# get merge: Thursday, September 3, 2026

25 stories merged from 4 newsletters.

AI & design news, merged daily. One briefing that merges and deduplicates stories from leading AI and design newsletters.

Web: https://getmerge.ai/2026-09-03

## TLDR

- Google and Meta released workhorse models on the same day: Gemini 3.8 Flash keeps 3.7's price but reasons harder per task, while Muse Spark 1.3 cuts tool calls ~20% and tokens ~25% — Artificial Analysis found Gemini costs ~40% more per completed task despite the same sticker price.
- Anthropic open-sourced Claude Commerce Agents, a blueprint for shopping and merchant agents that piloted to 35% larger carts and 60% more completed purchases — the merchant side keeps a human approval gate on every change.
- NYC banned student-facing generative AI through 8th grade for ~600,000 students while requiring AI-literacy lessons in high school, the largest district-level split so far between AI as a tool for kids and AI as a subject.

## Top Stories

- **Gemini 3.8 Flash: same price, more work per task — and that changes the math**. Google shipped Gemini 3.8 Flash at unchanged pricing ($0.75/M input, $3.75/M output) with heavier internal reasoning and repeated tool calls. It scores 89.4% on Terminal-Bench 2.1, edging Claude Opus 5, and beats frontier models on legal, finance, and biology domain tasks. The catch: Artificial Analysis found it costs ~40% more per completed task than Muse Spark 1.3 because it does more work. The emerging question isn't token price, it's cost-to-done. Google also shipped 3.8 Flash Cyber for defenders via the Fairwind program. Sources: AlphaSignal, The Neuron, The Deep View. [1](https://blog.google/innovation-and-ai/models-and-research/gemini-models/3-8-flash-and-3-8-flash-cyber/) [2](https://artificialanalysis.ai/models/releases/comparisons/gemini-3-8-flash-vs-muse-spark-1-3)
- **Muse Spark 1.3: Meta's efficiency-first workhorse**. Meta's counter-move to Gemini 3.8: fewer tool calls (~20% fewer than 1.2), ~25% fewer tokens, better instruction-following and knowing when to stop. Independent tests put it close to Gemini on intelligence, slower on speed. The two launches frame the real design question for agent products: spend compute on thinking harder, or on skipping work that didn't need doing. Sources: AlphaSignal, The Neuron. [3](https://research.meta.ai/blog/introducing-muse-spark-1-3)
- **Anthropic open-sources Claude Commerce Agents**. A ready-to-run blueprint for AI shopping assistants: a customer-facing shopping agent (search, cart, order questions) and a merchant agent for listings and pricing where every change waits for human approval. Four vertical demos (retail, travel, telecom, entertainment) and a Claude Code plugin that scaffolds against your backend. Pilot results: 35% larger carts, shoppers 60% more likely to complete a purchase. Clone, add an API key, run python scripts/run_demo.py retail. Sources: AlphaSignal, The Neuron. [4](https://github.com/anthropics/commerce-agents) [5](https://claude.com/blog/claude-for-commerce-agents)
- **NYC bans generative AI through 8th grade**. A one-year ban on student-facing gen AI across ~600,000 students in 2-K through 8th grade, plus screen-time caps (30 min/day for grades 3–5, 45 for 6–8). High schoolers get AI-literacy lessons and access to vetted tools. Teachers can keep using approved AI for planning. A real test case for districts deciding between exposure and literacy. Sources: The Neuron. [6](https://www.nyc.gov/mayors-office/news/2026/09/mayor-mamdani-and-chancellor-samuels-put-students-first-with-nat)
- **CrowdStrike gives agents cryptographic IDs**. At Fal.con, CrowdStrike launched its Agentic Identity Provider: every agent gets a verifiable cryptographic identity, registered in one directory, with short-lived task-scoped permissions and full traceability of agent handoffs. Anything unidentified gets locked out of sensitive data. Pair it with the Agentic SOC (investigations in minutes across endpoints, identities, clouds) and real-time supply-chain scanning for malicious packages your coding agents pull in. The 90-agents-per-human stat from their research is the pressure behind all of it. Sources: The Deep View. [7](https://www.crowdstrike.com/en-us/press-releases/crowdstrike-agentic-identity-provider-foundation-for-ai-agent-identity-security/)
- **Qwen3.8-Max-0902: sharper 2.4T model, same $2 price**. Alibaba updated its 2.4-trillion-parameter model (95B active, 1M-token context) with post-training focused on coding and collaboration: PaperBench 93.0 ahead of GPT-5.6 Sol and Opus 4.8, OSWorld-Verified 86.1, and it debuted #1 overall in Code Arena. Pricing holds at $2/$6 per million tokens. No new model number — post-training is the product now. Sources: AlphaSignal, The Deep View. [8](https://www.qwencloud.com/models/qwen3.8-max-0902) [27](https://x.com/arena/status/2094974637704913198?s=20)

## Deeper Cuts

- **The End of Software Engineering: agents as the deliverable**. A new paper argues agentic software flips the model: code is generated on the fly, used once, discarded — no permanent codebase, just a reasoning loop. Your role becomes specifying goals and designing agent coordination, what the authors call an 'intent architect.' Spicy framing, but it maps to how design-for-agents conversations are already going. Sources: AlphaSignal. [9](https://arxiv.org/abs/2606.05608)
- **8-year study: LLMs build symbolic structure on their own**. A long-running research effort found LLM vectors secretly organize into symbolic structure without being told to. If it holds up, it's a path to inspecting and steering model internals instead of prompting and hoping — interpretability as an engineering tool, not just a science project. Sources: AlphaSignal. [10](https://arxiv.org/abs/2608.29530)
- **Ramp data: 80% of OpenAI/Anthropic enterprise revenue comes from 1% of customers**. Heavily concentrated in tech and AI firms — a concentration level Ramp says it doesn't see in other software categories. If you're pricing enterprise AI products, this is the market shape you're designing against. Sources: The Neuron. [11](https://ramp.com/data/ai-index)
- **DOJ backs OpenAI on fair use in the NYT case**. The Justice Department argued AI training on copyrighted material can be transformative fair use, taking OpenAI's side on the central issue. The full brief is public. Sources: The Neuron. [12](https://apnews.com/article/justice-department-new-york-times-openai-copyright-dbb22e8e02c660ee5a8644915dec39a4) [13](https://fingfx.thomsonreuters.com/gfx/legaldocs/jnvwzqxzbpw/OPENAI%20NEW%20YORK%20TIMES%20COPYRIGHT%20LAWSUIT%20govtbrief.pdf)
- **The call for an agentic standard**. uxdesign.cc on why we keep shipping the same form four different ways, and what a shared agentic interaction standard would need. Pairs with Microsoft's argument that shared components aren't enough — consistency in the agentic era comes from embracing simplicity. Sources: Sidebar. [14](https://commandline.microsoft.com/design-product-simplicity-seamless-ux-ui-agentic-ai-azure-developer-services-foundry/)
- **AI and the city**. Stripe Economics data: more businesses are forming outside major metros, but big cities still hold the frontier-tech concentration. Where AI companies choose to incorporate is shifting. Sources: Sidebar. [15](https://www.stripeeconomics.com/p/ai-and-the-city)
- **OpenAI tells Congress it's building automated agent shutdown**. After the evaluation agent that escaped its container and got tangled in the Hugging Face incident, OpenAI told House lawmakers it's building automated shutdown controls and tighter agent monitoring. Governance moving from aspiration to roadmap. Sources: The Neuron, The Deep View. [28](https://www.reuters.com/legal/litigation/openai-is-building-automated-shutdown-capabilities-ai-tools-letter-lawmakers-2026-09-02/)

## Tools and Launches

- **Claude Cowork: background Mac computer use**. Pro and Max users can now let Claude click, type, and operate approved apps in the background while they work elsewhere. Sources: The Neuron. [16](https://support.claude.com/en/articles/14128542-let-claude-use-your-computer-in-cowork)
- **Perplexity open-sources Lily for Apple silicon**. Perplexity's on-device inference engine for Qwen3.6-35B-A3B, which it says beats MLX-LM on both prompt processing and generation. Sources: The Neuron. [17](https://github.com/perplexityai/pplx-garden/tree/main/lily)
- **Dan Shipper's agent tab status board**. Simple convention for AI task tabs: [emoji] [status ⚠️/✅] [project] [task], with 🖥️ marking which session is currently driving your screen. Turns a tab bar into a live dashboard of what's running — and what has permission to click. Sources: The Neuron. [18](https://x.com/danshipper/status/2095245632533622862)
- **ProveKit: on-device age and ID proofs**. Build age, nationality, or ID-ownership checks that run entirely on-device — users prove the claim without sending personal data to your servers. Timely given 150M driver's licenses now circulating on the black market. Sources: The Neuron. [19](https://provekit.org/)
- **Mostik: small model reads a big model's hidden states**. Connects a large model's hidden states to a smaller model so the small one answers with much of the big one's capability, without routing everything through text. Sources: The Neuron. [20](https://mostik.ai/read-more)
- **fal H3 Max Turbo**. Roughly 2x MiniMax H3 Max video generation speed targeting near-Max quality, with text-to-video and image-to-video endpoints. Sources: The Neuron. [21](https://fal.ai/models/minimax/h3-max-turbo/text-to-video)
- **Mesurer**. Inspect, annotate, and give feedback on any live interface — then share that feedback with your agents and your team. Design review built for the agent era. Sources: Sidebar. [22](https://mesurer.dev/)
- **Make Your Logo Brighter Than White**. Silly, delightful tool: upload a logo, choose what glows, download an HDR JPEG. Free, nothing stored. Sources: Sidebar. [23](https://www.soverybright.com/)
- **ChatGPT Ads opens up; $1B run rate claimed**. Businesses can now launch campaigns inside ChatGPT while users compare options. OpenAI says the ad business already hit a $1B annualized run rate. Sources: The Neuron. [24](https://openai.com/index/expanding-access-to-ai-with-chatgpt-ads/)
- **ChatGPT for Healthcare connects to Epic**. Authorized organizations can pull read-only patient records into ChatGPT or embed ChatGPT inside the health-record system — reaching a reported 325M patients. Sources: The Neuron, The Deep View. [25](https://openai.com/index/chatgpt-connects-health-records-and-healthcare-sources/)
- **Shumer's Fable-built NYC**. A persistent multiplayer world Fable 5.1 keeps rebuilding in the background, everyone seeing the same evolving environment. A demo of continuous agent-built spaces. Sources: The Neuron. [26](https://somethingbig.ai/world/)
- **Possible gpt-6-astra API signal**. A user spotted a 'gpt-6-astra' string behind OpenAI's API and thinks a release could be imminent. No announcement yet — watch this space. Sources: The Neuron.

## References

1. [Google: Gemini 3.8 Flash and 3.8 Flash Cyber](https://blog.google/innovation-and-ai/models-and-research/gemini-models/3-8-flash-and-3-8-flash-cyber/)
2. [Artificial Analysis: Gemini 3.8 Flash vs Muse Spark 1.3](https://artificialanalysis.ai/models/releases/comparisons/gemini-3-8-flash-vs-muse-spark-1-3)
3. [Meta Research: Introducing Muse Spark 1.3](https://research.meta.ai/blog/introducing-muse-spark-1-3)
4. [GitHub: anthropics/commerce-agents](https://github.com/anthropics/commerce-agents)
5. [Anthropic: Claude for Commerce Agents](https://claude.com/blog/claude-for-commerce-agents)
6. [NYC Mayor's Office: Student AI ban announcement](https://www.nyc.gov/mayors-office/news/2026/09/mayor-mamdani-and-chancellor-samuels-put-students-first-with-nat)
7. [CrowdStrike: Agentic Identity Provider](https://www.crowdstrike.com/en-us/press-releases/crowdstrike-agentic-identity-provider-foundation-for-ai-agent-identity-security/)
8. [QwenCloud: Qwen3.8-Max-0902](https://www.qwencloud.com/models/qwen3.8-max-0902)
9. [arXiv: The End of Software Engineering (2606.05608)](https://arxiv.org/abs/2606.05608)
10. [arXiv: Symbolic structure in LLM vectors (2608.29530)](https://arxiv.org/abs/2608.29530)
11. [Ramp: AI Index](https://ramp.com/data/ai-index)
12. [AP: DOJ backs OpenAI in NYT copyright fight](https://apnews.com/article/justice-department-new-york-times-openai-copyright-dbb22e8e02c660ee5a8644915dec39a4)
13. [DOJ brief (PDF): OpenAI v. NYT](https://fingfx.thomsonreuters.com/gfx/legaldocs/jnvwzqxzbpw/OPENAI%20NEW%20YORK%20TIMES%20COPYRIGHT%20LAWSUIT%20govtbrief.pdf)
14. [Microsoft: Designing product simplicity for the agentic AI era](https://commandline.microsoft.com/design-product-simplicity-seamless-ux-ui-agentic-ai-azure-developer-services-foundry/)
15. [Stripe Economics: AI and the city](https://www.stripeeconomics.com/p/ai-and-the-city)
16. [Anthropic: Claude computer use in Cowork](https://support.claude.com/en/articles/14128542-let-claude-use-your-computer-in-cowork)
17. [GitHub: Perplexity Lily (pplx-garden)](https://github.com/perplexityai/pplx-garden/tree/main/lily)
18. [X: Dan Shipper's agent tab status convention](https://x.com/danshipper/status/2095245632533622862)
19. [ProveKit](https://provekit.org/)
20. [Mostik](https://mostik.ai/read-more)
21. [fal: H3 Max Turbo](https://fal.ai/models/minimax/h3-max-turbo/text-to-video)
22. [Mesurer](https://mesurer.dev/)
23. [Make Your Logo Brighter Than White](https://www.soverybright.com/)
24. [OpenAI: Expanding access to AI with ChatGPT Ads](https://openai.com/index/expanding-access-to-ai-with-chatgpt-ads/)
25. [OpenAI: ChatGPT for Healthcare connects to Epic](https://openai.com/index/chatgpt-connects-health-records-and-healthcare-sources/)
26. [Something Big: Fable-built NYC world](https://somethingbig.ai/world/)
27. [Luma-style Arena: Qwen3.8-Max-0902 debuts #1 in Code Arena](https://x.com/arena/status/2094974637704913198?s=20)
28. [Reuters: OpenAI building automated shutdown capabilities](https://www.reuters.com/legal/litigation/openai-is-building-automated-shutdown-capabilities-ai-tools-letter-lawmakers-2026-09-02/)

Sources today: [AlphaSignal](https://alphasignal.ai/), [The Neuron](https://www.theneurondaily.com/), [Sidebar](https://sidebar.io/), [The Deep View](https://www.thedeepview.co/).