The OpenAI–Anthropic Benchmark Dispute Over ARC-AGI-3

OpenAI claims that its model, GPT-5.6 Sol, has outperformed Anthropic's Opus 5 on ARC-AGI-3, a highly challenging reasoning benchmark. According to The Decoder, GPT-5.6 Sol scored 38.3%, beating Opus 5's 30.2%.

However, there is an important caveat: OpenAI's strong score was achieved only in its **own API environment**, which comes with a mechanism that compresses context while preserving the reasoning process. When the same model is placed in the official test environment, the score drops to 7.8%. Opus 5's 30.2%, by contrast, is said to have been achieved without any such assistance.

In short, this is not an apples-to-apples comparison under identical measurement conditions, so it is hard to say outright that GPT-5.6 Sol is the better model. The episode has once again brought to the surface a long-debated problem in AI evaluation: which environment and settings a benchmark runs under can dramatically sway the results. For now, the sensible approach is to carefully check each company's measurement conditions before drawing any conclusions.

opencode.ai Blocked on Multiple Networks

The developer tool domain opencode.ai is reportedly being blocked across multiple routes. According to a Hacker News discussion, when the domain is resolved via Cloudflare, it is classified as malware and DNS returns 0.0.0.0. There are also reports that SFR, a French internet provider, is blocking access to the domain on the grounds of user protection.

Cases where a legitimate developer tool gets mistakenly flagged as harmful are drawing attention in the developer community as a side effect of automated security classification. The reason for the classification remains unclear at this point, and it bears watching whether it gets corrected going forward.

KADOKAWA and Hatena Announce RIKU, an AI Novel-Writing Editor

In Japan, KADOKAWA and Hatena have announced RIKU, an editor that uses AI to assist with writing novels. According to ITmedia AI+, the service is currently recruiting testers. As an effort to combine AI with creative writing support, it marks a step forward for practical adoption in Japan's content production scene.

Apple Heads Into Earnings Amid Cooling AI Investment

On the market side, Apple is drawing relative popularity as investors turn more cautious toward AI hype. According to Bloomberg, Apple shares rose 17% in July, marking the highest gain among the Magnificent Seven for full-year 2026 at 24%. Its market capitalization briefly topped the $5 trillion mark, reaching parity with NVIDIA. Investors wary of excessive AI spending are reportedly evaluating Apple as an "AI safety stock." Those expectations are set to be tested at Thursday's earnings report.

StatePlay: A World Model That Predicts a Game's Internal State

From the research world, StatePlay, a paper on world models for games, has been featured in Hugging Face's Daily Papers. In addition to conventional pixel-level visual generation, the approach explicitly predicts internal game states — HP, skill meters, timers, and the like — and uses them as constraints for frame generation. According to the paper, the model keeps the normalized error of state prediction below 0.06 and improves mechanics fidelity by 18.6% compared with models that do not handle internal state. It is drawing attention as a direction for improving the behavioral consistency of game AI.

---