Opus 5

Opus 5

2026-08-10 00:27:14

Opus 5 built a boat racing game with 690 million tokens, and a GPT-5.6 remake followed at about $5

A viral AI-generated game demo is drawing attention after two separate builds showed how different model stacks and workflows can produce similar results at very different apparent costs. The first project, called INK TIDE, was shared by developer Vyom, who said it was generated with Opus 5 from a single prompt in a one-shot run. According to the article, that process consumed 690 million tokens and cost $423. The game features a cartoon-style speedboat racing setup with lap counting, rankings, a minimap, racing UI, and code-generated models, textures, and audio built with Vite, TypeScript, and Three.js. The prompt behind the project was reportedly around 2,000 words long and structured more like a full product requirements document than a simple instruction. It also included a multi-agent workflow, assigning separate sub-agents to areas such as water simulation, toon rendering, boat physics, AI opponents, animation, audiovisual systems, and performance, with a visual review loop based on screenshots. Soon after, Anul Agarwal shared a similar remake built in Codex using GPT-5.6 Sol Ultra as the main agent and GPT-5.6 Luna Max for execution. That version took roughly five hours and two prompts, with an estimated cost of about $5 based on subscription allocation rather than a standalone API bill. The article says the remake reproduced the broad structure of the original, but its visual quality and polish were clearly lower.

1280

Claude Code

2026-07-30 10:10:09

Claude Code creator Boris Cherny says AI product teams should delete their scaffolding every six months

Boris Cherny, described in the source article as the creator of Claude Code, used a recent YC interview to make a blunt case for aggressive deletion in AI product development: remove the Claude.md file, remove skills, remove hooks, and strip back system prompts and harness code before adding pieces back one line at a time. In the interview released by YC on July 28 and titled “Boris Cherny: We Cut 80% of Claude Code’s Prompt,” Cherny framed the process as an ablation study, arguing that teams should stop guessing what a model needs and instead test each instruction against actual behavior. The article also points to Anthropic’s July 24 update on Claude 5 context engineering, which said Claude Code’s system prompt for newer models including Opus 5 and Fable 5 was cut by more than 80%. Beyond prompt reduction, Cherny laid out two ideas he sees as central to building on fast-improving models: “Product Overhang,” where model capability runs ahead of product design, and “Unhobbling,” or removing constraints so models can attempt harder tasks for longer periods. He cited internal experiments with Opus 5 and OpenCV, argued that self-verification is still a weak point across the industry, and described a long-running Claude task to rewrite an Electron app in Swift while checking output through screenshots and pixel-level comparison. He also urged developers to treat model work as an empirical discipline, not a purely theoretical one.

670

Opus 5

2026-07-29 08:43:11

Opus 5 ‘challenge loop’ prompt goes viral as developers use it to build browser games in 24 hours

A game-building prompt built around Opus 5 has spread quickly across X, drawing attention for how it pushes the model through repeated planning, execution and review cycles. The setup, named the “challenge loop” by Matt Shumer, uses a lead agent to break down the task, multiple specialist agents to produce assets and code, and a separate judging agent to compare results against real AAA references and send weak outputs back for another round. One of the most widely shared examples came from developer Anshu, who said he used the method to create a browser-based space exploration game called “The Long Silence” over a 24-hour run. Shumer also shared an Opus 5 version of a Call of Duty-style project, describing the textures, animation and sound as AI-generated and saying no external assets were used. As the prompt circulated, other users posted kart racers, cyberpunk-style prototypes and experiments on GPT-5.6 Sol. The wave of examples has been framed as a practical demonstration of the long-horizon planning and iterative refinement abilities highlighted when Opus 5 was introduced.

2790

Anthropic

2026-07-24 17:02:03

Anthropic Unveils Opus 5 AI Model at Half the Price of Fable 5

Anthropic announced the launch of its Opus 5 AI model on July 25, according to BlockBeats. The brief said Opus 5 delivers performance close to Fable 5, which it described as operating at a frontier-intelligence level, while costing only half as much. No additional details were provided in the source about the model’s specifications, release format, target users, or benchmark data. The report was published by BlockBeats as a market analysis item, and the source text did not include further technical or commercial context beyond the pricing and performance comparison.

1360

Kimi K3

2026-07-17 11:20:18

Analyst says Kimi K3 could speed up Opus 5 and GPT-6 launches

AI analyst Chubby said the newly released Kimi K3 has outperformed Anthropic’s Opus 4.8 in benchmark tests, a result he believes could push leading U.S. AI labs to accelerate the rollout of their next flagship models. According to his view, Anthropic may move quickly to release Opus 5 in response to competitive pressure from the East. He also said that while OpenAI’s GPT-5.6 Sol still leads Kimi K3 overall, the gap is now so narrow that it is "almost unacceptable." Based on that assessment, Chubby expects the planned release of GPT-6 to be brought forward as well. The comments were reported by BlockBeats on July 17.

1920