Skip to content

Weekly digest

31 August – 6 September 2026

23 posts

OpenAI put GPT-6 Astra on the OpenAI Platform and APIs a day after its September 3 release, at $10 per million input tokens and $50 per million output, the same API price Anthropic charges for Fable 5.1. Per OpenAI’s rollout update, access in ChatGPT Work and Codex may take a few days to reach Plus and all Business users.

Claude Fable 5.1 scores 66 at max effort on the Artificial Analysis Intelligence Index, the top result on that board. Per Artificial Analysis, it costs 20% more per task than Fable 5, despite the 75% cut to cache read pricing that shipped with it.

Claude’s 20x plan runs 10x on the weekly limit

The 20x in Claude’s top plan applies to the 5-hour limits, while the weekly allowance works out at 10x, so one label covers two meters that scale differently.

The comparison it’s set against is Codex. Per the quoted clarification, both Codex plans called 20X apply specifically to weekly usage limits, and there are no “5h limits for both Pro plans”. Codex’s Pro 20X is described there as “quite precisely 20X the usage of the Plus subscription”.

Hacker NewsDiscuss

Court finds Pentagon’s Anthropic ban unconstitutional

Federal Judge Rita Lin ruled in San Francisco that the Department of Defense violated the First Amendment by designating Anthropic a supply chain threat. The threat status remains in effect while litigation continues, Bloomberg reports.

The Pentagon placed Anthropic on the blacklist in March after negotiations broke down. Anthropic had sought guarantees that Claude would not be used in fully autonomous weapons systems.

Judge Lin stated that the Defense Department created the designation out of a desire to make a public example of the company.

Hacker NewsDiscuss

Claude Code 2.1.252 fixes Remote Control stalls

The update fixes Remote Control sessions hosted by Claude Desktop or VS Code that stalled for minutes after a tool finished when the connection to claude.ai was degraded. It also fixes Bash failures on some Macs and project settings that didn’t save “always allow.”

Claude Code ReleasesDiscuss

Claude iOS shows a Max effort usage warning

The reported Claude iOS interface puts a “1.5 more usage” warning beside Max effort, giving subscribers a stated usage cost before they choose it. Anthropic’s documentation says higher effort uses more tokens and reaches usage limits faster, but it doesn’t substantiate the reported multiplier or iOS change.

Testing Catalog ThreadsDiscuss

Fable 5.1 tops Artificial Analysis at 66, costs 20% more

Leading the Artificial Analysis Intelligence Index carries a bill: per task, Claude Fable 5.1 runs 20% more than Fable 5, despite the 75% cut to cache read pricing that shipped with it. At max effort it scores 66 on the index, the top result on Artificial Analysis’s board.

Artificial Analysis says it supported Anthropic with pre-release evaluation of Claude Fable 5.1.

Hacker NewsDiscuss

Computer use lands in Claude Cowork and Claude Code beta

Claude can work your desktop in the background now, clicking, typing and opening apps while you do something else. The beta runs on Pro and Max plans in the macOS desktop app, and it’s switched on in Settings → General → Computer use.

Claude ThreadsDiscuss

A $200bn Google finance machine for Anthropic, per the FT

The Financial Times pins a $200 billion figure on the Wall Street financing Google has assembled around Anthropic, calling it a finance machine in its headline. The rest of the report sits behind the FT’s paywall, so the breakdown of that sum stays with subscribers.

Hacker NewsDiscuss

Slash commands are coming to Claude Code on iOS

According to a Threads post, Claude Code on iOS will get support for / commands, with no rollout confirmation from Anthropic and the poster asking whether anyone has them already. Running a command in the mobile app means typing its full name: the GitHub issue tracker still listed the missing autocomplete as a known usability gap in March 2026.

Testing Catalog ThreadsDiscuss

GPT-6 Astra edges past Claude Fable 5.1 on WebDev Arena

As of September 5, OpenAI’s GPT-6 Astra sits on top of Code Arena’s WebDev leaderboard, 35 points ahead of the recently released Claude Fable 5.1. The gap reads 1,797 to 1,762 on a crowdsourced Elo-style rating.

The write-up credits Astra with reshaping the Pareto frontier as “the best-performing model at $40/Mtoken”, pricing that matches the latest Claude model. Fable 5.1 arrived days earlier and took first place in the Artificial Analysis Intelligence Index.

Testing Catalog ThreadsDiscuss

Claude Fable 5.1’s “Never” rules pulled into one document

An artifact called “Claude’s Nevers”, posted by @vivek, gathers the “Never do this” lines from what it presents as the Claude Fable 5.1 system prompt, so a developer building on Claude can read the hard refusals in one place. Neither the artifact nor its provenance is independently verified, and it reaches far beyond the song-lyrics section published on September 2. The file, ANTHROPIC/Claude-Fable-5.1.md, runs 2,196 lines and 40,046 words.

Hacker NewsDiscuss

AI-written GPU kernels ship without a line-by-line read

SemiAnalysis’ Jordan Nanos says OpenAI engineers scrolling their own kernel code couldn’t tell what it did line by line, and it didn’t matter to them because the AI wrote it. His description: “It’s not slop, because it performs, but it’s literally just AI-generated assembly basically puked out in Gluon.”

Hacker News CodexDiscuss

A full LibreOffice ships inside the ChatGPT desktop app

The ChatGPT desktop client carries everything it needs to open documents on its own machine. A dig through ~/.cache with OmniDiskSweeper turned up 1.7GB in a folder called codex-primary-runtime, left there by OpenAI’s Codex desktop app, since rebranded to just ChatGPT.

Inside sit a full Python installation, a full Node.js installation, and native binaries for Poppler, git, and the LibreOffice office suite. The runtime’s documents plugin folder holds the skills that tell Codex how to find and use those binaries.

Hacker News CodexDiscuss

US government sides with OpenAI in NYT lawsuit

Reuters reports that the U.S. government has backed OpenAI in the copyright case brought by the New York Times, putting Washington behind an AI lab in court against a newspaper. Reuters published the story on September 2 and doesn’t spell out what the government filed. OpenAI’s copyright docket also includes the German suit from publisher Carlsen over the children’s book “NEINhorn”, filed on August 19.

Hacker News CodexDiscuss

Two more newspapers take OpenAI to court

The Seattle Times and Newsday have sued OpenAI and Microsoft over copyright infringement, Reuters reported on September 5, putting two more publishers on a docket that already carries the New York Times case. The complaint was filed in the U.S. District Court for the Southern District of New York.

According to Channel News Asia, the newspapers allege that the companies’ AI products can reproduce passages from their reporting, closely paraphrase articles and provide answers that reduce users’ need to visit their websites or buy subscriptions. Most same-day coverage reproduces the Reuters report, and neither company’s response has surfaced in it.

Three days earlier, Reuters reported that the U.S. government had backed OpenAI in the New York Times copyright case.

Hacker News CodexDiscuss

Coding agents spread fastest among OpenAI researchers

Early internal data from OpenAI’s engineering blog shows its researchers leaning on coding agents through the day, often in concurrent sessions, with total usage growing faster than among the company’s other teams. The post covers agent usage, experiment velocity, task complexity and research acceleration.

OpenAI NewsDiscuss

Hybrid mode would split Computer between Mac and cloud

TestingCatalog reports that Perplexity is preparing a Hybrid mode for Computer on Mac, where the cloud orchestrates and delegates “suitable subtasks” to a model running on the user’s own machine. Gemma 4 is listed for 16 GB Macs, with Qwen3.6 and Perplexity’s own model for 32 GB. A component called Privacy Gate runs entirely on the Mac, inspecting anything headed to the cloud for personal or sensitive information and leaving the user to decide how to handle it before anything is sent. Perplexity’s existing local route, Portable Computer, runs on Nvidia hardware instead.

Testing Catalog ThreadsDiscuss

Agent Studio on GCP lists Gemini 3.8 Flash

Developers on Google Cloud can select Gemini 3.8 Flash inside Agent Studio, according to a post flagging the rollout. The pitch: complex multimodal data processing, coding use cases, and software engineering-related agentic tasks, with named uses including processing data with images and text, coding problems, and web research and application testing.

It follows Gemini 3.7 Flash, which the same account said shipped in the API, AI Studio and Antigravity at 50% cheaper than 3.6 Flash.

Testing Catalog ThreadsDiscuss

Android build of Grok Bot hits Google Play

xAI’s Grok Bot agent is downloadable on Android, putting the early beta on phones. Access runs through SuperGrok Heavy, Cursor Ultra, and Cursor Teams Premium subscriptions, and each bot gets its own dedicated cloud computer.

Testing Catalog ThreadsDiscuss

DeepSWE 1.1 score puts Muse Spark 1.3 above Opus 5

Meta says Muse Spark 1.3 scored above GPT-5.6 and Opus 5 on the DeepSWE 1.1 coding benchmark, and the model is already rolling out on Meta’s model APIs. The comparison comes from Meta’s own announcement.

Two more releases are queued: a next big model upgrade codenamed “Watermelon” and an open-weight version of Muse Spark. Muse Spark 1.3 landed on September 2, 2026, under a month after 1.2, which shipped August 5, 2026 and sits on OpenRouter’s contributor tier.

Testing Catalog ThreadsDiscuss

Gemini Live adds voice commands for Workspace tasks

Gemini Live can set up complex, multi-step tasks across Google Docs, Sheets, Drive, and the web by voice, putting those tools into one spoken workflow.

Gemini Live also has deep Gmail integration. Users can search, summarize, star, archive, or delete messages hands-free.

Testing Catalog ThreadsDiscuss

Lyria 3.5 lands in the Gemini app, AI Studio and APIs

Full-song generation is available in the Gemini app, Google AI Studio and the APIs, where Google has released Lyria 3.5. It’s the same surface set Google used for its recent Gemini model drops.

Google calls it an “advanced, full-song generative model with enhanced audio fidelity and vocal clarity, optimized for precise prompt adherence and rich, cohesive musical arrangements”. The model first shipped in Google Flow Music on July 29.

Testing Catalog ThreadsDiscuss

GPT-6 Astra lands on the API at $10 in, $50 out

Developers get OpenAI’s new flagship before some paying ChatGPT users do. Astra is live on the OpenAI Platform and APIs a day after its September 3 release, priced at $10 per million input tokens and $50 per million output.

Access in ChatGPT Work and Codex may take a few days to reach Plus and all Business users, according to OpenAI’s rollout update. Enterprise customers in the Daybreak program got the model first, and the API price matches what Anthropic charges for Fable 5.1.

Testing Catalog ThreadsDiscuss

Custom MCP servers reach Gemini Business in preview

Admins on the Business edition of Gemini Enterprise can hook up their own MCP servers, which puts private company data and internal tools within reach of Gemini. Google’s own listing marks the custom MCP server connection as a Public Preview feature, so it’s an early-access capability rather than a general release.

Model Context Protocol is the open standard for wiring external data sources and tools into an AI assistant. Google says a custom MCP server connection also covers custom internal tools and MCP-compliant third-party systems. The documentation says nothing about custom MCP support in consumer Gemini, and the setup page carries a last-updated date of July 29, 2026.

Testing Catalog ThreadsDiscuss