Skip to content

Weekly digest

10 – 16 August 2026

17 posts

OpenAI added Ultrafast, a new service tier in its API that runs GPT-5.6 Sol on Cerebras chips at up to 750 output tokens per second. OpenAI says that is up to 14x the model’s normal speed. The tier is in preview with a limited group of API customers.

Google introduced Gemini 3.7 Flash, per Testing Catalog posts on Threads, which claim it is 50% cheaper than Gemini 3.6 Flash and scores 65% on DeepSWE V1.1. The model is already live in the API, AI Studio and Antigravity.

Claude Code 2.1.227 stops pitching credits to Max subscribers

Three bug fixes and two interface tweaks in this release.

• A session started with an expired login token evaluated feature flags without accounting for the plan, so Max subscribers were wrongly prompted to enable credit billing for Fable • In claude-code-action with allowed_non_write_users, every Bash command failed on GitHub runners • /tui returned a conversation rolled back to a state before the first message • In the slash-command menu, only the selected row is marked in blue, matched characters are bold, and emoji and diacritics in names survive intact • Fewer event loop stalls on missing-file suggestions and on size checks for @ mentions

All of it is live in Claude Code 2.1.227.

Claude Code ReleasesDiscuss

OAuth login to MCP servers is fixed

Release 2.1.231 ships with a single fix in the changelog: OAuth login was failing with a redirect URI mismatch error on servers that use a pre-registered OAuth client, Slack among them.

Claude Code ReleasesDiscuss

Claude’s iOS app now has a Usage Credits section

Testing Catalog reports that a Usage Credits section, showing usage credits, has gone live in the Claude app for iOS.

The same source says profile photo support is in development. Testing Catalog previously reported that multi-account support is being prepared for Claude’s mobile apps.

Testing Catalog ThreadsTesting Catalog ThreadsDiscuss

Anthropic publishes its second Risk Report

Anthropic has released its second Risk Report, a document detailing the risks its systems pose and how prepared the company is to address them.

The reports come out on a regular schedule under Anthropic’s Responsible Scaling Policy. The new one is up on Anthropic’s site.

Anthropic TwitterDiscuss

Claude may get a side-by-side model comparison mode

Per Testing Catalog on Threads, Claude may get a Model Comparison interface, unannounced by Anthropic, that shows how responses from different models differ. The post claims users could turn memory on and off, “hold” responses, and, unlike an A/B testing UI, switch between normal chats and comparison mode on demand.

Testing Catalog ThreadsDiscuss

OpenAI’s new Ultrafast tier runs GPT-5.6 Sol on Cerebras chips

Ultrafast, a new service tier in the OpenAI API, pushes GPT-5.6 Sol to up to 750 output tokens per second. OpenAI says that’s up to 14x the model’s normal speed.

The tier is in preview and starts with a limited group of API customers. OpenAI plans to widen access as capacity grows.

OpenAI NewsOpenAI TwitterDiscuss

Codex adds an experimental thread/revert request

Codex’s app-server now has an experimental thread/revert request for rolling a thread back. The commit touched 24 files: 939 lines added, 21 removed.

• The loaded thread’s long-term paginated history is replaced with the slice preceding the beforeTurnId turn, and the thread ID stays the same. • The active turn is interrupted, history reloads without unloading the thread, mutable thread settings are preserved, and a thread/reverted event fires. • Backward pagination cursors are returned for the remaining turns and items. • Stale rollout paths are rejected after the revert, and local file changes are untouched.

The work landed in pull request #38440.

Hacker News CodexDiscuss

OpenAI’s revenue run rate tops $40 billion, per Bloomberg

OpenAI is now on an annualized revenue pace of more than $40 billion, roughly double its level at the end of 2025, Bloomberg reports. The outlet’s sources spoke anonymously because the figures aren’t public.

Growth picked up over the past few months on the back of AI coding tools, subscription sales, and an advertising business still in its early stages. The consumer side is growing too.

The numbers land as OpenAI preps for a public listing. Ahead of it, the company bought back roughly $7 billion in employee shares at an $852 billion valuation.

Hacker News CodexDiscuss

Microsoft claims a No. 2 Arena finish for MAI-Image-2.6. The leaderboard says otherwise.

Microsoft says MAI-Image-2.6 debuted on Arena and landed second in the text-to-image ranking, with access in MAI Playground and through the API promised for next week.

Public data doesn’t back that up. As of August 10, 2026, second place on the Arena board belongs to grok-imagine-image-2.0, and mai-image-2.5 sits ninth. There’s no separate 2.6 announcement in Microsoft’s public materials either; the second-place finish the company has documented was for MAI-Image-2.5 in the image editing category.

Microsoft did previously open a public preview of MAI-Image-2.5-Pro in Microsoft Foundry and MAI Playground, priced at $106 per 1M output tokens.

Testing Catalog ThreadsDiscuss

Codex and ChatGPT Work now share the same projects and chats

OpenAI says projects, chats, skills, and plugins are kept in a single synced state across Codex, ChatGPT Work, and other AI agents. Start a project in one, and it’s the same project in the other.

The project and chat sync between Codex and ChatGPT Work is described in OpenAI’s own post and in a Crypto Briefing writeup dated August 11.

Testing Catalog ThreadsDiscuss

xAI’s voice agent can now text you mid-call

Per a post on Threads, xAI’s voice agent builder picked up post-conversation actions: outbound SMS, email, and API calls.

• Outbound SMS through Twilio numbers: the agent texts the person it’s currently talking to, from an imported number, in the middle of the call. Meant for confirmations, links, and anything easier to read than to hear. • System variables in API tools: the caller’s number gets passed into the request so the agent can pull up their account and hand back personalized data. • Post-call email notifications: the message goes to up to three recipients and includes call duration, the channel, and a link to the recording.

The announcement says emails have a minimum call length setting so short calls don’t trigger notifications. We couldn’t find independent confirmation of the update in press coverage or in xAI’s own materials.

Testing Catalog ThreadsDiscuss

Codex limits reset again, alongside an unconfirmed 15 million user claim

Codex usage limits were zeroed out on August 13, according to services that track resets. The post announcing it credited a 15 million user milestone and separately called out Codex Realtime as a big opportunity. Neither the number nor the status of Codex Realtime is backed by any other source.

The publicly confirmed baseline is smaller. OpenAI said more than 5 million people were using Codex weekly as of June 2. Testing Catalog later put the combined weekly audience for ChatGPT Work and Codex at 10 million.

The previous limit reset landed on July 29, when an OpenAI employee flagged the return of the five-hour cap in Codex.

Testing Catalog ThreadsDiscuss

Gemini 3.7 Flash lands at half the price of 3.6 Flash

Google has introduced Gemini 3.7 Flash, according to posts from Testing Catalog on Threads. The claims:

• 50% cheaper than Gemini 3.6 Flash • 65% on the DeepSWE V1.1 benchmark, close to GPT-5.6 Terra • better tool use when driving Google Workspace apps • rolling out in Gemini Spark, which the same post says is available to Pro and Ultra subscribers in 160 countries

The model is already live in the API, AI Studio and Antigravity. Other products get it as the rollout continues.

Testing Catalog ThreadsTesting Catalog ThreadsDiscuss

ChatGPT Business Premium seats go out at $100 a month

OpenAI has started pulling companies off the waitlist for Premium seats in ChatGPT Business, priced at $100 per month. OpenAI describes the seat as 5x the usage of a Standard seat, with the five-hour limit removed and the quota resetting weekly on a predictable schedule.

Workspaces can mix Standard and Premium seats. Early participants can pick up as much as $500 in extra credits, calculated at $100 per Premium seat added. That five-hour window Premium drops is the same one OpenAI brought back to Codex in late July after temporarily scrapping it.

Waitlist applications tied to the promo credits are open through August 20.

Testing Catalog ThreadsDiscuss

Zhipu AI will release GLM-5.3 weights in stages after safety reviews

Zhipu AI announced GLM-5.3 in ZCode on August 14. The model is built for cyberdefense and coding work, and the company says API access and open weights will both be released in stages, after strict safety checks.

A post on Threads claims GLM-5.3 scored 84.5 on CyberGym, beating Mythos 5 and GPT-5.6 Sol, and that it was trained on top of a 743B-parameter base model. Zhipu’s official materials don’t confirm any of those numbers.

For reference on the same benchmark: Microsoft’s MDASH multi-agent scaffold scored 88.45%, according to the company’s own materials.

Testing Catalog ThreadsDiscuss

China gets an Apple-trained LLM, not Gemini

Reuters reports that Apple Intelligence in China will run on a large language model Apple trained itself, in partnership with Alibaba. Qwen features in the Chinese version too, with the Apple-trained model doing the primary work.

In the US and most other territories, Apple Intelligence uses Google Gemini technology, while its Foundation Models are entirely Apple’s. Apple has also shipped ChatGPT extensions for Siri, and none of those options, Gemini, ChatGPT, or Anthropic’s Claude, are available in China.

Apple InsiderDiscuss