OpenAI, Anthropic, and Google all shipped new models this week, and they showed up in Copilot within a day or two. Nvidia signed the Hugging Face deal that leaked last week. VS Code can now send an agent after a stuck pull request.
Python also locked 3.15 for an October 1 release. CISA added bugs in PaperCut, Starlette, and LiteLLM. Here is the week.
Top Stories This Week
OpenAI Ships GPT-6 Astra -
On September 3, OpenAI released GPT-6 Astra. It is the company’s strongest model so far, and the first one OpenAI has put in its highest cybersecurity risk class. OpenAI says Astra scores 98% on FrontierMath Tier 4, 99.9% on ARC-AGI-3, and 100% on ExploitBench when safety filters are off. On Terminal-Bench 4.0 it reports 57.9%, against 55.8% for Claude Fable 5.1 and 37.3% for GPT-5.6 Sol. Two days earlier, a safety post said Astra can find unknown bugs and write exploits without a person walking it through each step. In testing it found two zero-days on its own. OpenAI says it is telling those vendors.
You can use it in ChatGPT Plus and above as the rollout reaches you. The API id is gpt-6-astra on OpenAI, Azure, and Bedrock. Price is $10 per million input tokens and $50 per million output. Fast mode is twice the price and up to twice as fast. On Enterprise, an admin has to turn it on. GitHub Copilot added Astra on September 4 for Pro+, Max, Business, and Enterprise. OpenAI already said Astra will not go to Cursor. If you still pick GPT models in Cursor, that access ends November 12, and Astra is not next. Codex can now keep notes across long chats. That is off by default in config.toml for now.
OpenAI also said Astra’s written reasoning is harder to follow than GPT-5.6 Sol’s. If you review chain-of-thought logs, expect less of that text.
Claude Fable 5.1 Makes Cached Agent Runs Cheaper -
On September 1, Anthropic released Claude Fable 5.1 and Claude Mythos 5.1. They are the same model with different safety rules. Fable 5.1 is for everyone. Mythos 5.1 is only for approved cyber and life-science users. Anthropic reports 55.8% on Terminal-Bench 4.0 for Fable 5.1 and 60.9% for Mythos 5.1. On a science terminal bench, Fable 5.1 hit 52.6%, up from 24.7% on Fable 5. Cache reads fell 75% to $0.25 per million tokens. Pricing for new input and output is still $10 and $50 per million, same as Fable 5. Anthropic says most jobs get about 25% cheaper, and long agent jobs can get about 45% cheaper.
The API id is claude-fable-5-1. It is on the Claude API, Bedrock, Google Cloud, and Microsoft Foundry. Claude Code defaults to High effort. GitHub Copilot added it the same day for Pro+, Max, Business, and Enterprise. Business and Enterprise admins have to turn the policy on. Anthropic keeps prompts on this model by default so its safety checks can run. Some companies can keep zero data retention until a new setup ships this fall. If your agents already cache a big codebase, this is the first Fable release where that pattern is cheap enough to compare with Opus 5.
Popular reads
View All
Payment System Design: Ledger, Idempotency, and Settlement
Jul 18, 2026
Cursor Skills: How to Create and Use Agent Skills
Jun 23, 2026
The Complete HTMX Guide: From Zero to Production
Dec 22, 2025
Flash Sale System Design: Architecture, Scale, and Oversell
May 16, 2026
Transactional Outbox Pattern: Never Lose an Event Again
Apr 07, 2026
Local LLM Speed: RTX 3060, Qwen2 & Llama Benchmark Results
Jan 15, 2026
Nvidia Agrees to Buy Hugging Face for $12.93 Billion -
On September 3, Jensen Huang said Nvidia will buy Hugging Face for $12.93 billion. Last week’s leak is now a signed deal. Nvidia’s filing is dated September 2: about $11.9 billion to Hugging Face shareholders, plus up to $1 billion in stock for staff who join Nvidia. Close is aimed at the first half of 2027, after regulators look at it. Huang said the Hub stays open. You can still pick any model, framework, cloud, and chip. You do not have to use NVIDIA hardware. Nvidia says the Hub has more than 18 million users, 3 million models, 500,000 datasets, and 1 million apps.
Nothing you download this week has to change. The deal is not closed yet.
VS Code 1.136 Can Send an Agent After a Pull Request -
On September 2, VS Code 1.136 shipped with Agent Merge in preview. Turn on chat.agentMerge.enabled, then in the Agents window run Enable Agent Merge for Active Session, or click the Agent Merge button. The agent tries to fix review comments, failed checks, and merge conflicts, then reruns CI until the PR looks ready. Copilot and Claude chats can now span more than one folder in a workspace (still experimental). Agent sessions also run in a separate process, so closing a window does not kill the chat.
Update from Help, then Check for Updates. Try it on a test PR first.
Google Ships Gemini 3.8 Flash at the Same Price as 3.7 -
On September 2, Google launched Gemini 3.8 Flash and Gemini 3.8 Flash Cyber. 3.8 Flash is the public model. Intro price matches 3.7 Flash: $0.75 per million input tokens and $3.75 per million output through December 31, 2026. On January 1, 2027 that becomes $1.50 and $7.50. Google says it is better at long coding jobs while staying at Flash speed. 3.8 Flash Cyber is only for approved security teams, through a program Google calls Fairwind.
The API id is gemini-3.8-flash. It is in the Gemini API, AI Studio, Android Studio, Gemini Enterprise, and Antigravity. GitHub Copilot added it on September 3 for Pro and above. If you already use 3.7 Flash, change the id in a test agent before you change production. For anything that will still be running next year, budget the January price.
More from Tech news
View All
Dev Weekly Aug 24-30, 2026: OpenAI Cuts Off Cursor, AWS Buys DuckLabs, Grok Bot on Pro
Aug 30, 2026
Dev Weekly Aug 17-23, 2026: Go 1.27 Generic Methods, Cursor Origin, GitHub's Outage, and Stripe's OpenRouter Deal
Aug 23, 2026
Dev Weekly Aug 10-16, 2026: Gemini 3.7 Flash, Grok 4.6, Patch Tuesday's 400 Fixes, and Anthropic's $2 Trillion IPO
Aug 16, 2026
Dev Weekly Aug 3-9, 2026: Meta Ships Muse Code, GPT-5.6 Luna Becomes the Free ChatGPT Default, Qwen 3.8-Max Goes GA, Next.js 16.3, and a Black Hat GitHub-Issue RCE in Claude Code, Gemini CLI, and Codex
Aug 09, 2026
Dev Weekly Jul 27-Aug 2, 2026: Node.js Ships Emergency Security Releases, Redis Patches AI-Found Zero-Days, DeepSeek V4 Flash Gets an Agentic Upgrade, Loco 1.0, Rails Active Storage RCE
Aug 02, 2026
Dev Weekly Jul 20-26, 2026: Anthropic Ships Claude Opus 5 at Half the Price, Google Floods the Market With Three Gemini Flash Models, Cursor Router, OpenAI Voice-Drives Codex, SharePoint Flaw Exploited
Jul 26, 2026
Python 3.15 Hits Its Last Release Candidate -
On September 1, Python 3.15.0rc2 shipped. It is the last planned candidate. About 144 fixes landed since rc1. Wheels built on rc2 should work on the final 3.15.0 release. Final is set for October 1. New in 3.15: lazy imports (PEP 810), frozendict and sentinel builtins, UTF-8 as the default encoding, unpacking in comprehensions, and a JIT that Python says is 8-9% faster on x86-64 Linux.
If you ship a C extension or wheels, build on rc2 now. Get it from python.org. Keep production on a stable release. Put rc2 in CI so October 1 is a version bump.
Copilot Can Approve Pull Requests -
On September 1, Copilot code review started saying whether a PR looks ready to approve. Admins can also let Copilot click Approve, and that count can meet a required-review rule. It is off by default. It is in public preview for Pro and above. You can set it at the company, org, or repo, including which file paths Copilot may approve. A new commit drops the approval, same as a human review.
Leave this off unless you really want an AI approval to count. The comment is still useful as a second opinion.
Developer Tools & Platforms
Google Makes Video Analysis Cheaper in Gemini Flash -
On September 1, Google turned on a new video mode on Gemini 3.7 Flash, 3.6 Flash, and 3.5 Flash-Lite. The model can search frames, audio, and captions instead of reading the whole clip at a fixed frame rate. Google says that can cut tokens by up to 88%, cut cost by up to 66%, and raise accuracy by up to 7% versus 1 frame per second. It works on uploads and YouTube links in the Gemini API. Price is the usual token rate. There is no extra fee.
Set "processing": "agentic" on the video part of the request. Use it for long talks and “when did X happen” questions.
Rust 1.98.1 Fixes a Compiler Bug -
On September 3, Rust 1.98.1 fixed a bug in 1.98.0 that could put a null pointer in a trait object. That can crash, or worse. Run rustup update stable and rebuild anything you shipped on 1.98.0 that uses trait objects.
Kubernetes Can Run the Kubelet as a Normal User -
On September 4, Kubernetes posted that rootless kubelet is beta in 1.37 (that release shipped August 26). The feature flag is on by default, but existing clusters that run as root do not change. To try it, run kind or minikube on rootless Docker or Podman. This is for a laptop cluster, not a production rewrite.
Security
CISA Says PaperCut Bugs Are Being Exploited -
On August 31, CISA added two PaperCut NG/MF bugs to its known-exploited list: CVE-2026-81578 and CVE-2026-82078. US federal agencies have until September 14 to patch. Huntress saw the attack in the wild. PaperCut says the fix is 24.1.10, 25.0.13, or 26.0.5. If a print server is on the internet, patch it and check the logs. Take the admin page off the public web.
CISA Also Flags Starlette, LiteLLM, Kestra, and Artifactory -
On September 2, CISA added seven more bugs. For app teams, the important ones are Starlette CVE-2026-48710 (fixed in 1.0.1), LiteLLM CVE-2026-59822 (use 1.84.0 or later), Kestra CVE-2026-49869, and JFrog Artifactory CVE-2026-82329. The rest are Sangoma Switchvox and two SonicWall SMA1000 bugs.
If you run FastAPI or anything on Starlette, pin starlette>=1.0.1 and rebuild. If you run a LiteLLM proxy, pin litellm>=1.84.0. Patch Artifactory and Kestra if they face the internet.
Funding & Industry Deals
Anthropic Signs a Reported $35B Cloud Deal With Lambda -
On August 31, the Wall Street Journal reported that Anthropic signed a $35 billion cloud deal with Lambda. Nvidia holds the lease on a Hut 8 data center in Nueces County, Texas. Bloomberg and Reuters put the site at about 350 megawatts. None of the companies commented. Anthropic also recently signed a reported $45 billion deal with Nscale. This does not change the Claude API this week.
AIR Raises $50M to Check What AI Agents Install -
On September 1, TechCrunch reported that AIR raised $50 million in two seed rounds. Sequoia led $10 million. Greenoaks led $40 million. The product watches the skills and plugins that agents install.
Lyte Raises $165M to Help Robots See -
On September 2, Lyte raised a $165 million Series C led by Maverick Silicon. The company is now worth $1.6 billion. Total raised is $272 million. It is based in Sunnyvale and builds sensors for robots.
Wafer Raises $40M to Tune Inference on the Fly -
On September 1, Wafer raised a $40 million Series A co-led by Marathon and Chemistry. AMD Ventures, Wing, Outset, Fifty Years, and Y Combinator also joined. Wafer says it will keep retuning models and hardware against live traffic, instead of tuning once at deploy time.
Empirik Launches With $21M to Predict Outages -
On September 1, Empirik launched as its own company with $21 million from Sequoia, Canapi, and Alumni Ventures. It tries to predict outages before they happen.
Layoffs: Uber and The Trade Desk
- Uber: On September 2, TechCrunch reported about 3,300 cuts, about 10% of staff. Bloomberg had it first. CEO Dara Khosrowshahi said Uber will have about 20% fewer managers, will merge engineering, science, and delivery teams, and will end almost all remote jobs. Reuters tied the cuts to robotaxis. Khosrowshahi did not blame AI.
- The Trade Desk: On September 3, a filing said the ad company will cut about 15% of staff and take $39 million to $51 million in cash charges this quarter. Business Insider put that at more than 500 people, based on 3,843 staff at the end of 2025. CEO Jeff Green said the company wants smaller teams.
Popular reads
View All
Payment System Design: Ledger, Idempotency, and Settlement
Jul 18, 2026
Cursor Skills: How to Create and Use Agent Skills
Jun 23, 2026
The Complete HTMX Guide: From Zero to Production
Dec 22, 2025
Flash Sale System Design: Architecture, Scale, and Oversell
May 16, 2026
Transactional Outbox Pattern: Never Lose an Event Again
Apr 07, 2026
Local LLM Speed: RTX 3060, Qwen2 & Llama Benchmark Results
Jan 15, 2026
The Numbers That Matter
- $12.93 Billion Nvidia’s signed price for Hugging Face
- $10 / $50 GPT-6 Astra and Claude Fable 5.1 standard API rates per million input and output tokens
- $0.25 Fable 5.1 cache read price per million tokens, down 75% from Fable 5
- $0.75 / $3.75 Gemini 3.8 Flash intro price through December 31, 2026
- 3,300 Uber roles cut, about 10% of staff
- 15% The Trade Desk workforce cut
- $35 Billion Reported Anthropic commitment to Lambda
- October 1, 2026 Python 3.15.0 final. rc2 is the last candidate
Quick Hits
- Anthropic / Lambda - August 31. Reported $35 billion cloud deal. Nvidia holds the Texas lease.
- PaperCut KEV - August 31. CVE-2026-81578 and CVE-2026-82078. Patch 24.1.10, 25.0.13, or 26.0.5. Due September 14.
- Claude Fable 5.1 - September 1.
claude-fable-5-1. Cache reads $0.25 per million tokens. - Copilot Fable 5.1 - September 1. Pro+ and above. Admin policy. Data retention on by default.
- Copilot PR approvals - September 1. Off by default. Can count toward required reviews if enabled.
- Python 3.15.0rc2 - September 1. Final RC. ABI frozen. Final due October 1.
- Gemini agentic video - September 1. Set
processingtoagenticon Flash video requests. - AIR $50M - September 1. Sequoia then Greenoaks. Agent plugin supply chain.
- Wafer $40M - September 1. Series A for continuous inference optimization.
- Empirik $21M - September 1. Sequoia-incubated outage prediction.
- GPT-6 Astra Path post - September 1. Critical cyber tier confirmed ahead of the September 3 launch.
- Gemini 3.8 Flash - September 2.
gemini-3.8-flash. Intro price through year-end. - VS Code 1.136 - September 2. Agent Merge preview.
chat.agentMerge.enabled. - Uber 10% - September 2. About 3,300 people. Fewer managers, almost no remote.
- Lyte Series C - September 2. $165 million at $1.6 billion.
- CISA seven bugs - September 2. Starlette, LiteLLM, Kestra, Artifactory, Sangoma, SonicWall.
- Nvidia / Hugging Face - September 3. Signed $12.93 billion. Close targeted H1 2027.
- GPT-6 Astra - September 3.
gpt-6-astra. $10 / $50. Not coming to Cursor. - Copilot Gemini 3.8 - September 3. Pro and above.
- Rust 1.98.1 - September 3.
rustup update stable. Fixes a compiler crash bug. - Trade Desk 15% - September 3. $39 million to $51 million in cash charges.
- Copilot Astra - September 4. Pro+, Max, Business, Enterprise.
- Kubernetes rootless kubelet - September 4. Beta in 1.37. Existing root clusters do not change.
The week was three new models in Copilot, a signed Nvidia deal for Hugging Face, and a VS Code feature that will try to finish your PR. Astra, Fable 5.1, and Gemini 3.8 Flash are all live if your plan includes them. Hugging Face still works the same until 2027 at the earliest. If you run Starlette, LiteLLM, or PaperCut, patch those. If you ship Python packages, build 3.15 wheels on rc2. Next week, watch whether Astra’s extra cyber access opens, and whether Python finds one last 3.15 bug. See you then.
