22AugSaturday2026Index中文EN

Daily AI Digest

MODEL QUICK LOOK · CHECKED DAILYWhich model should you use today?Compare capability, price, release date and strengths across 30 general modelsOpen latest ranking →

Top Stories


DeepSeek · V4 Flash Vision Exp · Multimodal API

DeepSeek adds image input to V4 Flash across three API formats

DeepSeek has released the experimental V4-Flash-Vision-Exp. It retains V4 Flash's text, reasoning and agent capabilities while adding image understanding. DeepSeek says its multimodal-agent performance approaches Claude Opus 4.8, but this is a vendor-reported comparison that still needs independent replication.

Developers can send mixed text and images through Chat Completions, Messages or Responses. Images may arrive as base64, external URLs or Files API references, consume up to 384 tokens each and use V4 Flash pricing. The Files API is free and lets an uploaded image be reused by file_id; DeepSeek Harness 0.1.1 supports the model.

Verdict: Teams building screenshot understanding or visual agents can begin controlled tests, but should benchmark their own tasks and restrict both external image URLs and tool permissions.

DeepSeek · Aug 21 main window · 2 sources
Anthropic · Claude Mythos 5 · Cyber defense

Claude Mythos 5 begins scanning enterprise code as Anthropic commits $35 million to open-source defense

Anthropic has added Claude Mythos 5 to the Claude Security public beta. Claude Enterprise customers can scan repositories they are authorized to test and receive findings with CWE categories, confidence, severity and suggested fixes. Scans use standard token billing under the existing plan, the model is not exposed directly, and every patch still requires human approval.

Anthropic is also integrating Mythos 5 into partner security products and plans to expand its Cyber Verification Program for vetted defenders. The new Defender Advantage Fund will provide $35 million in Claude credits for open-source projects patching active vulnerabilities, building reusable scan-and-fix automation and pursuing broader security improvements.

Verdict: Enterprise defenders should start with read-only scans and human-approved patches. Understaffed open-source maintainers should watch the fund, but model findings are leads rather than final security judgments.

Anthropic · Aug 21 · 2 sources

Tools and Products


Claude Code · v2.1.239 · Billing and session fixes

Claude Code 2.1.239 fixes proxy-driven double billing and expands cross-machine sessions

Claude Code 2.1.239 fixes a billing bug triggered when some proxies stripped Content-Type from Bedrock streaming responses, causing every turn to be retried in non-streaming mode. Cost estimates now include the 1.1-times U.S.-only inference premium, and a depleted monthly-spend message also shows session or weekly reset timing.

The release repairs deleted-workdir startup, long-conversation titles, cloud plan-mode recovery and MCP reconnects after transient 5xx errors. Windows now supports cross-session SendMessage and ListAgents, while mobile images uploaded to remote sessions include their saved paths.

Anthropic GitHub · Aug 21, 12:54 PM Pacific
Claude Academy · Free courses · Official learning hub

Anthropic launches Claude Academy as a free home for courses, tutorials and team use cases

Claude Academy serves people exploring AI, users starting with Claude and organizations rolling it out to teams. Anthropic maintains AI Fluency material, guided starts, tutorials and use cases in one searchable learning site.

It targets the gap between being able to chat with a model and using it consistently at work. Teams can use the official material as a common baseline, then add their own data permissions, review processes and task templates.

Anthropic · Aug 20, 8:56 PM Pacific

Community Pulse


Anna's Archive · Destructive scanning debate · 517 HN points

Books dismantled for AI training revive fears that public knowledge may survive only in corporate copies

Anna's Archive says AI companies have bought secondhand books through intermediaries, cut them apart for scanning and destroyed them to obtain pre-2022 text uncontaminated by machine-generated material. It points to Anthropic's Project Panama as a disclosed case and asks volunteers to preserve rare and fragile collections first.

This is an advocacy post from a shadow library. Buying, destructive scanning and disposal appeared in Anthropic copyright litigation, but the article's broader claims about company motives, scale and knowledge monopolies are not automatically established facts. The high-volume HN debate focused on the collision between training demand, preservation and copyright.

Hacker News · Aug 20, 7:38 PM Pacific · 517 HN points when checked
AI writing fatigue · Workplace documents · 252 HN points

A developer says low-effort AI documents now trigger automatic inattention

Rafal Cymerys describes losing focus when a design document, marketing deck or requirements brief reads like low-effort AI output. The problem is broader than stock phrases: every detail is framed as a breakthrough and simple decisions arrive wrapped in long, uncertain reasoning, forcing readers into follow-up exchanges.

He calls the response AI-blind, analogous to banner blindness. It is a personal essay rather than user research, but the operational warning is concrete: generating more text does not transmit more information, and the sender still has to remove filler and own the judgment.

Hacker News · Aug 21, 4:48 AM Pacific · 252 HN points when checked
Self-hosted agent factory · Isolated hardware · 77 HN points

An isolated home server built, tested and deployed an HTTPS app from one prompt

Jake Saunders assembled an agent software factory on a dedicated home PC. Hermes calls Codex while Forgejo, CI runners, Coolify, Postgres and internal DNS let one instruction create a repository, write code and tests, pass CI, provision a database and deploy an HTTPS service.

The main safeguard is structural rather than trust: the agent runs on separate hardware with no public ingress, so the worst case is rebuilding the box. Inference, Tailscale, DNS and certificate services still leave the machine, which is why the author calls it almost self-hosted and does not present the post as a complete copy-and-paste guide.

Hacker News · Aug 21, 9:28 AM Pacific · 77 HN points when checked

Builder Perspectives


OpenAI Codex product lead Thibault Sottiaux says Codex did not quietly change ordinary subscription limits. Many affected users were using sub2api to convert subscription access into API traffic and re-serve it to multiple users, behavior flagged by fraud controls; official clients and open-source clients such as Pi or OpenCode remain supported when they use Sign in with ChatGPT. He also showed that GPT-Image-2 can now create transparent images in ChatGPT and the API.

Usage clarification →Transparent images →

Box CEO Aaron Levie argues that applied-AI companies can use post-training to reduce cost and improve accuracy once they understand a domain and have enough repeated work. Rewarding efficient tool use and reasoning can cut inference tokens at comparable quality, but bespoke training is unnecessary when task volume is low or a general model is already sufficient.

View on X →

GitHub Trending


GitHub Trending · Salesforce sf-skills 1.41.0

Salesforce adds 26 agent skills and updates 11 existing ones

sf-skills is Salesforce's application-development skill collection for Agentforce Vibes and other AI tools. Version 1.41.0 added 26 skills and updated 11 inside the window. It had 892 stars when checked and uses Apache-2.0.

Agent skills · 892 stars when checked · Apache-2.0
GitHub Trending · FastVideo · MiniMax H3

FastVideo adds few-step MiniMax H3 generation and cuts encoder and decoder overhead

FastVideo is an inference and post-training framework for video generation. In-window commits added a MiniMax FastH3 Preview few-step path and reduced H3 text-encoder memory and VAE decoding cost. It had 4,022 stars when checked and uses Apache-2.0.

Python · 4,022 stars when checked · Apache-2.0
GitHub Trending · Google ADK Samples · Virtual try-on

Google ADK Samples adds a retail virtual-try-on skill

Google's ADK sample repository collects agents built with Agent Development Kit. An in-window commit added a retail/virtual-tryon skill as a reference for virtual try-on workflows. It had 10,202 stars when checked and uses Apache-2.0.

Python · 10,202 stars when checked · Apache-2.0
Editor's note

The common thread is putting stronger models inside constrained workflows. DeepSeek reduces repeated image transfer with Files API, Anthropic delivers Mythos 5 through scans, partner tools and human-approved outputs instead of direct access, and a self-hosted experiment puts isolated hardware ahead of trusting the agent. As models gain agency, access, permissions, billing and review have to be designed together.