Sep 23 · AI Daily

OpenAI has brought higher hit rates, diagnostic tools, breakpoints, warm-ups, and reasoning changes that retain caches to GPT-6 prompt caching. Cache reuse can cut cached input token costs by up to 90%, and a new dashboard shows cache hit rates alongside cached and uncached token volumes. The author notes that high hit rates remain an application design problem, and developers need to keep stable instructions and tools for long-running agents reusable.

Huawei has released version 1.2.0 of the Cangjie programming language under the STS channel with a 6-month maintenance cycle. STS 1.1.x is scheduled to reach end-of-life on Oct 30, 2026. Compared to 1.1.3, this version adds Android 23 (API 23) aarch64 and iOS LTO support, enhances ObjC/Java cross-language interoperability, and fixes a batch of semantic ICE issues.

The team behind the AI/ML API tested the new Claude Opus 5.5 using GPT-6 Sol.

These tests involved generating a set of one-shot 3D scene prompts, starting with "school of animated fish."

According to their results, across four scenes, Opus 5.5 generated finer-grained renders for a total cost of $4.37, compared to $0.34 for GPT-6 Sol.

At the 2026 Snapdragon Summit, Qualcomm introduced two chips: the Snapdragon 8 Elite and the Snapdragon 8 Elite Extreme Edition. Both share largely identical maximum CPU frequencies, X105 5G modem, Wi-Fi 8/7/6, and graphics API specifications.

At the 2026 Snapdragon Summit, Qualcomm announced that Xiaomi, vivo, Red Magic, REDMI, OPPO, OnePlus, Motorola, iQOO, and Honor will be among the first to adopt the Snapdragon 8 Elite Extreme Edition and Snapdragon 8 Elite mobile platforms in their flagship products. The Snapdragon 8 Elite Extreme Edition is officially promoted as the "world's fastest mobile CPU." Qualcomm stated that this dual-flagship strategy aims to bring agentic AI experiences to a broader range of flagship devices.

Artificial Analysis calculates the cost per Intelligence Index task for Claude Opus 5.5 (max) at $5.98, close to the $5.86 of Opus 5 (max). Higher token usage alone would have pushed costs for Opus 5.5 to $10.51, but a 20% base price cut by Anthropic (from $5/$25 down to $4/$20) reduced it to $8.41, and a further 29% reduction in cache reads down to $0.20/1M brought it to $5.98.

Microsoft has pushed optional cumulative preview update KB5124006 for Windows 11 26H1, raising the build number to Build 28000.3086 after installation. The taskbar gains bottom, top, left, and right four-side docking along with a small taskbar mode, and the Start menu adds small and large layouts alongside "Automatic." This version also introduces default-disabled Admin Guard and Microsoft Execution Container (MXC) process isolation capabilities for AI coding agents.

FoxDev Studio allows Visual FoxPro projects to be opened and run on modern machines without rewriting, converting, or exporting, directly reading original files for projects, forms, class libraries, menus, and reports.

Tech analyst and former a16z general partner Benedict Evans points out that AI may automate much of the daily routine work of junior lawyers, consultants, bankers, and advertising professionals. He believes that work at the bottom of the pyramid will be reshaped, and those who have never worked at a law firm or Bain, BCG, or McKinsey will find it hard to truly understand how this change unfolds.

Meta's new agent Muse has been criticized for retracing the path of Facebook M, which similarly targeted tasks like restaurant reservations and theater ticket purchases. Launched in 2015, Facebook M was restricted to about 10,000 users due to high operating costs and faced reports of human intervention behind the scenes, ultimately being canceled in January 2018 less than three years after launch.

Lose yourself in your side project.

One of the developers from our GPT-6 hackathon brought us real star maps and Astra to help you find your way home.

At the 2026 Snapdragon Summit, Qualcomm CEO Cristiano Amon stated that the agentic era has arrived, reasoning has become critical, the edge will play a major role, and Snapdragon is the processor built for the agentic era. He stated that agentic experiences should be personalized, proactive, responsive, trustworthy, and effortless for users, seamlessly coordinating intelligence across phones, PCs, cars, glasses, and other devices.

The TechCrunch Founder Summit will take place on November 4 at the SoWa Power Station in Boston, and the official agenda has been announced, covering themes such as fundraising, hiring, and AI strategy. Lior Div, co-founder and CEO of 7AI, will share how to build an AI-native company from day one, while Chase Garbarino, co-founder and CEO of HqO, will discuss his experience building a company spanning over 30 countries out of Boston.

Unreal Labs has released Unreal Agent, which manages tool calls in a fully asynchronous manner, eliminating the overhead of model processing waits, polling, and heartbeats. The company states it saves up to 40% in costs compared to Codex and 20% compared to Pi on real-world workloads and agent benchmarks.

METR has published its pre-deployment evaluation summary for Claude Opus 5.5, concluding that the model offers a slight speedup in AI R&D over Fable 5.1 as an incremental improvement rather than a quantum leap, making it unlikely to fully automate AI R&D. Its development process may experience a roughly 1.5x speedup from AI (with about a 30% chance of reaching 2x), but it is unlikely to be massively accelerated.

Stanford professor and World Labs co-founder Fei-Fei Li has called for more oversight of AI by independent bodies and the public sector, arguing that increasingly powerful AI systems should not be left entirely to development companies to self-evaluate for safety. She stated that internal researchers at AI companies can evaluate individual models, but that cannot replace broader standards formulated jointly by governments, industry, and academia, and that industry and government must share responsibility.

A project named LLM Ass Bench scored 117 points on Hacker News, hosted at assbench.com. The page provides no specific information such as model versions, parameter counts, or evaluation scores, presenting itself simply as an "LLM-assisted bench."

Artificial Analysis published a post noting that GPT-6 Sol and Luna score similarly to their respective predecessors in the Intelligence Index while seeing token pricing drop by roughly 50%, cutting per-task costs in half. Their chart illustrates the trade-off curves between intelligence and per-task cost across generations of OpenAI models.

Grab the wheel and tear across open terrain. #PixVerseWorldModel changes the world without stopping. Reshape environments, summon rainbows, keep driving. The world responds to your actions.

Meet PixVerse R2, our brand-new real-time world model. Explore living worlds. Control and edit them with prompts. Shape stories. Meet characters that remember and respond.

arXiv math submissions surged 33.5% in 2026.

During the first 8 months of 2026, Mathematics recorded 47,127 arXiv listing entries.

This is not driven by a single popular topic. Out of 30 mathematics subfields, 29 showed growth, and the data also shows more authors publishing repeatedly at unusually high frequencies.

The paper suggests that research in the AI era now requires verifying as aggressively as generative scaling.