The Brief Brief
the week in AI, briefly. then briefly again.
Artificial Intelligence — briefly, then briefly again · tbb.ceo
listen tothe weekly #017 0:00 –:––
#017 ·29 AUG 2026 ·SATURDAY ·8 MIN READ ·10 STORIES + 20 EXTRAS

The Week Cursor Picked a Side

August 22-29: OpenAI cut Cursor off at the model layer, two Chinese labs open-sourced multimodal models within hours of each other, Claude Code shipped a flag that means no, and Amazon switched off the humans who had been pretending to be artificial intelligence since 2005.

01 / The Ten

The week, ranked

10

OpenAI cuts Cursor off at the model layer

OpenAI notified SpaceX on August 28 that it is winding down the contract supplying its models to Cursor, with direct access ending November 12, 2026 and future models including Astra excluded entirely. OpenAI said it could not be confident SpaceX would honour its terms of service, citing prior experience with Musk-owned companies. SpaceX closed its $60 billion Cursor acquisition on August 15.

Why it mattersThe model layer just stopped being a commodity and started being an alliance. Coding tools now inherit their owners' feuds, and the developers holding ten weeks of migration work are the collateral.
01 openai.com↗

Grok Bot lands inside Cursor two days before the cutoff

xAI made Grok Bot, launched in beta on August 11, available to SuperGrok, Cursor Pro and all Cursor Teams plans on August 26. The next day it put Grok 4.6 on Microsoft Foundry with a 500k context window, configurable reasoning levels and long-running agent support, plus Google's Enterprise Agent Platform.

Why it mattersCursor's replacement supply chain was wired up before OpenAI announced the divorce. The acquisition was never about GPUs alone -- it was about never needing a rival's models again.
02 x.ai↗

Z.ai open-sources GLM-5.3-Flash under MIT after a week undercover

Z.ai released GLM-5.3-Flash on August 26: a 320B-total, 18B-active hybrid-attention MoE, natively multimodal across text, image and video, with a 1M-token context window and MIT-licensed weights. It spent its first week running anonymously as 'Ox Alpha' on OpenCode and OpenRouter, served entirely on domestically produced Chinese chips. Z.ai claims it beats GLM-5.2 across its evaluations at one tenth the price.

Why it mattersA frontier-adjacent multimodal model, served on Chinese silicon and given away under the most permissive licence there is, is a pricing attack and a sovereignty statement in the same release. The anonymous week was a benchmark run with the flag taken off.
03 siliconangle.com↗

Alibaba ships Qwen3.8-Flash-Next the very same day

Alibaba's Qwen team released Qwen3.8-Flash-Next on August 26, a 125B-parameter natively multimodal MoE activating just 6B parameters per token and billed as a preview of the Qwen4 architecture. Alibaba positions it as competitive with Anthropic's Opus 4.6 and DeepSeek's V4-Flash at a lower price point. It follows the 2.4-trillion-parameter Qwen3.8-Max on August 3 and the laptop-class Qwen3.8-27B in mid-August.

Why it mattersTwo Chinese labs shipped competitive open multimodal models within hours of each other, unprompted and uncoordinated. Release cadence, not benchmark scores, is now the weapon.
04 marktechpost.com↗

Claude Code ships a flag that means no

Claude Code v2.1.248 landed August 28 with a --restricted flag (or CLAUDE_CODE_RESTRICTED=1) that strips out every built-in tool that runs commands or code, removes WebFetch unless explicitly named, fences file tools to the working directory, refuses bypassPermissions and ignores user, project and local settings files. The same release added cross-session messaging on Bedrock, Vertex and Foundry, a per-agent prompt-cache TTL, and /usage-credits for enterprise orgs.

Why it mattersOne flag that guarantees an agent cannot shell out is Anthropic conceding that agents now run unattended on machines their owners do not control. Sandboxing stopped being a config exercise and became a product promise.
05 github.com↗

Codex turns /import into a defection tool

OpenAI expanded Codex's /import command to migrate Cursor and Claude Code settings, MCP servers, plugins, sessions, commands and project-scoped memories, added import of Cursor-managed skills, and now syncs changes to imported Claude and Cursor conversations without creating duplicates. The same cycle stabilised multi-agent V2 -- spawn_agent, send_input, resume_agent, wait_agent and close_agent, each sub-agent with its own model, reasoning level, concurrency cap and sandbox profile -- plus opt-in support for the MCP 2026-07-28 protocol.

Why it mattersAccumulated agent configuration was the last real switching cost in coding tools. OpenAI spent a release cycle deleting it, aimed squarely at the rival it is cutting off at the model layer in November.
06 gradually.ai↗

Amazon shuts down Mechanical Turk after 21 years

AWS confirmed this week that Mechanical Turk closes on September 30, 2026, taking SageMaker Ground Truth with it, after stopping new customer signups in July. AWS gave no specific reason beyond an internal assessment. Research published in 2023 found as many as 46% of MTurk workers were already using AI models to complete their assigned tasks.

Why it mattersThe marketplace that supplied the labelled data underpinning modern AI is being switched off by the thing it trained. Insurance and travel firms still running data work through it have four weeks to find somewhere else.
07 cnbc.com↗

Claude's memory goes cross-surface, with an off switch

Anthropic announced on August 25 that Claude now carries shared memory across chat and Claude Cowork, updating as you talk, with controls to view, edit, delete, pause or reset anything saved. Sensitive categories such as health and personal beliefs are excluded from memory by default and must be opted into in Memory settings.

Why it mattersPersistent memory is the feature that makes an assistant sticky and the feature regulators will read most closely. Shipping health data default-off is what compliance-shaped product design looks like.
08 claude.com↗

Gemini Omni 1.1 Flash stretches AI video to 40 seconds

Google's August 27 update lets scene extension analyse up to ten seconds of existing footage instead of just the last second, chaining 10-second increments to 40 seconds total. Developers can supply up to three seconds of external video as a style reference, and a new 360p draft mode runs up to 60% faster at a third of 720p's cost. Per-second pricing is $0.03 at 360p, $0.10 at 720p, $0.15 at 1080p and $0.30 at 4K.

Why it mattersContinuity across cuts, not raw clip length, is what has kept AI video out of real production. Draft-then-upscale pricing is the first release that assumes creators iterate rather than roll the dice once.
09 the-decoder.com↗

Engineers crossed the daily-agent threshold

Temporal's 2026 State of Development Report found 80.8% of respondents now use AI agents daily or more often, up from 47.3% a year earlier -- a 70.8% leap. The survey ran April 29 to May 25, 2026 across 550-plus software engineers, architects and engineering leaders, two thirds US-based and one third UK/EMEA.

Why it mattersDaily use is the line where tooling choices stop being preferences and become infrastructure. It explains why every fight this week -- Cursor, Codex imports, restricted mode -- was about who controls the harness rather than who has the best model.
10 temporal.io↗
02 / Also

Worth knowing

20
Fast inference is a security problem
An OpenAI researcher posting as 'roon' warned on August 27 that a misaligned frontier-class model running 50x faster could infiltrate systems far quicker than human responders can react, arguing for autonomous detection and shutdown rather than monitoring.
x.com ↗
Fireworks AI raises $1.5B
The enterprise inference startup closed the largest AI round of the period at $1.5 billion.
news.crunchbase.com ↗
River AI launches with $1.1B
xAI co-founder Igor Babuschkin's full-stack AI company announced $1.1 billion in funding on August 21, days before his former employer's models started replacing OpenAI's inside Cursor.
techstartups.com ↗
Gatik tops the week's funding table at $200M
The autonomous trucking company led the August 24-28 round-up, followed by Stability AI at $76 million and Runable's $21 million Series A.
analyticsinsight.net ↗
OpenAI closes the enterprise gap
Ramp data on 70,000-plus US businesses shows Anthropic still leading at 43.5% of companies against OpenAI's 39.7%, but OpenAI's quarter-over-quarter growth rate (82) now outpaces Anthropic's (76).
techcrunch.com ↗
A2A joins MCP under one roof
Google's Agent2Agent protocol became a hosted project of the Linux Foundation's Agentic AI Foundation, which has grown from under 40 members to more than 250 since December 2025, putting the agent-to-agent and agent-to-tool layers under the same neutral governance.
forbes.com ↗
Copilot cloud agents move into Teams
GitHub added Copilot cloud agent sessions to Microsoft Teams, letting anyone in a channel, thread or DM mention @GitHub to start and steer a session, with repo write access required to trigger actual changes.
github.blog ↗
Copilot's agent harness goes stable
The GitHub Copilot Agent is now released and stable for .NET and Python in Microsoft Agent Framework, usable as the execution engine for custom agents while keeping the framework's observability and human-in-the-loop approvals.
devblogs.microsoft.com ↗
DALL-E retires on August 30
OpenAI is shutting down the official DALL-E GPT inside ChatGPT tomorrow, folding image generation entirely into the main model.
help.openai.com ↗
ChatGPT loads long chats in pieces
An August 21 update loads long web conversations in smaller sections rather than retrieving the whole thread, and interactive content now appears progressively as it is generated instead of waiting for completion.
help.openai.com ↗
Claude Code reaches v2.1.250
Releases this month added automatic feedback-report drafting when issues occur, per-turn completion timing, clearer warnings on overly broad permission rules, and a fix for a startup crash hitting Linux users on recent glibc.
releasebot.io ↗
NotebookLM is now Gemini Notebook
Google's rebrand gives every notebook a secure cloud computer for native code execution and data analysis, pushing the source-grounded research tool toward a working environment for decisions rather than a summariser.
blog.mean.ceo ↗
Terminal-Bench 2.1 leaderboard
Claude Code and Codex lead at 89.5% and 89.1%, with Grok 4.6 close behind at 88.4% -- a spread narrow enough that switching costs matter more than capability.
morphllm.com ↗
Gemini 3.7 Flash halved its own price
Google's coding-and-agents workhorse launched at $0.75 and $3.75 per million input and output tokens, scoring 43.6% on FrontierCode 1.1 Main against 34.4% for 3.6 Flash, and 65.3% on DeepSWE v1.1 against 49%.
venturebeat.com ↗
Qwen3.8-27B runs on a laptop
Alibaba's mid-August open-weight release trims to 27 billion parameters explicitly targeting consumer hardware, answering Meta's push into small local models.
cnbc.com ↗
Releases have outrun testing
Fourteen models shipped from eight providers in August alone, eleven of them inside a single 20-day stretch -- faster than anyone can meaningfully evaluate them.
llm-stats.com ↗
Google Cloud names agent security the gate
An August 24 post tied to its State of AI infrastructure report frames agent security as the top blocker keeping pilots out of production.
agentic.ai ↗
Agent startups pulled in $1.8B
Roughly $1.8 billion went into about a dozen AI agent deals across July and August 2026.
agentic.ai ↗
Codex logs into Bedrock
OpenAI added Amazon Bedrock login, custom endpoint and authentication support to Codex, with GPT-5.6 Sol as the default Bedrock model, plus audio inputs and tool outputs in common local formats.
releasebot.io ↗
RoboColiseum opens in Shanghai
A standardised simulation platform for embodied AI launched August 24, giving robotics labs a common arena to benchmark policies against each other.
agentic.ai ↗
Be subscriber #013 one email a week · no spam · unsubscribe anytime
Back issues

The Archive

#017The Week Cursor Picked a Side#016The Week Nvidia Bought the Stack#015Exit Velocity#014The quarter Anthropic won#013Proofs and Exploits#012The agents are talking to each other#011The Week the Agents Went Rogue#010Escape velocity#009Proofs, Suits, and Export Controls#008The week the gate went up#007The week the ceiling came into view#006The week everyone built a moat#005Cold Feet, Hot Models#004The Week the Export Hammer Dropped#003IPOs, Agents, and Two Apologies#002New Brains, New Tiers, New Tickers#001AI's trillion-dollar week
Every Friday · 10:00

Get the brief

One email a week. The ten things in AI that mattered, and why. Choose your channel.