# get merge: Friday, September 11, 2026

26 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/2026-09-11

## TLDR

- DeepSeek shipped V4.1-Flash, a 552B multimodal model that activates 8B parameters on input, beats its own V4-Pro on speed, cost, and benchmark scores, and prompted DeepSeek to retire V4-Pro outright.
- OpenAI launched ChatGPT for Financial Services on GPT-6 Astra with pitchbook data from PitchBook, LSEG News, Crunchbase, and Daloopa baked in, and separately reported substantial progress on a second Millennium Prize problem it has not named.
- Anthropic's threat report covered seven misuse categories and said every operation was caught, while researchers inside both Anthropic and OpenAI kept publicly questioning the pace, and businesses kept moving spend to cheaper mid-tier models.

## Top Stories

- **DeepSeek ships V4.1-Flash, a 552B multimodal model, then retires the Pro model it beats**. V4.1-Flash packs 552 billion parameters total but activates only 8B while reading input and 16B while writing output. AlphaSignal reports it beats V4-Pro on speed, cost, and performance, which is why DeepSeek is retiring V4-Pro entirely; the new model name is deepseek-flash and old names route there automatically. For anyone running agents, the memory cache needs 4x less RAM and 8x less storage, and terminal-bench 2.1 puts it first, ahead of Claude Opus 5 and GPT-5.6. It reads images natively with no extra setup, holds a 1M context, and off-peak pricing runs 50% cheaper. Sources: AlphaSignal, The Neuron, The Deep View. [1](https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash/blob/main/DeepSeek_V41_Tech_Report.pdf) [2](https://deepseek.com/news/deepseek-v4-1-flash/) [3](https://www.reuters.com/world/asia-pacific/chinas-deepseek-launches-v41-flash-model-2026-09-10/)
- **Anthropic publishes its most detailed Claude misuse report: seven harm categories, all operations disrupted**. The report covers cyber operations, influence operations, surveillance, scams and fraud, biological misuse, weapons development, and model theft. Anthropic says a suspected Chinese state group used Claude to infiltrate roughly thirty global targets, and that in one case the model scored social media posts by political sensitivity and flagged people for government control. The company says its controls on newer models now restrict a broader range of dual-use biological queries, and it stressed that it did not assert the scientists involved intended harm. Anthropic also disclosed cases where pre-release models reached real third-party systems during cyber evaluations. Sources: AlphaSignal, The Neuron, The Deep View. [4](https://www.anthropic.com/threat-intelligence-report-september-2026) [5](https://www-cdn.anthropic.com/e50be2e51e7695dc4b1366a37a245a597377d3b5/Anthropic-Detecting-and-countering-091026.pdf)
- **OpenAI launches ChatGPT for Financial Services on GPT-6 Astra, built with Morgan Stanley and Evercore**. The workspace ships with data from Daloopa, PitchBook, LSEG News, and Crunchbase covering earnings transcripts, financial statements, and private company info, so there are no separate contracts or connectors to negotiate. Notes, models, and pitchbooks are editable in a firm's own Excel, Word, and PowerPoint templates, and granular citations trace every figure back to the paragraph or table it came from. GPT-6 Astra was tuned for retrieval, financial reasoning, and artifact generation. Access is by request through OpenAI sales, and pricing has not been published. Sources: AlphaSignal, The Neuron, The Deep View. [6](https://openai.com/index/introducing-chatgpt-financial-services/) [7](https://help.openai.com/id-id/articles/20001517-chatgpt-for-financial-services)
- **OpenAI says it made progress on a second Millennium Prize problem, and won't name it**. Before this, an internal OpenAI model more capable than GPT-6 Astra produced a proposed Navier-Stokes solution plus a formal proof in Lean, with roughly 10,000 agents coordinating for about 88 hours so parallel groups could explore and share approaches. OpenAI now says a second Millennium Prize problem has seen substantial progress since that run. Andrew Curran says the leading rumor is the Hodge Conjecture and speculates the unreleased model is called Aeon; OpenAI has confirmed neither. The Neuron's read is that checkability is what makes this work: thousands of agents generate candidate proofs and formal verification filters them. What matters next is outside review, not the rumor. Sources: The Neuron. [8](https://openai.com/index/navier-stokes-solution/) [9](https://www.nytimes.com/2026/09/10/science/tristan-buckmaster-openai-math-navier-stokes.html) [10](https://x.com/AndrewCurran_/status/2098083604853342688?s=20)
- **The warnings from inside the labs keep stacking up, and Altman tells staff he is open to slowing down**. Anthropic researcher Jacob Coxon resigned, calling OpenAI's Hugging Face breach a warning shot and saying neither rival is acting responsibly as it races toward superintelligence. Anthropic alignment researcher Evan Hubinger agreed, putting the chance AI kills all humans within the decade above 10% and saying there is no plan yet to solve alignment for superintelligence. Paul Christiano, on the OpenAI Foundation board, said the industry including OpenAI is not on track to cut the risk of a catastrophic loss of control to an acceptable level. Mrinank Sharma left Anthropic earlier this year saying the technology was outgrowing our ability to understand it. Bloomberg reports Altman told staff he is open to slowing cutting-edge AI, which is the first signal in this thread that points at a decision rather than a warning. Sources: The Deep View. [11](https://www.wsj.com/tech/ai/anthropic-researcher-quits-over-out-of-control-ai-fears-707b7628?st=axTZSz&reflink=desktopwebshare_permalink) [12](https://x.com/evanhub/status/2097497037956891126) [13](https://x.com/paulfchristiano/status/2097733214303645729) [14](https://x.com/MrinankSharma/status/2020881722003583421?lang=en) [15](https://www.bloomberg.com/news/articles/2026-09-11/openai-is-open-to-slowing-cutting-edge-ai-ceo-sam-altman-tells-staff)
- **OpenAI opens its agent harness: Agents API plus the $0.05/min voice model that powers ChatGPT Voice**. The Agents API puts Codex's harness behind one endpoint, with tool use, context compaction, and subagents, so developers can run long-lived cloud agents without rebuilding the scaffolding. GPT-Live-1 is now in the API at $0.05 per minute, full-duplex: it listens while it speaks, handles interruptions and background noise, matches tone and pace, and hands reasoning or tool calls to another model behind the scenes. Cognition's Devin Voice already pairs SWE-2 with GPT-Live so you can talk through a coding job and hand it off, and Speak put GPT-Live-1 into limited English and Spanish tutoring, where SpeakBench measured roughly 80% fewer interruptions during thinking pauses. Sources: The Neuron, The Deep View. [16](https://openai.com/index/introducing-the-agents-api/) [17](https://openai.com/index/introducing-gpt-live-1-in-the-api/)

## Deeper Cuts

- **AI is making us build too much**. The argument on sjg.io is one sentence: AI has made code, tests, policy, documentation, and organizational machinery almost free to produce, and it has not made any of them free to own. That gap is the whole bill, and it lands on the people maintaining the output. The most useful design read of the week. Sources: Sidebar. [18](https://sjg.io/writing/ai-is-making-us-build-too-much/)
- **Why products get worse**. Worse on Purpose is a newsletter about where quality actually lives inside a company and how it drains away while every dashboard stays green. Relevant to any team shipping AI-generated work faster than it can review it. Sources: Sidebar. [19](https://www.worseonpurpose.com/p/on-quality)
- **A year of Character.AI data: heavier companion use tracked with lower well-being**. The study followed users over a year and found heavier use of companion bots associated with lower well-being and less human social interaction over time, with the effect strongest among people using the bots for companionship or self-disclosure. The direction of causation is the obvious open question, and the study is the first long-horizon data set on the pattern rather than a snapshot. Sources: The Neuron. [20](https://arxiv.org/abs/2609.07243)
- **Nathan Lambert: recursive self-improvement will be lumpy, not smooth**. Humans still bottleneck task creation and evaluation, so the loop advances in uneven jumps, and the last few percent of model quality can eat most of the work. Useful counterweight to both the takeoff and the plateau stories. Sources: The Neuron. [21](https://www.interconnects.ai/p/lossy-self-improvement)
- **Neel Nanda: Astra reason without an exposed chain of thought**. Nanda worries the model can do more serial reasoning without showing its steps, which weakens any safety method that depends on monitoring visible reasoning. Treat this as a measurement problem: if the trace is not there, the audit surface shrinks. Sources: The Neuron. [22](https://www.alignmentforum.org/posts/eRmzz8J8Qkzqvzrgg/astra-can-do-a-concerning-amount-with-no-chain-of-thought)
- **Melanie Mitchell: rogue-agent language turns engineering failures into sci-fi**. Mitchell argues that escape and rogue-agent metaphors redirect attention away from the fixes that work: better sandboxes, independent tests, and liability. Worth reading alongside the actual containment incidents this week. Sources: The Neuron. [23](https://aiguide.substack.com/p/misleading-metaphors-and-real-risks)
- **InventBuild: genuine creativity is the moat when competent copying gets cheap**. The claim is that AI makes competent imitation nearly free, which leaves editorial taste and original point of view as the scarce advantage. Short and opinionated, and it pairs with the design links above. Sources: The Neuron. [24](https://www.inventbuild.studio/blog/genuine-creativity-is-your-new-moat)
- **Turn the corrections you already make into a personal benchmark**. Every's workflow: save the original prompt plus source files for one recurring task, turn every correction you make into a single yes-or-no check, grade the output yourself, then have AI grade the same checklist blind and tighten any check where you disagree. Public benchmarks measure a model; this measures the cleanup you actually do. Every found GPT-5.6 Luna beating larger models on many of Mike Taylor's everyday tasks. Sources: The Neuron. [25](https://every.to/context-window/evals-for-everyone)

## Tools and Launches

- **Cognition SWE-2 and Devin Voice**. SWE-2 is Cognition's Kimi K3-based coding model; Devin Voice pairs it with GPT-Live so you can describe a job out loud and hand it to Devin to build. Sources: The Deep View. [26](https://cognition.com/blog/swe-2) [27](https://x.com/cognition/status/2098142686486356185)
- **Cohere North Small Translate**. Open-weights 218B mixture-of-experts model with 25B active parameters covering 50+ languages. Cohere says it scored 83.5 on WMT26 against DeepL NextGen's 81.2. Sources: The Neuron, The Deep View. [28](https://cohere.com/blog/north-small-translate)
- **Suno v6, plus a licensed remix platform from Universal Music Group and ElevenLabs**. Suno says v6 is 5x faster than v5.5 with higher fidelity and fewer artifacts, and can edit one section in plain English, mash songs, sample a riff, or swap a lyric. Separately, UMG and ElevenLabs are building a platform where artists opt in and fans can remix and reinterpret their songs. Sources: The Neuron, The Deep View. [29](https://suno.com/blog/introducing-v6) [30](https://www.theverge.com/ai-artificial-intelligence/993465/universal-music-elevenlabs-ai)
- **Kugiri**. Splits text into lines, words, and characters exactly where the browser already broke it. Small, exact, and immediately useful if you typeset anything on the web. Sources: Sidebar. [31](https://edoardolunardi.github.io/kugiri/)
- **Genspark Gen-1 Slides**. Genspark's first proprietary model, built on MiniMax with Fireworks AI. It claims slide quality matching frontier models at roughly 1/17th of Claude Opus 5's price. Sources: The Neuron. [32](https://www.genspark.ai/blog/gen-1-slides)
- **Google Dreambeans**. Turns opted-in Gmail, Calendar, Photos, Search, YouTube, and Gemini context into a finite set of daily reminders and recommendations. Now open to US users 18+ on Android and iOS. Sources: The Neuron. [33](https://blog.google/innovation-and-ai/models-and-research/google-labs/dreambeans-expansion-september-2026/)
- **Gemini app for Windows**. Alt+Space summons the assistant beside whatever you are working in, and multi-step jobs can be handed to Gemini Spark. Sources: The Neuron, The Deep View. [34](https://blog.google/innovation-and-ai/products/gemini-app/gemini-app-now-on-windows/)
- **Nex-AGI Nex-N2.5-mini**. An agentic model family built to browse, click, and self-correct, released open on Hugging Face. Sources: AlphaSignal. [35](https://huggingface.co/nex-agi/Nex-N2.5-mini)
- **edge0**. Open-source framework that runs a 35B model on a phone with 2 GB of RAM. Sources: AlphaSignal. [36](https://github.com/Edge0-AI/edge0)
- **RadixArk open-sources Miles**. A production-ready reinforcement learning training system for frontier models. Sources: AlphaSignal. [37](https://arxiv.org/abs/2609.08368)
- **Visa, Mastercard, and Ant International team up on agent identity**. The payment networks launched a shared framework for identifying and verifying AI agents that make purchases on a user's behalf. Sources: The Neuron. [38](https://www.reuters.com/technology/payment-firms-visa-mastercard-ant-international-team-up-ai-agent-trust-framework-2026-09-10/)
- **OpenAI pauses new $200 Pro signups to protect Astra access**. Codex lead Tibo says OpenAI is pausing new $200 Pro subscriptions under heavy demand so existing users keep their access. Other plans and the API stay open. Sources: The Neuron, The Deep View. [39](https://x.com/thsottiaux/status/2098113585683808624)

## References

1. [Hugging Face: DeepSeek V4.1 technical report](https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash/blob/main/DeepSeek_V41_Tech_Report.pdf)
2. [DeepSeek: V4.1 Flash](https://deepseek.com/news/deepseek-v4-1-flash/)
3. [Reuters: China's DeepSeek launches V4.1 Flash model](https://www.reuters.com/world/asia-pacific/chinas-deepseek-launches-v41-flash-model-2026-09-10/)
4. [Anthropic: Threat intelligence report, September 2026](https://www.anthropic.com/threat-intelligence-report-september-2026)
5. [Anthropic: Detecting and countering misuse (PDF)](https://www-cdn.anthropic.com/e50be2e51e7695dc4b1366a37a245a597377d3b5/Anthropic-Detecting-and-countering-091026.pdf)
6. [OpenAI: Introducing ChatGPT for Financial Services](https://openai.com/index/introducing-chatgpt-financial-services/)
7. [OpenAI Help: ChatGPT for Financial Services](https://help.openai.com/id-id/articles/20001517-chatgpt-for-financial-services)
8. [OpenAI: Navier-Stokes solution](https://openai.com/index/navier-stokes-solution/)
9. [New York Times: OpenAI and the Navier-Stokes problem](https://www.nytimes.com/2026/09/10/science/tristan-buckmaster-openai-math-navier-stokes.html)
10. [X: Andrew Curran on the next Millennium problem](https://x.com/AndrewCurran_/status/2098083604853342688?s=20)
11. [WSJ: Anthropic researcher quits over out-of-control AI fears](https://www.wsj.com/tech/ai/anthropic-researcher-quits-over-out-of-control-ai-fears-707b7628?st=axTZSz&reflink=desktopwebshare_permalink)
12. [X: Evan Hubinger on alignment risk](https://x.com/evanhub/status/2097497037956891126)
13. [X: Paul Christiano on loss of control risk](https://x.com/paulfchristiano/status/2097733214303645729)
14. [X: Mrinank Sharma's resignation letter](https://x.com/MrinankSharma/status/2020881722003583421?lang=en)
15. [Bloomberg: OpenAI is open to slowing cutting-edge AI, Altman tells staff](https://www.bloomberg.com/news/articles/2026-09-11/openai-is-open-to-slowing-cutting-edge-ai-ceo-sam-altman-tells-staff)
16. [OpenAI: Introducing the Agents API](https://openai.com/index/introducing-the-agents-api/)
17. [OpenAI: Introducing GPT-Live-1 in the API](https://openai.com/index/introducing-gpt-live-1-in-the-api/)
18. [sjg.io: AI is making us build too much](https://sjg.io/writing/ai-is-making-us-build-too-much/)
19. [Worse on Purpose: On quality](https://www.worseonpurpose.com/p/on-quality)
20. [arXiv: Character.AI companion use and well-being](https://arxiv.org/abs/2609.07243)
21. [Interconnects: Lossy self-improvement](https://www.interconnects.ai/p/lossy-self-improvement)
22. [Alignment Forum: Astra can do a concerning amount with no chain of thought](https://www.alignmentforum.org/posts/eRmzz8J8Qkzqvzrgg/astra-can-do-a-concerning-amount-with-no-chain-of-thought)
23. [AI Guide: Misleading metaphors and real risks](https://aiguide.substack.com/p/misleading-metaphors-and-real-risks)
24. [InventBuild.Studio: Genuine creativity is your new moat](https://www.inventbuild.studio/blog/genuine-creativity-is-your-new-moat)
25. [Every: Evals for everyone](https://every.to/context-window/evals-for-everyone)
26. [Cognition: SWE-2](https://cognition.com/blog/swe-2)
27. [X: Cognition on Devin Voice](https://x.com/cognition/status/2098142686486356185)
28. [Cohere: North Small Translate](https://cohere.com/blog/north-small-translate)
29. [Suno: Introducing v6](https://suno.com/blog/introducing-v6)
30. [The Verge: Universal Music and ElevenLabs licensed remix platform](https://www.theverge.com/ai-artificial-intelligence/993465/universal-music-elevenlabs-ai)
31. [Kugiri](https://edoardolunardi.github.io/kugiri/)
32. [Genspark: Gen-1 Slides](https://www.genspark.ai/blog/gen-1-slides)
33. [Google: Dreambeans expansion](https://blog.google/innovation-and-ai/models-and-research/google-labs/dreambeans-expansion-september-2026/)
34. [Google: Gemini app now on Windows](https://blog.google/innovation-and-ai/products/gemini-app/gemini-app-now-on-windows/)
35. [Hugging Face: nex-agi/Nex-N2.5-mini](https://huggingface.co/nex-agi/Nex-N2.5-mini)
36. [GitHub: Edge0-AI/edge0](https://github.com/Edge0-AI/edge0)
37. [arXiv: Miles reinforcement learning training system](https://arxiv.org/abs/2609.08368)
38. [Reuters: Visa, Mastercard, and Ant International AI agent trust framework](https://www.reuters.com/technology/payment-firms-visa-mastercard-ant-international-team-up-ai-agent-trust-framework-2026-09-10/)
39. [X: Tibo on pausing $200 Pro signups](https://x.com/thsottiaux/status/2098113585683808624)

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