Skip to content

Weekly digest

17 – 23 August 2026

15 posts

OpenAI is cutting API and credit pricing for GPT-5.6 Sol by over 20%, and OpenAI says the lower rate holds for the next 3 months. Sol was the model left out when Luna and Terra got cheaper on July 30, and Pro, Plus and Business subscription usage stays as it was.

Claude Code v2.1.237, which went out on GitHub on August 20, fixes prompt caching for sessions running through an LLM gateway or a custom base URL. The same release adds a built-in Concise output style, selected under Output style in /config.

Perf win of the day: Claude Code CLI now uses 2x less CPU at p99. Bun’s garbage collector was running on a fixed tim…

AcBun’s garbage collector waits for idle in Claude Code

Mid-turn, the CLI was losing CPU to its own runtime: Bun’s garbage collector ran on a fixed timer, so it kicked in right when Claude Code was busiest. The GC now holds off until the process is idle, and the perf note claims 2x less CPU at p99.

The note doesn’t name the Claude Code release carrying the change. Claude Code has run on the Rust port of Bun since v2.1.181, which shipped June 17.

ClaudeDevs TwitterDiscuss

Editable artboards land in Claude Code via /design

Claude Code takes on UI design work through /design, a research preview skill that carries Claude Design’s artboard workflow into the CLI and Desktop. Run the command to get editable artboards for your interface, pick one, tweak it, then have Claude implement it. The skill is built on artifacts.

Anthropic lists /design on the Pro, Max, Team, and Enterprise plans, and Claude Code has to be updated before the command shows up.

ClaudeDevs TwitterClaudeDevs TwitterDiscuss

Claude Desktop boots ~2x faster after a timer fix

Starting in the background used to throttle the app’s timers and drop the JS engine into power-saving mode. Anthropic says it now boots at full speed even while the window’s still hidden, ~2x faster than a month ago, along with smaller perf fixes.

ClaudeDevs TwitterDiscuss

Claude Code v2.1.237 adds a Concise output style

Sessions running through an LLM gateway or a custom base URL get prompt caching fixed in Claude Code v2.1.237, and the same release adds a built-in Concise output style: Claude leads with results and skips preamble and narration, while doing the work just as thoroughly. Select it under Output style in /config. The release went out on GitHub on August 20.

Claude Code ReleasesDiscuss

Claude Managed Agents get an AG-UI cookbook

A new Anthropic cookbook pairs Claude Managed Agents with AG-UI, CopilotKit’s open Agent-User Interaction Protocol, so one agent can drive many different front ends. The adapter maps each chat thread to a managed session and streams text, tool calls and thinking into the interface.

ClaudeDevs on Twitter points to the sample code, which sits in Anthropic’s claude cookbook repository on GitHub.

ClaudeDevs TwitterClaudeDevs TwitterDiscuss

Claude Academy opens with free courses and tutorials

Free courses and tutorials on working with AI are live at academy.claude.com, with paths for people figuring out what AI is and for those already using Claude every day. Anthropic published a companion blog post on how it approaches teaching and learning AI and where Claude Academy sits in that.

Claude ThreadsClaude TwitterDiscuss

Cheaper rivals outdraw Anthropic’s best model, FT says

According to the Financial Times, Anthropic’s strongest AI model is struggling to attract users while cheaper tools thrive. The article sits behind the FT’s paywall, and the visible text names neither the model nor the competing tools.

Bloomberg earlier reported preliminary second-quarter revenue above $11.5 billion for Anthropic, against $4.73 billion in the first quarter of 2026.

Hacker NewsDiscuss

18% quarterly growth disappoints OpenAI investors

The Wall Street Journal reports that OpenAI’s second-quarter results landed badly with some shareholders who wanted faster progress against Anthropic. The company told investors revenue grew 18% from the first quarter, reaching $6.7 billion in the three months ended in June.

Losses deepened over the same stretch. OpenAI’s operating margin sank further into the red, people familiar with the matter said, pushing it away from profitability ahead of a much-anticipated IPO. Anthropic’s annualized run rate hit $65 billion at the end of July, a number Bloomberg first reported and CNBC confirmed.

Hacker News CodexDiscuss

Over 20% off GPT-5.6 Sol, but only for 3 months

OpenAI is dropping API and credit pricing for GPT-5.6 Sol by over 20%, and the discount runs on a clock: the company says it holds for the next 3 months. Sol was the model left out when Luna and Terra got cheaper on July 30.

The lower rate is live on the API and rolling out across eligible plans for ChatGPT Work and Codex credits. Pro, Plus, and Business subscription usage stays as it was.

Hacker News CodexDiscuss

AGENTS.md request stays open on Claude Code tracker

A feature request on the Claude Code tracker asks for support of AGENTS.md, the shared Markdown file Codex, Amp and Cursor are standardizing around, arguing that CLAUDE.md “feels too specific to Claude Code” when you collaborate with developers who don’t use it. Issue #6235 was still listed as open with a last update of July 21, 2026, while a related request, #50778, was closed as a duplicate.

Hacker NewsDiscuss

ChatGPT ads cross into 31 European countries

Free and Go users in 31 European countries will start seeing ads in ChatGPT next week, OpenAI says, extending the format that reached those same plans this month. The ads run as standalone widgets below the answer, and the paid tiers above Go carry none.

Advertisers in the region are also expected to get access to the Ads platform. Digiday notes the European launch lands at the six month mark for OpenAI’s ads business, which started as a U.S. pilot in February.

Testing Catalog ThreadsDiscuss

Plan mode spotted in AI Studio’s Build section

An unconfirmed sighting has Google AI Studio’s Build section picking up a Plan mode that creates a plan before making changes. Google hasn’t documented it there. Per Google’s blog, Plan mode currently ships in Gemini CLI as a read-only mode that restricts the agent to safe exploration tools.

Testing Catalog ThreadsDiscuss

Transparent backgrounds arrive for gpt-image-2

Developers can pull images out of OpenAI’s API with the background already gone, so a cutout step drops off the pipeline. OpenAI lists product cutouts, logos, stickers and presentation graphics as the supported cases for gpt-image-2.

Testing Catalog ThreadsDiscuss

Vision-Exp brings image input to DeepSeek’s API

DeepSeek opened API access to DeepSeek-V4-Flash-Vision-Exp, a multimodal model that reads text and images in the same prompt. The company says it delivers performance close to Opus 4.8 and matches DeepSeek-V4-Flash on text.

Calls go through Chat Completions, Messages and Responses, the endpoint set the V4-Flash line picked up during the 0731 beta. The model sits on the DeepSeek API Platform, nine days after the V4-Pro general availability build went live on August 12, 2026.

Testing Catalog ThreadsDiscuss