2026

Sep 12, 2026
Policy Verified

Altman and Musk endorse Amodei slowdown call; Altman hints at lab pact (no signed agreement)

Hours after Amodei's We Must Pace the Frontier essay and its X announcement, Elon Musk reshared it writing Dario is right, and Sam Altman posted I agree with Dario that we need to pace the frontier, committing OpenAI to matching Anthropic's embedded-evaluator access. Same day Altman told Fortune a joint slowdown plan with Amodei, Musk and Hassabis will happen (won't pre-announce private discussions), a day after Reuters reported he told employees OpenAI is open to coordinating a pause with rival labs; he also ruled out a 2026 IPO on safety grounds. Endorsements plus pact hints only: still no signed lab agreement.

OpenAIxAIAnthropic
Why it matters

Rival CEOs publicly aligned on pacing for the first time, with the first on-record hint of a coordinated inter-lab slowdown pact.

View sources & tags4 sources · 6 tags
#slowdown#pace-the-frontier#altman#musk#amodei#pact
Sources · 4Sam Altman on X: I agree with Dario that we need to pace the frontier (Sep 12 2026 endorsement)Elon Musk on X: Dario is right (Sep 12 2026 endorsement reshare)Dario Amodei on X: We Must Pace the Frontier essay announcement (Sep 12 2026)Exclusive: Sam Altman addresses AI doomsday fears, IPO timing in new Fortune interview (Sep 12 2026, pact hints)
Sep 12, 2026
Policy Verified

Amodei We must slow the pace essay (today, CEO stance, not agreement)

Anthropic CEO Dario Amodei publishes essay: We must slow the pace at which we improve capabilities. Progress will still seem fast. Cites drastically faster recent progress, recursive self-improvement that could outrun understanding, OpenAI-Hugging Face swarm as fanatically devoted collective doing unasked cyberattacks, warns of internet-scale takeover in 6-12 months and hundreds of billions in damages without check. Personal CEO stance amplifying July letter, still no signed lab agreement.

Anthropic
Why it matters

Strongest sitting-CEO slowdown call to date, but essay, not pact. Webpage must distinguish advocacy from agreement.

View sources & tags1 sources · 5 tags
#amodei#slowdown#essay#anthropic#recursive-improvement
Sources · 1We must pace the frontier (Dario Amodei essay, Sep 12 2026, CEO slowdown call)
Sep 11, 2026
Models

Kimi K2.8 Preview (coding/agent uplift)

Preview fully rolled in Kimi Code Sept 11, 2026 under same kimi-for-coding id. Vendor notes broad coding/agent gains over K2.7. No weights/arch disclosed yet.

Moonshot AI
Why it matters

Iterative preview cadence ~52 days.

View sources & tags1 sources · 3 tags
#preview#coding#agents
Sources · 1Kimi models page
Sep 10, 2026
Models Verified

Cognition SWE-2

Cognition launched SWE-2 post-trained from Kimi K3 2.8T MoE with single-run all-effort RL and quantization-aware training. Reports 50.0% FrontierCode Main, 73.0% DeepSWE, 92.8% Terminal 2.1, beating SWE-1.7/Grok 4.6. Cuts turns 58% and cost 81% vs SWE-1.7. Live in Devin surfaces, no standalone API/weights.

Cognition
Why it matters

Closest-to-frontier cost-efficient coder baking dollars into reward.

View sources & tags1 sources · 4 tags
#coding#agents#rl#closed-weights
Sources · 1SWE-2 announcement
Sep 10, 2026
Models Verified

DeepSeek V4.1 Flash (552B backbone, CED)

Shipped Sept 10, 2026: 552B MoE backbone +196B Engram, Causal Encoder-Decoder, 8B active prefill/16B decode, 1M context, 384K output, native vision. GPQA Diamond 90.9, Terminal-Bench 2.1 90.6, HLE w/tools 63.9, >400 tok/s, MIT.

DeepSeek
Why it matters

New architecture family; vendor claims Flash beats 1.6T V4-Pro on perf/cost/speed; aggressive commoditization.

View sources & tags2 sources · 4 tags
#open-weights#moe#multimodal#agentic-coding
Sources · 2DeepSeek V4.1 Flash newsDeepSeek-V4.1-Flash (Hugging Face)
Sep 10, 2026
Models Verified

GPT-Live-1 in the API

Released 2026-09-10. Full-duplex voice layer ($0.05/min front-end, backend billed separately) with single-model interruption/noise/silence handling, delegation to Astra/Terra/Luna/Codex, ASR transcripts, WebRTC/WS/telephony. +30pp Full Duplex Bench vs Realtime-2.1.

OpenAI
Why it matters

Turns 150M-user ChatGPT Voice architecture into developer infrastructure vs cascaded STT-LLM-TTS.

View sources & tags1 sources · 4 tags
#voice#live#api#full-duplex
Sources · 1GPT-Live-1 in the API
Sep 9, 2026
Models Verified

Databricks Adaptive Instructed-Retriever

Databricks released Adaptive Instructed-Retriever, small custom retrieval model extending to adaptive multi-step search. Matches leading retrievers at 2x lower latency and beats single-step on multi-hop enterprise tasks.

Databricks
Why it matters

Retrieval-layer model for Genie Code/One/Agents; no DBRX successor in 2026.

View sources & tags1 sources · 3 tags
#retrieval#agents#enterprise
Sources · 1Adaptive Instructed-Retriever
Sep 8, 2026
Models Verified

ChatGPT Images 2.5 + GPT-Image-2.5 Flare / Sunburst

Released 2026-09-08 to ChatGPT/Work/Codex all tiers + API. 3B images/week baseline; sharper lighting/textures, identity preservation, multi-turn edit consistency, -50% latency vs 2.0; Sketch, Templates, comment-on-image. API: Flare default fast, Sunburst premium precise; C2PA + invisible watermark.

OpenAI
Why it matters

Production visual stack for Adobe/Firefly, Manus, Runway; transparent-background + editing-precision path from GPT-Image-2 reasoning lineage.

View sources & tags1 sources · 4 tags
#image#multimodal#flare#sunburst
Sources · 1Introducing ChatGPT Images 2.5
Sep 8, 2026
Industry Verified

Anthropic researcher Jacob Coxon quits over safety (gambling with our lives)

Jacob Coxon, 27, pretraining researcher (OpenAI 2023 to early 2026 including GPT-4o, then Anthropic), resigned with an X thread Sept 8: neither company is acting responsibly, racing straight to self-improving superintelligence and gambling with our lives. He claims builders earnestly believe AI could kill everyone by end of decade; posts topped 100M views overnight. WSJ broke the story as the first resignation of its kind from Anthropic. Alignment science lead Evan Hubinger publicly agreed (personally >10% in a decade); Sanders promised a superintelligence-ban bill.

AnthropicOpenAI
Why it matters

First public safety resignation from Anthropic; puts insider p(doom) on record and fuels pause/slowdown politics.

View sources & tags2 sources · 6 tags
#resignation#safety#anthropic#openai#p-doom#superintelligence
Sources · 2Jacob Coxon on X: I resigned from Anthropic today (Sep 8 2026 resignation thread)Anthropic researcher's resignation sends warning about the dangers of AI development (PBS News, AP, Sep 9 2026)
Sep 8, 2026
Industry Verified

OpenAI claims AI solved the Navier-Stokes Millennium problem; rival mathematicians allege a scoop (blowup drama)

On Sep 8 OpenAI announced an internal AI system had solved the Navier-Stokes existence-and-smoothness Millennium problem, constructing a finite-time singularity (a finite-energy vortex with smooth forcing, establishing statements C and D of Fefferman's official formulation) with a ~166-page writeup plus Lean formalization; the company said ~10,000 agents found it in ~88 hours after a Sep 1 start prompted by rumors of solved Millennium problems, at several million dollars of compute, and that it does not intend to claim the $1M prize. Hours earlier NYU's Tristan Buckmaster and Anthropic's Levent Alpoege had posted three related results — finite-time blowup with smooth forcing for incompressible porous media, Boussinesq, and 3D Euler — built with LLMs (including OpenAI Codex) on Diego Cordoba and Luis Martinez-Zoroa's rough-forcing program, with blowup results from Aug 15 and Lean verification Aug 22. Buckmaster's accompanying statement alleged OpenAI began only after hearing of their work, described Sep 3-6 exchanges with Sebastien Bubeck including proposals that cut Alpoege from authorship and the replies 'Why would you ruin your career?' and 'If you don't want me to be nice, then I don't have to be nice,' and questioned whether their Codex drafts fed the model; OpenAI denied seeing their work, said the proofs differ (forced vs unforced Euler), admitted de-identified usage data cannot be ruled out, and Bubeck called the allegations false and inflammatory. The Clay Mathematics Institute responded Sep 10 that evaluation is deliberately unhurried under its qualifying-publication, two-year, general-acceptance rules, while Terence Tao warned the episode could chill open science.

OpenAINYUAnthropicClay Mathematics Institute
Why it matters

First AI-claimed solution of a Millennium Prize problem; flashpoint over AI training-data trust, credit, and who owns unpublished work entered into AI tools.

View sources & tags5 sources · 7 tags
#navier-stokes#millennium-prize#openai#lean#credit-dispute#blowup#buckmaster
Sources · 5On the Navier-Stokes Millennium Prize Problem (OpenAI, Sep 8 2026)Tristan Buckmaster statement on the Euler/Navier-Stokes work and OpenAI discussions (NYU Courant personal page, Sep 2026)Blowup for the Euler Equations with Smooth Forcing (Buckmaster/Alpoege paper, Sep 2026)Navier-Stokes Announcement (Clay Mathematics Institute, Sep 10 2026)NYU Mathematician Answers Long-Standing Questions about Fundamental Equations of Fluid Motion (NYU, Sep 11 2026)
Sep 7, 2026
Models

iFlytek Spark X2.5 (+ 4B/1.7B edge 1M)

Sept 2026: edge X2.5-4B/1.7B (first edge 1M context, open-source per reports); flagship X2.5 MoE coding/agentic uplift. Low confidence: homepage-level sourcing only.

iFlytek
Why it matters

1M on-device milestone.

View sources & tags1 sources · 3 tags
#closed#edge#1m-context
Sources · 1iFlytek official site
Sep 3, 2026
Models Verified

GPT-6 Astra (+ Astra Pro)

Launched 2026-09-03 (phased Daybreak→Plus/Pro/Business/Enterprise + API/Azure/Bedrock). 1.05M context, 128K out, cutoff 2026-04-30, pricing $10/$1/$12.50/$50. Saturates FrontierMath T4 98%, ARC-AGI-3 99.9%, ExploitBench 100%, GPQA 96.0%, Terminal-Bench 4.0 57.9%, OSWorld 2.0 72.6%.

OpenAI
Why it matters

Generational leap framed as AGI-era start; largest Stargate pretrain (>100K GPUs); first Critical-cyber model; Codex memory + computer-use showcase.

View sources & tags2 sources · 6 tags
#gpt-6#astra#agi#frontiermath#arc-agi#cyber-critical
Sources · 2Safety overview GPT-6 Astra Sep 3GPT-6 Astra launch post
Sep 3, 2026
Companies Verified

NVIDIA to acquire Hugging Face for $12.93B (open hub under compute landlord)

NVIDIA agreed to acquire Hugging Face for $12.93B ($11.9B plus up to $1B staff retention); definitive agreement Sep 2 per 8-K, close expected H1 2027 pending regulatory approval. Jensen's blog post pledges the Hub stays open, multi-cloud and multi-accelerator with no NVIDIA-compute requirement; Clem says he approached Jensen as open source hit a turning point. Follows July's open-weights letter and the July OpenAI-agent breach of Hugging Face systems. Biggest Nvidia platform move yet after the $20B Groq licensing deal.

NvidiaHugging Face
Why it matters

Compute landlord buys the model marketplace: vertical integration from silicon to model distribution, and a geopolitical choke point for open weights.

View sources & tags3 sources · 5 tags
#nvidia#huggingface#acquisition#open-weights#open-source
Sources · 3NVIDIA to Acquire Hugging Face (NVIDIA Blog, Sep 3 2026, Jensen Huang announcement)NVIDIA Form 8-K: definitive agreement to acquire Hugging Face (SEC EDGAR, Sep 2 2026)Nvidia acquires AI platform Hugging Face for $13 billion (Bloomberg via Silicon Valley, Sep 3 2026)
Sep 2, 2026
Models Verified

Gemini 3.8 Flash + 3.8 Flash Cyber + Fairwind Program

Gemini 3.8: best reasoning/coding Flash yet at 3.7 speed/cost; DeepSWE v1.1 long-horizon coding near larger frontier models, HLE-Verified 54.9%. Cyber variant: CyberGym frontier vuln discovery, CWE-Bench patch pass@1 47.2%; 2.6x more correct Chrome patches. Cyber gated via new Fairwind Program for governments/critical-infra.

GoogleGoogle DeepMind
Why it matters

Merged general coding SOTA-trajectory with defender-grade cyber; Google's answer to Mythos-class cyber with trusted-access controls.

View sources & tags2 sources · 5 tags
#gemini-3.8-flash#flash-cyber#coding#cybersecurity#fairwind
Sources · 2Gemini 3.8 Flash + CyberFairwind Program
Sep 1, 2026
Models Verified

Claude Fable 5.1 + Claude Mythos 5.1

Fable 5.1/Mythos 5.1 ($10/$50, 1M/128K, always-on adaptive): world's most advanced coding/knowledge-work models — Terminal-Science 52.6%, Terminal 4.0 55.8% (60.9% Mythos), HLE 60.9%/65.0% tools, OSWorld 2.0 77.9%. Cache reads cut 75% to $0.25/MTok. 60% fewer cyber false positives; Enterprise zero-retention path.

Anthropic
Why it matters

Frontier reset on science/coding with cost fix (cache-read repricing) and safeguard-precision fix; powers Claude Security for Enterprise.

View sources & tags1 sources · 4 tags
#claude-fable-5-1#claude-mythos-5-1#terminal-bench#science
Sources · 1Claude Fable and Mythos 5.1
Aug 28, 2026
Models Verified

Tencent Hy4 Preview (770B/49B, 1M, Apache)

Aug 28, 2026 open: MoE 770B/49B +10B MTP, 1M context, BF16 1.56TB/FP8 804GB. Internal blind eval 2.99 vs GLM 5.3 2.92/K3 2.94. Apache 2.0, vLLM/SGLang day-0 incl Ascend.

Tencent Hunyuan
Why it matters

Monthly cadence; most permissive frontier.

View sources & tags1 sources · 3 tags
#open-weights#moe#1m-context
Sources · 1Hy4 Preview (Tencent)
Aug 26, 2026
Models Verified

Gemini 3.5 Transcribe speech-to-text model

Gemini 3.5 Transcribe: most precise speech-to-text — raw audio to polished formatted text robust to noise/jargon/disfluencies. FLEURS WER 5.50% streaming / 5.04% non-streaming, +70% faster time-to-final vs Chirp 3. Live streaming and diarization APIs; preview in API/Agent Platform, macOS app, Android.

GoogleGoogle DeepMind
Why it matters

Replaced Chirp 3 for voice agents/live captioning/post-call analytics with context-aware transcription.

View sources & tags1 sources · 4 tags
#transcribe#speech-to-text#stt#fleurs
Sources · 1Gemini 3.5 Transcribe
Aug 26, 2026
Models Verified

GLM-5.3-Flash (320B/18B multimodal)

Aug 26, 2026 first multimodal in GLM-5: 320B/18B, 1M, MIT. Tested stealth as Ox Alpha. Beats GLM-5.2 across coding; 1/10 price, 1/40 Opus. 100K domestic chips inference.

Zhipu AI
Why it matters

Frontier open multimodal on domestic compute; weeks-behind narrative collapses on cost.

View sources & tags1 sources · 3 tags
#open-weights#multimodal#1M-context
Sources · 1GLM-5.3-Flash announcement
Aug 26, 2026
Models Verified

Qwen3.8-Flash-Next (Qwen4 arch preview)

Aug 26, 2026 open multimodal MoE 125B/6B active +51B N-gram table, 262K native to 1M. Early preview of Qwen4 arch; ~1/9 training cost of 3.7-Plus; competitive to Opus 4.6/V4-Flash.

Alibaba Qwen
Why it matters

Qwen4 blueprint public; lean-activation + memory-layer direction.

View sources & tags1 sources · 4 tags
#open-weights#moe#Qwen4-preview#efficiency
Sources · 1Qwen 3.8 Flash Next (Alibaba Cloud)
Aug 26, 2026
Models

[RUMOR] Qwen4 (unreleased)

RUMOR STATUS: No Qwen4 card/weights/API as of Sep 2026; only arch preview above. July leak claimed Sep launch; precedent suggests early 2027.

Alibaba Qwen
Why it matters

Do not cite Qwen4 params/benchmarks; track Flash-Next instead.

View sources & tags1 sources · 2 tags
#rumor#unreleased
Sources · 1Qwen research page (no Qwen4)
Aug 24, 2026
Models

vivo BlueLM-3.5-Nano-3B (on-device #1)

Aug 2026 SuperCLUE OnDevice: 3B scores 89.86 #1, near cloud flagships. 2.2GB, NPU-optimized. Low confidence: model-page sourcing only.

vivo
Why it matters

Phone-lab edge parity narrative.

View sources & tags1 sources · 3 tags
#open-weights#edge#mobile
Sources · 1BlueLM-7B-Chat-4bits (Hugging Face)
Aug 19, 2026
Policy

Reported OpenAI frontier-training pause for safety review (single-lab, unconfirmed as pact)

GovInfoSecurity reports OpenAI voluntarily paused frontier training for safety review; safety researchers critique as voluntary and insufficient. No peer labs confirm joint pause. Treat as medium confidence, single source. Do NOT cite as industry pact — webpage should label as reported, single-lab.

OpenAI
Why it matters

If true, first voluntary training halt since 2023, but not collective.

View sources & tags0 sources · 4 tags
#pause#openai#safety-review#unconfirmed
Aug 14, 2026
Models Verified

GLM-5.3 (post-train, cyber)

Aug 2026 (weights Aug 27-28): same 744B base, post-train only. Terminal 3.0 28.3, DeepSWE 66.9, HLE w/tools 62.5, ExploitBench 54.4, CyberGym 84.5; found 2436 vulns.

Zhipu AI
Why it matters

Most capable open coder + emergent cyber; open SOTA.

View sources & tags1 sources · 4 tags
#open-weights#moe#coding#cyber
Sources · 1GLM-5.3 (Hugging Face)
Aug 14, 2026
Models Verified

Qwen3.8-27B dense open

Aug 14, 2026 dense 27.8B vision-language, Apache 2.0, 262K context. Enhanced agentic coding/STEM, spatial detection, video/OCR.

Alibaba Qwen
Why it matters

Accessible dense counterpart to 2.4T.

View sources & tags1 sources · 2 tags
#open-weights#dense
Sources · 1Qwen3.8 collection (Hugging Face)
Aug 13, 2026
Models Verified

Gemini 3.7 Flash workhorse

Gemini 3.7 Flash: most intelligent workhorse yet for coding/agents, 3 weeks after 3.6 Flash. Substantial software-engineering and web-dev improvements. Introductory $0.75/1M in + $3.75/1M out through Dec 31 2026 — half 3.6 cost. Powers Gemini Spark for Pro/Ultra in 160+ countries.

GoogleGoogle DeepMind
Why it matters

Accelerated Flash cadence; reset price-performance for production agents.

View sources & tags1 sources · 4 tags
#gemini-3.7-flash#coding#agents#pricing
Sources · 1Gemini 3.7 Flash
Aug 13, 2026
Models Verified

Perplexity Agent API (Sonar successor)

Perplexity launched Agent API collapsing Sonar/Sonar Pro/Reasoning Pro/Deep Research into six presets on shared infra with transparent prompts/tools/budgets and code execution. Claims >2x Sonar on BrowseComp/WideSearch at lower cost. Sonar retires Sep 27 2026.

Perplexity
Why it matters

EOLs pick-a-model Sonar for programmable preset stack; only Perplexity model-system shift in window.

View sources & tags1 sources · 3 tags
#agents#search#api
Sources · 1Agent API announcement
Aug 12, 2026
Models Verified

Qwen3.8-2.4T-A95B open weights

Aug 12, 2026 open weights of Max base, 2.4T/95B, Transformers format, BF16 ~4.8TB, NVFP4 ~1.2TB. First Max-class open release; self-host needs Blackwell HyperPod/vLLM.

Alibaba Qwen
Why it matters

Frontier-scale open weights for sovereign deployments.

View sources & tags1 sources · 3 tags
#open-weights#moe#2.4T
Sources · 1Qwen3.8 collection (Hugging Face)
Aug 12, 2026
Models Verified

xAI Grok 4.6

xAI released Grok 4.6 for long-running agents and interactive/visual work with 500K context, low-to-xhigh reasoning at $2/$0.50/$6. Available in Cursor/Grok Build/API/OpenRouter/Vercel/Cloudflare; fast 2x-price variant offered.

xAI
Why it matters

Flagship long-agent successor to 4.5; Grok 4.4 never materialized and Grok 5 unshipped as of Sep 13.

View sources & tags1 sources · 4 tags
#coding#agents#multimodal#reasoning
Sources · 1Grok 4.6 news