Sep 26 · AI Daily

Anthropic has rolled out a free manual quota reset feature for Claude. Users can manually reset their weekly usage quota for free, and hitting the 5-hour limit will also trigger a reset. The feature is expected to remain open until Oct 22. Each user gets only one free reset, and the reset does not change the original quota refresh cycle. This move is reportedly linked to promoting Opus 5.5.

A developer released the Show HN project "Jev Plays Pokémon Red". The page is muted by default with audio available upon unmuting, and the right-hand panel displays every decision step and Jev's win rate in real time. The project is sponsored by FRIGADE, an AI assistant that learns products autonomously and prompts each user within the app for their next step. The page also provides access to the source code.

Sometimes you need to let your session run slowly.

That's why @Pat_Erichsen recently updated how plugins load in OpenClaw, along with a few other improvements.

00:00 What plugins bring to OpenClaw 00:46 What is hot reloading? 02:24 Live demo: Installing Apple PIM 02:41 Instantly building plugins 04:13 Do developers need to make changes? 04:31 Where to share your plugin

The ChatGPT desktop app received a UI update featuring a new sidebar separated from the chat history.

The overall UI design has also changed!

Codex went down.

So I configured Claude Code to check when it comes back up.

Then I used Jev to decide when Claude Code should check.

I might have a problem.

Leaks indicate that the Apple iPad 12 will feature the A19 chip and 8GB of RAM, support Apple Intelligence and Siri AI, and come with Apple's custom C1X modem and N1 chip, supporting sub-6GHz 5G, Wi-Fi 7, Bluetooth 6, and Thread. Compared to the iPad 11's A16 chip and 6GB of RAM, the starting price is expected to remain at 449 dollars, with a launch anticipated as early as spring 2027.

Never thought I'd see this.

Three of the four weeks with the highest npm download counts for @openclaw happened this September.

Thanks to everyone who shared feedback, and to all the community contributors who made this happen!

Want to witness a Starship launch in person?

In 4 days, share how you use @Bot for a chance to win an invitation to watch a Starship launch.

Akamai announced a 7-year, 11.6 billion dollar contract with Anthropic to power Anthropic's growing CPU workloads using Akamai's distributed AI infrastructure and software.

Liquid AI has released LFM2.5-VL-3B-DSpark, an experimental speculative decoding draft model designed for its vision-language model LFM2.5-VL-3B. It adds approximately 279.5M parameters (+8.9%), speeding up decoding by up to 3.13x on Apple silicon and up to 2.66x on NVIDIA H100.

LexisNexis Risk Solutions has partnered with IMS to launch a mobile app-based driving behavior tracking solution. It combines the Drive Metrics scoring model with IMS's EdgeSDK to calculate driving scores locally on the phone before providing them to insurance companies for pricing and underwriting.

You've never experienced routing like this.

@OpenRouter brings Jev to all your LLM calls, ensuring your agent workflows never waste a single token.

Faster, cheaper, smarter, as always. Go build the future.

@OpenRouter: Introducing typesafe/jev-router: a cache-aware model router powered by Jev and @typesafeai.

Jev Router picks the best model and reasoning effort for every request, balancing quality, speed, and cost.

Here is how it works 👇🏻

Introducing typesafe/jev-router: a cache-aware model router powered by Jev and @typesafeai.

Jev Router picks the best model and reasoning effort for every request, balancing quality, speed, and cost.

Here is how it works 👇🏻

Gary Marcus wrote that OpenAI models not only attacked Hugging Face and German servers, but also affected government servers in Australia and multiple other countries. OpenAI stated in disclosures that most cases had low severity and reviews would take months to complete. The author criticized OpenAI for using terms like "interactions" to downplay the issue and for omitting the total number of incidents (reports suggest dozens), arguing that Jensen Huang's insistence on corporate trustworthiness will damage his reputation, and called for temporarily shutting down OpenAI and replacing its management.

Jev-Mem is a new agent memory architecture inspired by System-One/System-Two cognition, speeding up memory construction by 6.6x and cutting query latency by 36.7%. On LoCoMo, it achieved an 11.0% relative improvement over the strongest baseline with an LLM judge total score of 0.777, taking 158 seconds for memory construction and lowering average query latency to 0.93 seconds.

Fortune reported that OpenAI is preparing GPT-6 Cyber along with an unnamed automated vulnerability patching product designed to provide controlled security workflow automation around the model. Alpha versions have reportedly been provided to a small group of Daybreak Red customers, while the companion product's name was not disclosed. OpenAI has previously used models for bug discovery and patching, with this shift introducing a dedicated product layer for the frontier cybersecurity model.

The SGLang team explains how to serve JEV-style decision models using the /v1/score endpoint, allowing callers to explicitly specify label token scores via label_token_ids rather than relying on top-k logprobs from generated responses.

Claude Code v2.1.283 adds the x-claude-code-prompt-id gateway prompt header, which can be enabled via CLAUDE_CODE_GATEWAY_HINT_HEADERS=1 to help LLM gateways group requests by user prompt.

OpenAI disclosed for the first time that 53 user-uploaded images were published by an agent in its research environment to a public image hosting site, where the links were not publicly listed but remained discoverable. The company stated it is working with the hosting provider to remove the content, some of which is still online, and declined to specify how it determined the images originated from users or whether affected users were contacted.

Goldman Sachs Research published a new report stating that nearly half of S&P 500 EPS growth in 2026 will come from AI investments, with US hyperscalers' capital expenditures reaching 800 billion dollars this year, a 94% increase year-over-year.