The Week's Big Shake-Up: GPT-6 Soul and Luna at Half Price as Opus 5.5 Claims the Crown

The last 48 hours turned the AI news cycle into a simultaneous fireworks show, and WorldofAI's roundup tries to catch every burst in one place — the host himself notes this is his third video today, which says a lot about the tempo. After OpenAI's flagship GPT-6 Astra at the start of September, two lighter siblings take the stage, Anthropic pushes the frontier with the new Opus 5.5 family, Alibaba and Tencent fire back-to-back announcements from China, and American politics even tries to rename the field. Packing such different tracks into one window makes the video a panorama of the week rather than a single launch story.

GPT-6 Soul and Luna — Same Generation, Half the Bill

OpenAI's new duo, GPT-6 Soul and GPT-6 Luna, brings Astra-grade GPT-6 intelligence to everyday workloads. The promised gains cover coding, factual accuracy, computer use and professional tasks, but pricing steals the headline. Soul is listed at $2 per million input tokens and $10 per million output, while Luna lands at $0.10 input and $0.50 output. Both are roughly 50% cheaper than the prior GPT-5.6 promotional rates, and in the host's telling that translates into one line: same-generation brains at a far lower bill, with the barrier for high-volume use coming down visibly.

Price alone would not be news without proof in professional workflows, and the video leans on Automation Bench for that. The benchmark runs realistic business workflows across multiple apps. There Soul manages to beat Anthropic's prior flagship Opus 5 at maximum effort while cutting cost per task to about 9%. Luna follows a similar arc, adding 5.4 percentage points over its predecessor at high effort while costing 58% less per task. The takeaway the host repeats is straightforward: OpenAI is not just cheaper, it delivers more work per dollar, and that ratio is being placed at the center of buying decisions.

Curiosity then turns to real-world cost and what independent evaluators see. Via AIML API the video generates four distinct three-dimensional scenes in one shot each — a moving fish shoal and a tsunami-like set among them. Opus 5.5 spends about $4.37 for the four, Soul does the set for $0.34, and while Opus leads visually the cost gap approaches 13x. On the global intelligence index Soul ranks third overall behind Astra, while Artificial Analysis shows only a one-point move from GPT-5.6 Soul to the new Soul, prompting the host to say pricing thrills but the jump is smaller than he hoped. In the same block hallucination rates are said to fall from roughly 92% to 60% for Soul and 93% to 77% for Luna, a drop that stands out at Luna's scale.

Opus 5.5 — Is There a New King

Anthropic drives the video's boldest claim. Claude Opus 5.5 arrives as the first member of the Claude 5.5 family, and in the host's framing it matches or beats Fable 5.1 across most tasks while running about 40% cheaper than Opus 5. Ranked at the top in independent tallies at the time of recording, it also ships with a one-time usage-limit reset valid until October 22. In a side-by-side New York skyline built with Three.js, Soul delivers a static single frame without a proper render loop while Opus 5.5 erects the full scene with textures, shadows, lighting and interactive camera movement; this is where the host's line that the gap spans 'many domains, not two' becomes visible. Yet when the prompt is more detailed, Soul can also assemble a browser-based OS-like interface with calendar, lock screen and SVG app icons fairly competently, showing how much prompting detail matters.

Opus 5.5's show of force continues with a night-train simulation. A train cutting through a northern landscape, described as Alaska-like, is coded end to end in Three.js with tracking cameras and atmospheric lighting, and the host reads it as a clear visual step ahead of GPT Astra. The point is not only aesthetics; the whole scene runs as an interactive simulation without the plan falling apart over a long horizon, which the host links to stronger planning and coherence.

From Minecraft to Elden Ring — A One-Prompt Game Factory

As the demo marathon lengthens, scale grows. A Minecraft-like sandbox world is rebuilt with shaders, and Opus 5.5 at maximum effort runs for 1 hour 37 minutes to produce richer lighting and textures than the host expects. Next a playable Spider-Man Miles Morales-inspired city arrives at medium effort, complete with swing physics, animations and textures in one prompt. The pivot is an Elden Ring-style action game: produced at peak inference effort, this clone does more than mimic textures and mood, it builds a playable loop up to a boss fight. Here the host crystallizes the thesis: the model is no longer just a landing-page builder but a full-game generator, and the separation shows up in conceptual depth rather than surface scaffolding.

Opus 5.5 is also a family signal, not a solo star. Anthropic confirms on its Opus 5.5 blog post that Claude Haiku is back and that Sonnet 5.5 and Haiku 5.5 should land within weeks, with the expectation that the efficiency and safety gains seen in Opus 5.5 carry down to lighter siblings. A parallel side demo asks for a macOS-like system inside the browser: calendar, lock screen, SVG app icons, working links and even a music player appear, although dragging folders does not yet work and the host notes further prompt refinement is needed.

A Double Salvo From China — Qwen 4 and HY Image 3.5

Alibaba clarifies the Qwen 4 roadmap around its Apsara conference. The lineup lists Qwen 4 Max at the top, followed by Qwen 4 Plus and Qwen 4 Flash, with the most anticipated being a locally runnable Qwen 4 27B that the host describes as near-frontier when run on device. The roadmap does not stop there; Qwen 4.5 and Qwen 5 are floated in the 5 to 10 trillion parameter range, which would mark a major scale leap if delivered. Tencent answers on the image front with HY Image 3.5 Preview: a 30% human-preference win over HY Image 3.0, consistency improvements, text-to-image and image-to-image support and generation up to 2K. The sharpest data point is price: 2 cents per image on Tencent Cloud with no fee for reference images, signalling a push where quality and cost are optimized together.

Trump's Rename Push — From Artificial to Super

The video's most political segment reaches the White House. Former President Donald Trump argues the word 'artificial' makes the technology sound fake and asks for a rename. He first puts three options to a vote: Superior Intelligence, Extreme Intelligence and Supreme Intelligence; in the last poll on X, Superior Intelligence wins narrowly with 52% of 184,000 votes. In a September 22 address to the United Nations General Assembly, he says U.S. government documents will adopt 'super intelligence' or SI as the more accurate term, phrasing it as 'from this point forward, all U.S. documents and hopefully the world's.' Experts quoted by the BBC counter that SI is premature and misleading for today's systems and that the industry prefers to reserve it for a higher tier; some figures in Washington already use SI, but broad adoption looks unlikely and the host jokes he may need to rename his channel World of Super Intelligence.

Stepping back, the video signals a shift in the ecosystem rather than a single-model victory. Price lists have become the first row of the spec sheet, efficiency is marketed almost as loudly as capability, locally runnable 27B models and 2-cent images point to broader access, while Opus 5.5's game-factory demos prove the frontier still separates. When production cost falls nearly 13-fold yet visual quality stays 'not far behind,' the rational move for many teams is not to pick the best at any cost but to pick the best balance of cost and capability; the host's note that Soul felt like Opus 5.5 in his own prompt library carries weight for that reason. The open questions are whether discounts prove durable or promotional and whether independent tests confirm the same picture; until then the prudent lens is not sticker price but cost per completed task.