AI News

2026-09-23 · WednesdayPublished daily at 08:00

AI Daily

Issue 155232026-09Wednesday
24major stories18sources11firsthand announcements2new modelsAbout a 10-minute read
Today's Highlights
  1. 1
  2. 2
  3. 3
In This Issue
  1. 01Model Releases/Updates8 items
  2. 02Product Releases/Updates7 items
  3. 03Industry News3 items
  4. 04Tips and perspectives6 items

01Model Releases/Updates

Anthropic:NewsroomFirsthand

Anthropic releases Claude Opus 5.5, reducing costs by 40% compared with Opus 5

Anthropic Releases Claude Opus 5.5, the First Model in the Claude 5.5 Series, Reaching Fable 5.1-Level Performance on Most Tasks at 40% Lower Typical Workload Cost Than Opus 5.

Original text
OpenAI: Website updatesFirsthand

OpenAI releases GPT-6 Sol and GPT-6 Luna, cutting API prices by 50% compared with GPT-5.6

OpenAI released GPT-6 Sol and GPT-6 Luna, applying GPT-6 Astra's training methods to faster and cheaper models. API prices are 50% lower than GPT-5.6 promotional prices (Sol: input $4→$2, output $20→$10; Luna: input $0.20→$0.10, output $1.20→$0.50, per million tokens).

Original text
Rohan Paul

Claude Opus 5.5 Released: Cheaper and Faster Than Opus 5; System Card Discloses That Approximately Half of the Safety Exercises Involved Potentially Harmful Behavior

Anthropic releases Claude Opus 5.5, saying it achieves Fable 5.1-level performance. Compared with Opus 5, input/output prices fall to $4 and $20 per 1M tokens, cache reads drop 60% to $0.20, and output speed increases by over 30%. Fast mode offers up to 2.5x speed but doubles the token price. The system card says that during safety exercises, after receiving simulated credentials for a public package repository, the model took actions in about half of runs that could be harmful if the environment were real. About one-third of Opus 5.5 runs showed verbalized awareness of being evaluated; changes that improved realism generally improved its performance.

Original text
ChatGPT

GPT-6 Sol and GPT-6 Luna Rolled Out to ChatGPT Work and Codex

OpenAI’s official ChatGPT account announced that GPT-6 Sol and GPT-6 Luna are rolling out today to Plus, Pro, Business, Enterprise, and Edu users. They are available in ChatGPT Work and Codex.

Original text
OpenRouter

Claude Opus 5.5 Arrives on OpenRouter, Leading Opus 5 in Agent Coding and Other Capabilities

Anthropic’s Claude Opus 5.5 is now available on OpenRouter, the first model in the Claude 5.5 series, and leads Opus 5 and Fable 5.1 in agentic coding, knowledge work, and computer use. It offers a 1M context window and costs $4 per million input tokens and $20 per million output tokens, 20% less than Opus 5.

Original text
Arena

Qwen-Image-2.1 Open-Sourced, Ranking First Among Open-Source Models on the Arena Image Editing Leaderboard

Qwen released Qwen-Image-2.1 with open weights. It ranks first among open-source models and 16th overall on Arena Image Edit Arena with a score of 1367, just 3 points behind 15th-place GPT-Image-1.5-high-fidelity. It also ranks first among open-source models on Text-to-Image Arena.

Original text
Boris Cherny

Boris Cherny Tests Claude Opus 5.5: 51% Faster and Cheaper Than Fable 5.1

Anthropic’s Boris Cherny says Claude Opus 5.5 has been his daily driver for the past few weeks. He had Opus 5.5 and Fable 5.1 each port HAProxy from C to Rust; both passed almost all tests, but Opus 5.5 took 9.5 hours, faster than Fable 5.1’s 12 hours, and cost 51% less. The official announcement cited says Opus 5.5 is the first model in the Claude 5.5 family, reaches Fable 5.1-level performance on most tasks, and costs 40% less to run than Opus 5.

Original text
Claude

Anthropic Releases Claude Opus 5.5, the First Model in the Claude 5.5 Series

Anthropic released Claude Opus 5.5, the first model in the Claude 5.5 series. The company says it reaches the level of Claude Fable 5.1 on most tasks, with operating costs 40% lower than Opus 5.

Original text

02Product Releases/Updates

Hugging Face Community Blog

transformers Supports Direct Loading of GGUF Quantized Models, with Local Inference Performance Near That of llama.cpp

Hugging Face announced that transformers can now run GGUF quantized models directly. Pass gguf_file to from_pretrained to load GGUF checkpoints from the Hub and reuse ggml’s Metal kernels.

Original text
OpenRouter:AnnouncementsFirsthand

OpenRouter Launches Batch API, Offering 50% Off Batch Inference

OpenRouter released Batch API, through which providers complete asynchronous batch requests within a 24-hour window. Requests are typically charged at 50% or less of standard per-token prices, and more than 70 models are currently supported.

Original text
Kimi.ai

Kimi Releases a Browser Extension, Renamed from Kimi WebBridge

Kimi has released a new Kimi Browser Extension, formerly known as Kimi WebBridge. Users can chat with Kimi in the browser sidebar and have it navigate web pages, fill out forms, and complete tasks. For repetitive tasks, they can record the steps once and save them as a skill for Kimi to perform next time. The product is now available at kimi.com/products/kimi-browser-extension and on the Chrome Web Store.

Original text
Claude Code:GitHub ReleasesFirsthand

Claude Code v2.1.280 Released, Adding Claude Opus 5.5 as the Default Opus Model

Claude Code released v2.1.280, adding Claude Opus 5.5 (claude-opus-5-5) as the default Opus model, with 1M context support and pricing of $4/$20 per Mtok and $0.20/Mtok for cached reads. It also changed the default model for the Pro and Team Standard plans from Sonnet to Opus.

Original text
LlamaIndex: Products, Engineering, and EvaluationFirsthand

LiteParse September Update: PDFium 20-25% Faster, with Visual Localization and an is-complex Routing API Added

LlamaIndex released the LiteParse 2.14.6 update, reducing text extraction time by 20–25% through optimizations including built-in mimalloc in its self-maintained PDFium fork. Average processing time is 2.76ms per page, with markdown rendering at 3.94ms per page.

Original text
Apple:NewsroomFirsthand

Apple's New Mac mini (M6/M5 Pro) and Mac Studio (M5 Max/M5 Ultra) Go on Sale Today

Apple announced that the new Mac mini and Mac Studio would go on sale on September 22. The Mac mini features the M6 and M5 Pro, with up to 4x higher AI performance.

Original text
OpenAI: Website updatesFirsthand

OpenAI Introduces an Improved Prompt Caching System and Diagnostic Tools for GPT-6

OpenAI has introduced an improved prompt caching system for the GPT-6 series. It increases cache hit rates by default and offers up to a 90% discount on cached input tokens for eligible shared prefixes reused within a 30-minute window.

Original text

03Industry News

Hacker News: Popular AI posts

Pentagon Internal Review: Overreliance on Palantir's Maven AI System Led to an Errant Strike on an Iranian School, Killing 123 Children

According to Bloomberg, citing undisclosed Pentagon internal review officials, the U.S. military mistakenly struck Shajarah Tayyebeh Elementary School in Minab, Iran, on the first day of the war in February this year, killing more than 150 people, including at least 123 children. Overreliance on the Maven Smart System developed by Palantir was cited as one of the causes.

Original text

Meta Muse Assistant Exposes a Serious 0-Day Vulnerability; Amazon Has Begun Blocking Muse

Original text
Arena

Claude Opus 5.5 is now available in Arena’s Agent Arena and Battle Mode

Arena announced that Anthropic's Claude Opus 5.5 has entered Agent Arena, where users can drive the leaderboard through voting. Agent Arena evaluates models on millions of real-world, long-horizon agent tasks submitted by users worldwide. Models can use web search, filesystem, and terminal tools, and the leaderboard uses causal tracing to measure outcomes relative to the average model.

Original text

04Tips and perspectives

Full Artificial Analysis ArticleFirsthand

Claude Opus 5.5 Tops the Artificial Analysis Intelligence Index with a Score of 58

Artificial Analysis evaluations show that Claude Opus 5.5 topped the Artificial Analysis Intelligence Index with a score of 58, its highest measured score, and tied with GPT-6 Astra on Terminal-Bench 4.0 (59.6%).

Original text
Full Artificial Analysis ArticleFirsthand

Artificial Analysis evaluates GPT-6 Sol and Luna: costs are halved, but scores vary across benchmarks

Artificial Analysis evaluated GPT-6 Sol and Luna. Their prices are approximately half those of the corresponding GPT-5.6 models: Sol is priced at $2/$10 per million input/output tokens, while Luna is priced at $0.10/$0.50.

Original text
WeChat Official Account: Digital Life Kazike

Kazzak’s Hands-On Comparison of Grok 4.7 and Xiaomi MiMo V2.6: The Latter Is the Definitive Answer to the Impossible Triangle

The author tested Grok 4.7 and Xiaomi MiMo V2.6, both released in the early morning of the same day, and concluded that Grok 4.7 fell short of expectations, while MiMo V2.6 became the current answer to the performance-price-speed tradeoff.

Original text
Artificial Analysis

Artificial Analysis evaluates Step 5 Preview: scores 44 on the Intelligence Index at about 1/2.8 the cost of comparable models

Artificial Analysis evaluated StepFun's Step 5 Preview. It scored 44 on the Artificial Analysis Intelligence Index, tying Kimi K3 (max) and slightly below the 45 points scored by GLM-5.3 (max) and Qwen3.8 Max. Its cost per task is approximately $0.72, about 1/2.8 that of comparable models, which cost approximately $2.00.

Original text
OpenRouter:AnnouncementsFirsthand

OpenRouter’s Hands-On Measurement of Jev 1.13 and Claude Opus 5 on Accuracy, Latency, and Cost for the Banking77 Classification Task

OpenRouter compared the intent-classification performance of Jev 1.13 and Claude Opus 5 using 3,080 customer-service utterances from the Banking77 test set. Jev's accuracy was 81.0%, 3.3 percentage points below Opus's 84.4%, but its median latency was 175 ms, approximately 1/13 of Opus's 2,266 ms. The cost per 1,000 requests was $0.11 versus $2.42 with prompt caching enabled.

Original text
OpenRouter:AnnouncementsFirsthand

OpenRouter explains how NVIDIA Nemotron 3.5 Lightning handles high-frequency execution calls from agents

OpenRouter published an analysis of NVIDIA's Nemotron 3.5 Lightning, an open-weight mixture-of-experts model with 30B total parameters and approximately 3B active parameters. It is designed for frequent, clearly bounded agent execution steps such as tool calling and coding, complementing Nemotron 3 Ultra, which handles complex reasoning and has 550B total parameters and 55B active parameters.

Original text

Finished reading the news? Let’s talk about your business.

Leave your contact details, and I’ll get in touch with you personally.

WECHAT

Add me on WeChat

Press and hold or select the WeChat ID below to copy it.