAI's model released in July 2026, described by reviewer Wes Roth as an Opus-class coding model that is fast and inexpensive. All performance figures in circulation from his video are reviewer-reported, not independently verified.

• Is Grok 4.5 better than Fable 5? No, not on overall capability. Roth found a sizable gap to Fable 5, his name for the current top-tier model, though Grok 4.5 won on his specific 3D sailing game test. He argues the comparison matters little because most work does not need the top tier.

• How much does Grok 4.5 cost? Roth reports it is substantially cheaper than the models it matches, and his demo city build cost about eight dollars in API spend for 1.35 million tokens. Check xAI's current pricing page for exact numbers before budgeting.

• What is the architect-and-construction-crew pattern? It is a two-model workflow where a strong expensive model writes a specification with acceptance checks and a cheaper model executes the build. Roth's version used Fable 5 for the spec and Grok 4.5 for construction, cutting his estimated cost by roughly ninety percent.

• Can I trust the launch-day benchmarks? Treat them as provisional. The video cites Deep SWE and terminal bench placements without showing reproducible runs, and the parameter estimate is unconfirmed. Wait for the model to appear on maintained public leaderboards.

Turn Your Own Watch Sessions Into Articles

This article exists because one 35-minute video contained a benchmark reading, a reproducible workflow, and a cost argument worth preserving in searchable form. If you sit on the same kind of knowledge inside YouTube videos, whether interviews, tutorials, or your own commentary, that material does not need to stay locked in a timeline.

Skalablog converts a YouTube video into a structured written article: paste the URL, get a transcription, and edit a draft that keeps the evidence and drops the filler. The routing logic applies to content too, since your existing recordings already hold the expensive thinking. Visit skalablog.com to run your first video through it.

If you build software and want structured practice material alongside reading, CrazyStack Typescript collects project-based lessons in one place. Its author, Gustavo Dev Doido, publishes walkthroughs that pair well with the build-and-review habits this article describes.

Source video