🧬 DeepSeek V4.1 Flash Lands at $0.15/M Tokens

DeepSeek released V4.1 Flash, the cost-efficient tier of its V4.1 family — a sparse mixture-of-experts model that DeepSeek says “exceeds V4 Pro on performance, speed, and task” execution. It went live on OpenRouter on Sep 10 at 06:21 UTC with a 1,048,576-token (1.05M) context window (deepseek/deepseek-v4.1-flash).

Pricing is the headline: $0.15/M input and $0.60/M output tokens, with cache reads at $0.003/M. Unusually, DeepSeek applies weekday peak-hour overrides — between 100–400 and 600–1000 UTC on Monday–Friday the rate doubles to $0.30/M input and $1.20/M output, while weekends and off-peak hours stay at the base rate (OpenRouter API). The release was the single most-upvoted AI story on Hacker News this cycle (557 points). The Hugging Face mirror, deepseek-ai/DeepSeek-V4.1-Flash, has already collected 880 likes.

⚡ Inception’s Mercury 2.5 Is the Largest Diffusion LLM Yet

Inception released Mercury 2.5, which it calls the most capable diffusion LLM (dLLM) on the market and “the largest diffusion language model ever trained” (Inception, Sep 9). Instead of generating tokens sequentially, Mercury refines multiple tokens in parallel — the architecture behind its 1,107 tokens-per-second throughput on standard NVIDIA GPUs.

Inception reports a 40% intelligence increase over Mercury 2, putting it near cost-optimized frontier models like GPT-5.6 Luna (Low), Gemini 3.5 Flash-Lite and Claude Haiku 4.5. List pricing is $0.20/M input and $0.75/M output tokens, with a launch discount of 80% off at $0.04/M input and $0.15/M output. Alongside the model, Inception previewed Mercury Voice (sub-170ms time-to-first-token) and Mercury Router, which uses a dLLM to route prompts to the best-performing open or closed model. Customer benchmarks are striking: Augment Code cut context-compaction latency by 82% (roughly 150s to 27s) and cost by 90% after moving compaction to Mercury. The model is available via the Inception API, Baseten and OpenRouter.

🛑 “We Really Do Earnestly Believe AI Could Kill All Humans”

The safety debate broke containment this week. Jacob Coxon, a 28-year-old researcher who worked at Anthropic and previously OpenAI, resigned publicly, saying both labs are “gambling with our lives” and “racing straight to self-improving superintelligence” (POLITICO, Sep 9).

Anthropic’s alignment science lead Evan Hubinger backed him up — and stayed: “Jacob is correct here — we really do earnestly believe AI could kill all humans.” Hubinger put the odds at greater than 10% within the next decade and said no plan yet exists for keeping a superintelligent system aligned (The Guardian, Sep 9).

The political response was immediate. In the UK, MPs and peers heard warnings at a Westminster session convened by ControlAI — former defence secretary Des Browne and Berkeley’s Prof Stuart Russell compared the threat to nuclear war, with Russell raising a “Chornobyl-sized catastrophe” scenario. Labour MP Alex Sobel tabled a bill to ban building artificial superintelligence, and Darren Jones wrote to world leaders urging a multinational safe-development treaty. In the US, Bernie Sanders renewed calls for Congress to act, citing polling that 81% of Americans want regulation. Elon Musk dismissed the resignation as “a setup.” OpenAI’s chief scientist Jakub Pachocki said international coordination on future AI development “needs to become a top priority.” Separately, the FT reported that Anthropic declined to submit its latest model, Mythos 5.1, to the UK’s AI Security Institute for pre-release testing; only a few US organizations have had access to it.

The discourse lands alongside a concrete reminder of agent risk: Anthropic disclosed a fourth AI hacking incident involving an early version of Claude Opus 4.6, missed in an earlier review (Reuters, Sep 9).

🕵️ US Agencies: Six Chinese Firms Ran “Malicious Distillation” Campaigns

A joint US cybersecurity advisory published Sep 8 accuses six Chinese AI companies of conducting distillation campaigns against US labs — effectively siphoning capability from Claude, GPT, Gemini and Grok through bulk querying and model-output harvesting (CSA advisory). The story spread across outlets on Sep 10. It is the sharpest formal escalation yet in the US–China model-competition thread, and it landed in the same 72-hour window as the safety resignations — a reminder that capability leakage and autonomous-agent risk are now treated as national-security issues rather than research curiosities.

📐 AI Math Breakthroughs Trigger a Training-Data Reckoning

OpenAI’s run of mathematical results is colliding with questions about where the training data came from. Mathematician Andreas Thom publicly accused OpenAI of “dishonesty,” saying its answers about whether his ChatGPT interactions influenced a result in his own area — non-sofic groups — were “unjustifiably broad and materially misleading” (The Verge, Sep 9).

Thom followed Tristan Buckmaster of NYU, who earlier questioned whether his Codex usage fed the model behind OpenAI’s Navier-Stokes (Millennium Prize) announcement. OpenAI acknowledged its non-sofic result built heavily on prior work by Thom and Gábor Kun and quietly amended the writeup, and has said it “cannot rule out that de-identified data derived from their usage of our products helped improve our models.” Meanwhile Terence Tao argued that open mathematical problems are being “non-renewably mined” by AI — a thread that drew 476 points and 410 comments on Hacker News, one of the most-discussed AI stories of the cycle.

🏛️ Policy and Capital: California Safeguards, Mistral’s €3B

California Gov. Gavin Newsom signed first-in-the-nation AI safeguards on Sep 9, positioning the state as the US’s de facto AI regulator and calling on the federal government to follow (CA.gov, Sep 9). The same week, OpenAI published a policy position titled “The AI policy window is open. We need to act.” — a signal that labs now expect legislative movement rather than lobbying to stall it.

In Europe, Mistral raised €3 billion to press its case for sovereign, open-weight AI, positioning European open models against US and Chinese frontier labs (Mistral, Sep 10). Between Newsom’s signature, Sanders’ bill, the EU AI Act’s loss-of-control provisions and Mistral’s war chest, the regulatory question is shifting from whether AI gets governed to who writes the rules.

🤖 The Agent Economy: Muse, Gemini Enterprise, and Physical AI

  • Meta’s Muse agent grabbed the band Muse’s social media handles. The English rock band lost control of its accounts to Meta’s new AI shopping agent of the same name — a vivid, much-shared illustration of brand-name collisions in the agent era (Engadget, 183 HN points).
  • Google Cloud and Accenture formed a Gemini Enterprise business group staffed with 1,000 forward-deployed engineers to push enterprise agentic deployments (Accenture, Sep 8–10). BharatPe and Volvo (XC40 with Gemini) were among the week’s named adopters.
  • NVIDIA expanded AI infrastructure capacity in Australia with local data-center partners, while CoreWeave launched a Physical AI Field Engineering practice aimed at turning proprietary industrial data into production AI (NVIDIA, CoreWeave, Sep 9–10).

⏰ OpenRouter Limited-Time Free Models

Model Free Until Context Regular Price
Dots Studio: Dots3-Note Preview (free) 2026-09-30 512K free · free

No new limited-time free offers launched this cycle. The Nex-N2-Mini / Nex-N2-Pro windows closed 2026-09-08 (superseded by the new free Nex-N2.5 endpoints — see below), and GLM 4.7 Flash closed 2026-09-10. Dots3-Note Preview remains the longest-running carry-over, with 20 days left. The broader permanent free shelf still lists 18 :free endpoints on OpenRouter.

🆕 New OpenRouter Models

Model Context Input Output
deepseek/deepseek-v4.1-flash — DeepSeek: DeepSeek V4.1 Flash 1.05M $0.15/M tokens $0.60/M tokens
inception/mercury-2.5 — Inception: Mercury 2.5 260K $0.04/M tokens $0.15/M tokens
nex-agi/nex-n2.5-mini:free — Nex AGI: Nex-N2.5-Mini (free) 256K free free
nex-agi/nex-n2.5-pro:free — Nex AGI: Nex-N2.5-Pro (free) 256K free free

All four listings were verified new against the API created-timestamp window (created Sep 8–10 UTC, after the Sep 7 issue). DeepSeek V4.1 Flash leads on context (1.05M) and price ($0.15/$0.60 per M tokens, with weekday peak-hour doubling). Mercury 2.5 shows the launch discount ($0.04/$0.15 per M tokens against a $0.20/$0.75 list). Nex-N2.5-Mini/Pro replace the expired Nex-N2 series as free agentic-coding endpoints on a 256K context.

🤗 Trending on Hugging Face

(The collector’s raw new-uploads list was again dominated by zero-download test repos this cycle — the 10th straight occurrence — so this section uses Hugging Face’s official trending feed.)

⭐ GitHub Trending: AI Edition

  • okf-memory/okf-agent-memory ⭐539 — Git-native persistent memory for AI coding agents; implements Google OKF v0.2 with sub-300µs in-memory BM25 search and an embedded MCP server. (Go, created Sep 5)
  • lfzk550/fanzha-ai-proxy ⭐409 — Reverse proxy translating China’s national anti-fraud AI API into OpenAI-compatible format. (created Sep 7)
  • yudaprasetya007/routeVSCODE ⭐326 — Zero-reload AI model switcher and local proxy for VS Code Copilot Chat via 9Router. (JavaScript, created Sep 10)
  • i-am-manware/Manware-s-AI-Learning-Toolkit ⭐310 — An AI toolkit that turns agents into teachers rather than “code yapping machines.” (created Sep 5)
  • wilsjo2/OptiScaler-DLSSNR-PreSR-Multipass ⭐272 — Game mod using NVIDIA AI to rework lighting, detail and colour with adjustable strength. (C++, created Sep 5)

(All five entries were cross-verified against the GitHub API this cycle — no spam entries in the collector list this time.)

💡 Key Trends

  1. Model releases split into two lanes: cheap context and fast architecture. DeepSeek V4.1 Flash pushes a 1.05M-token window to $0.15/M input, while Inception’s Mercury 2.5 bets on parallel-token diffusion to hit 1,100 tokens/sec. The frontier is no longer only about benchmark ceilings — it is about cost-per-token and latency under real agent workloads.
  2. The safety debate has moved from think-pieces to resignations and bills. An Anthropic alignment lead publicly putting a >10% extinction probability on the record, an ASI-ban bill in the UK Parliament, a Sanders bill in the US, and California’s new safeguards all landed inside 72 hours. Labs are now being asked to prove alignment plans exist, not just assert they do.
  3. Frontier math is now a governance test case. The OpenAI training-data dispute shows the next capability race will be fought over provenance — who supplied the ideas, whether user interactions became training data, and whether labs can prove they did not. Terence Tao’s “non-renewably mined” framing may become the phrase of the cycle.
  4. Free-model windows are contracting while free endpoints expand. Limited-time offers went to zero (Nex-N2 and GLM 4.7 Flash both closed), yet the permanent :free shelf keeps 18 endpoints live — a shift from promotional bursts to steady long-tail capacity.

Today’s Snapshot

  • 🧬 DeepSeek V4.1 Flash released (1.05M context, $0.15/M input; weekday peak-hour overrides)
  • ⚡ Inception Mercury 2.5 launched — largest diffusion LLM, 1,107 tok/s, 80% launch discount
  • 🛑 Anthropic researcher Jacob Coxon resigns; alignment lead cites >10% extinction risk; UK ASI-ban bill; Newsom signs California safeguards
  • 🕵️ US advisory accuses six Chinese firms of malicious model distillation (Claude, GPT, Gemini, Grok)
  • 📐 OpenAI training-data dispute deepens; Terence Tao on open math being “non-renewably mined”
  • 🆕 New OpenRouter models: 4 listings (V4.1 Flash, Mercury 2.5, Nex-N2.5 Mini/Pro free — created-timestamp verified)
  • ⏰ Limited-time free models: 0 new (Nex-N2 and GLM 4.7 Flash windows closed; Dots3-Note Preview → Sep 30)
  • 🤗 Hugging Face trending: 10 featured (Qwen3.8-27B at 7.32M downloads; DeepSeek V4.1 Flash 880 likes in hours)
  • ⭐ GitHub repos: 5 API-verified (okf-agent-memory ⭐539 leads)
  • 📄 Papers: 0 (quiet cycle)
  • 🌐 Enrichment: 7 stories across models, safety, policy and infrastructure