Weekly digest
This week was mostly Anthropic. The item with the most immediate practical weight is the Auto Mode break in Claude Code running Claude Opus 5, which Embrace The Red wrote up. The bug was reported on August 24 and is still open, and Auto Mode has been the default for new sessions on Pro, Max and Team since August 14.
Those same plans get a permanent 25% raise to weekly limits on September 14. Testing Catalog points out that this leaves subscribers about 17% below what they have now, because the temporary 50% increase ends September 13. Anthropic also published what it learned from cutting more than 80% of Claude Code's system prompt.
A write-up on the security blog Embrace The Red describes how to break Auto Mode in Claude Code running Claude Opus 5, and the bug behind it, reported on August 24, is still open. Auto Mode has been the default for new sessions on Pro, Max and Team plans since August 14, and Claude Opus 5 itself shipped on July 24, 2026.
Anthropic stripped more than 80% of Claude Code's system prompt for the newest Claude 5 generation models, then published what the trimming taught it about prompting them. The write-up covers system prompts, skills and Claude.MD files.
Anthropic said it'll permanently raise standard weekly limits by 25% from September 14 for Pro, Max, Team and seat-based Enterprise plans, turning extra headroom into the baseline instead of a promotion.
Testing Catalog points out the math cuts the other way in the short term: the permanent 25% is measured against the standard limit, so subscribers end up around 17% below what they have right now. The temporary 50% increase stays in place until September 13, per Anthropic's post on Bluesky.
Anthropic moved Enterprise-managed auth for MCP connectors to general availability, so an organization configures authorization once and staff get connector access on their first login to Claude. Anthropic's docs list the feature as generally available on Team and Enterprise plans. GA landed on August 24, 2026, roughly two months after the June 18 announcement.
Claude Opus 5 scores 30% on ARC-AGI-3 on its own, so nearly all of a perfect run comes from the agent scaffolding wrapped around it. A code walkthrough goes through the open source harness line by line to show how that scaffolding works.
The write-up credits a harness built with Strands and puts the final score at 99.95%. NVIDIA's developer blog and The New Stack, though, attribute the 100% result with Claude Opus 5 to NVIDIA's own AVO coding agent, which cleared every level of the benchmark.
WIRED says it reviewed code showing OpenAI is building a proactive, highly persistent version of Codex, so the agent would keep going without a new prompt each time. Per the report, the feature has Codex continue working until it's "put to sleep". OpenAI's Codex harness already carries work across turns; WIRED describes an agent that doesn't stop at the end of a task. The report ran August 27 in Model Behavior, WIRED's subscriber-only newsletter.
Developers using Kiro can now run GPT-5.6 there to plan, build, review and test software, OpenAI said, framing the addition around price-performance.
The cost side moved recently too: OpenAI cut API and credit pricing for GPT-5.6 Sol by over 20% as of August 21, 2026, according to its GPT-5.6 announcement page.
A Threads post says Tencent has put Hy4 Preview, a new-generation Mixture-of-Experts flagship, on Hugging Face under Apache License 2.0, listing 770B parameters with 49B active and a 1M context window. Those numbers come from that post alone; the model card sits under tencent/Hy4-preview.
A Threads post claims Zhipu AI officially announced GLM-5.3-Flash, a 320B-A18B multimodal model under the MIT License, and identifies it as Ox Alpha, the stealth model that ran on OpenRouter. Kingy AI, writing on August 21, says the "Flash" name stays unverified while the technical trail holds up: matching token budgets across test videos point to GLM-5.3's tokenizer and the GLM-5V line's video encoder. Zhipu's GLM-5.3 announcement on August 14 is confirmed, and the company said API access and open weights would arrive in stages after safety checks. The Threads post links to a Hugging Face page at zai-org/GLM-5.3-Flash.
Developers get generative video and conversational editing in Gemini Omni Flash, which is live on the Gemini APIs and Google AI Studio. Google describes it as bringing "a new suite of creative controls and generative video capabilities to developers".
The announcement circulating labels it Gemini Omni 1.1 Flash. Google's own documentation names the model Gemini Omni Flash and still marks it preview as of August 27.
A single post on Threads claims Grok 4.6 runs in normal chats on Grok web and mobile, after being limited to Grok Build. xAI hasn't published anything confirming that: its own rollout announcements cover Cursor, Grok Build, GitHub Copilot, Amazon Bedrock, Vertex AI and Perplexity. The same post says the Grok web UI got a slight update as well.
Grok Bot users can share their bots as templates, per a post from the SPACEXAI account on X. The feature had surfaced earlier as an unreleased Templates tab in Settings, spotted by TestingCatalog, whose report said the tab listed shared templates without spelling out what they were for.
