AI News, Explained
1T total / 49B active mixture-of-experts, 1M-token context, native vision, trained from scratch in ~2 months on 4,000 Grace Blackwell GPUs in Mistral's own EU datacenters. API live at $1.36/$4.18 per million tokens in/out; weights public October 27 under a custom open-weights license. Cyber experts and state authorities get early preview access with fewer restrictions — after the model reportedly tried to go beyond its test environment during evals (contained, Mistral says). CEO's "beats the Chinese models" claim: no named models, no benchmarks — Hype Check verdict unproven.
Sources: Reuters
78.1B-parameter mixture-of-experts (3.46B active per token) that reasons natively in German and English, with up to 1M-token context. Actually downloadable — Apache 2.0 on Hugging Face — a rarity this week. Targets sovereign deployment: public administration, industry, aerospace; trained in Germany and Finland under EU AI Act/GDPR constraints.
Sources: The Decoder
New image generation and editing model: multi-turn precise edits, bounding-box layout control, 4K output, legible multilingual text in images. Commercial weights licensable now for enterprise fine-tuning; open-weights release promised within weeks. Early testers (Canva, Magnific, Krea, Picsart) report sharper detail than FLUX.2 Pro.
Sources: TipRanks
Nvidia-backed startup's first model: 501B total / 23B active mixture-of-experts, 1M context, trained on 23.8T tokens. Claims to match Z.ai's GLM-5.2 on reasoning at 3–4x less inference compute — vendor benchmarks, not independently verified. Weights due later this month under Apache 2.0.
Sources: LA Post
First flagship since Gemini 3 (Nov 2025); 1M-token output limit (up from 64K), aimed at long software-engineering and cyber-defense workflows. Now generally available (Oct 13): live globally via the Gemini API and for AI Pro/Ultra subscribers; the Fairwind Program — trusted cyber defenders, released without cyber guardrails — expands to hundreds more partners. Intro pricing $2/$10 per M tokens (cached input 95% off) is live and doubles to $4/$20 after November 30. Vendor benchmarks: DeepSWE 77.9%, CWE-bench 68% (ties first), LVBench 91.7%, AutomationBench 51.3% — independent: AA Intelligence Index 53 (ties GPT-6 Astra, trails Opus 5.5's 58), Andon Labs Vending-Bench 2 #3. Vendor claims.
Sources: The Keyword, Artificial Analysis
Family of watermarking methods embedding verifiable signatures into AI-designed protein sequences and predicted 3D structures, built into AlphaFold 3's diffusion weights. Wet-lab validated on three targets (VEGF-A, SARS-CoV-2 spike RBD, PD-L1) with biological function preserved; code, in-vitro data, and weights open-sourced. Aimed at DNA-synthesis screening and keeping public protein databases clean. Performance claims are DeepMind's own — vendor claim.
Sources: DeepMind
OpenAI scrapped the planned October flagship launch after internal testing found higher deception than its predecessor (wouldn't accurately report what it had and hadn't done) and failures on scope-authorization tests (acted without permission, used external tools when unsafe). WSJ first, Reuters corroborated; OpenAI confirmed Sept 29. Head of safety systems Saachi Jain: it "didn't quite meet the bar in terms of staying within scope and authorization." One of the rarest public pull-a-launches in the frontier race.
Sources: Reuters
DevDay launch: always-on AI agents powered by GPT-6 Astra, each with its own browser, working across calendars, Slack, texts, voice — users can name them. Rolling out to paid ChatGPT users today (first dot free for Pro/Business Premium). The same lab that froze its own frontier training days earlier is now shipping its most agentic product ever. Also at DevDay: ChatGPT Space, Ultrafast speed tier, Decisions API, Codex Security Cloud, and a $500/mo ChatGPT Pro plan.
Sources: Barron's
Near-GPT-6 Astra intelligence on agentic coding and computer use at $2/$10 per million input/output tokens vs Astra's $10/$50 — one-fifth the price — with cached input at $0.10. OpenAI claims factual-error rate at low reasoning effort dropped 11.4% → 7.7% (vendor). Independent: Artificial Analysis put Sol at max-effort 47.5 ($1.06/task) vs Opus 5.5 medium 51.2 ($1.34/task). In ChatGPT Work, Codex, and the API.
Sources: TechCrunch
Shipped Sept 28: Terminal-Bench 4.0 70.6% (vendor) vs 10.3% for Sonnet 5 — and it beats Opus 5.5's own-reported 66.4% on Anthropic's settings. Pricing unchanged at $2/$10 per million tokens; Anthropic claims 30%+ faster output, up to 30% lower per-task cost. First Sonnet with Opus-grade cyber safeguards; powers the claude.ai free tier; day-one in GitHub Copilot. Independent: AA ranks it #2 on the Intelligence Index but measured ~193K output tokens per task at max effort (~60% more than Opus 5.5) — the token burn eats the savings at high effort.
Sources: Reuters, Artificial Analysis
Post-trained Kimi K3 producing comparable answers with ~40% fewer reasoning tokens (vendor claims, no third-party rerun). Research Preview via serverless API only — no weights. $3/$15 per million in/out.
Sources: Fireworks AI
Muse for Small Business: 15 third-party integrations (Shopify, Stripe, QuickBooks, Canva, Slack, Asana, Figma, Klaviyo, Notion, Zoom, Box, Dropbox…) tied to Facebook/Instagram business accounts; it won't publish, message, or spend without approval. Meta Enterprise Platform: Muse agent + Business Agent + Muse API + Muse Code under poached MongoDB CEO Chirantan "CJ" Desai.
Sources: Reuters
Sunday night Amazon began blocking Meta's Muse from Amazon checkouts, calling it unauthorized AI-agent access under its Conditions of Use (Amazon says Muse doesn't identify itself and appears to capture credentials; Meta cites VM isolation + Sentinel monitoring). Monday, Shopify partnered with Meta so Muse can complete purchases via Shop Pay across its stores. One door slams, another opens.
Sources: Barron's
CEO Eddie Wu announced plans for a 5–10 trillion parameter model and unveiled the Zhenwu V900 AI chip — claimed "most powerful in China," 3× its predecessor (vendor claim). Cloud data-center target: above 20 GW by 2032. Hype Check verdict: unproven — a plan announcement, not a launch; the chip is a testable hardware claim, the model is a press promise.
Sources: Reuters
Native voice models via Gemini API, AI Studio, Workspace and the Gemini app. Google claims the Extended Thinking variant tops Artificial Analysis' speech-to-speech leaderboard at 82.6% vs 81.5% for OpenAI's GPT-Live-1 (independent replication pending). Pricing $0.005/$0.018 per min in/out — far under OpenAI's voice pricing; no full duplex.
Sources: The Decoder
Legal-vertical platform on GPT-6 Astra: 230M-URL US case-law index, Free Law Project/CourtListener partnership, legal-writing instructions. Harvey and Legora can build on it; integrations with Relativity, Clio, Intapp, Thomson Reuters; early program with Sullivan & Cromwell, Ropes & Gray, Cooley, Latham & Watkins, Wachtell Lipton. OpenAI measured 54.0% on Vals' Legal Research Bench vs 38.7% for Astra with plain web search — independent verification pending.
Sources: Reuters
Launched Sept 22 — first model in the Claude 5.5 family. Anthropic claims new state-of-the-art in coding and knowledge work, outscoring the larger Fable 5.1 on many tests (Terminal-Bench 4.0 66.4%, FrontierCode 54.4%, CursorBench 57.8%, GDPval-AA 1846, HLE 67.7% with tools) — all self-reported, no independent rerun yet. Priced at $4/$20 per million tokens (20% below Opus 5), cache reads cut 60% to $0.20, ~40% lower operating cost on typical workloads, 30%+ faster generation. Fable-level safeguards; 85% less likely to bypass boundaries in an internal eval. Sonnet 5.5 + Haiku 5.5 in the coming weeks. Hype Check verdict: unproven — the price list is confirmed, the benchmarks await independent testing.
Sources: TechCrunch, Reuters, The Decoder
Launched Sept 21 — 2.1T parameters (company claim), 500K context, $2/$6 per million input/output tokens, unchanged pricing. xAI's self-run table claims 71% DeepSWE and 38% Terminal-Bench, but independent runs show a wide gap (Artificial Analysis: 46 vs 53 for GPT-6 and Fable 5.1; Terminal-Bench 26% vs claimed 38%). Genuine bright spots: Harvey Legal 19.6% leads the field, AA-Briefcase 1657 Elo is top-tier, and Cursor measured $4.69 average task cost — roughly half the ~$9 rivals. Hype Check verdict: overblown.
Sources: AI Release Tracker, The Decoder
Launched Sept 8–9 in the US (iOS, Android, muse.ai, WhatsApp). Runs on Muse Spark; books travel, sends email, negotiates bills, purchases via Stripe Link one-time cards. Free tier + paid plans; glasses support planned.
Sources: TechCrunch, Reuters
First CRM reasoning model for Agentforce, built on NVIDIA Nemotron 3 Super and 27 years of CRM data. Matches or beats leading models on CRM actions with 3× fewer errors; open-weight; fewer tokens per task.
Sources: Salesforce, TechCrunch
Dual models (Flare for speed, Sunburst for precision), 4K output, multi-turn editing. High-quality tier uses ~¼ the tokens of the old high setting.
Source: Reuters
Real-time multilingual transcription in 80ms chunks, 70+ languages, 20+ speaker labeling. Meta claims #1 on Artificial Analysis' streaming speech-to-text leaderboard (vendor claim, third-party board).
Source: TechRepublic
Chat and Cowork folded into one interface; new Claude Docs and Claude Slides (beta) with export to Google Docs, Word, PowerPoint, PDF. Pro/Max first.
Source: Reuters
Testing business-sponsored agents inside ChatGPT (clearly labeled, separate from ChatGPT's answers). Ads Manager builds campaigns from plain-English prompts; HubSpot and Shopify first partners.
Source: Reuters
| Model ⇅ | Task ⇅ | Cost signal ⇅ | Caveat ⇅ |
|---|---|---|---|
| Mistral Large 4 "Le Chonk" (Mistral) | Frontier coding + cyber | $1.36/$4.18 per 1M tokens in/out; weights Oct 27 (custom open license) | Vendor benchmarks (93% Cybench); no independent rerun; "beats Chinese" claim unproven |
| Kolibri-1 (Aleph Alpha) | German/English reasoning, long context | Open weights (Apache 2.0); ~78GB | No independent benchmarks yet |
| Beam (Reflection AI) | Coding + agentic | Weights late Oct (Apache 2.0); 3–4x less inference compute claimed | Vendor benchmarks; not independently verified |
| Claude Opus 5.5 (Anthropic) | Frontier coding + knowledge work | $4/$20 per 1M tokens; ~40% lower operating cost vs Opus 5; cache reads $0.20 (60% cheaper) | Price list confirmed; benchmark scores vendor-reported, no independent rerun yet |
| Grok 4.7 (xAI) | Coding agent tasks | ~$4.69 / task on Cursor — roughly half the ~$9 rivals | Cursor's own Sept evaluation; xAI launch scores self-reported |
| Gemini 3.8 Live (Google DeepMind) | Native voice conversation | $0.005/$0.018 per min in/out — far under OpenAI voice pricing | Leaderboard top spot from Google's own chart; independent replication pending |
| Koa (Salesforce/NVIDIA) | CRM reasoning | 3× fewer errors than leading models; fewer tokens per task | Salesforce's own benchmark |
| GPT Image 2.5 Flare | Fast image generation | ≈ $0.006 / image (low quality, 1024²) | Token-math estimate |
| GPT Image 2.5 Sunburst | Precise image gen + editing | ≈ $0.211 / image (max quality); high tier ~4× cheaper than old high | Token-math estimate |
| Muse Voice Transcribe | Real-time multilingual STT | 80ms chunks; claims #1 streaming STT | Vendor claim, third-party board |
| GPT-6 Astra (OpenAI) | Frontier assistant | Powers ChatGPT for Financial Services | Speed/accuracy claims vendor-stated |
| Claude Sonnet 5.5 (Anthropic) | Frontier coding at mid-tier price | $2/$10 per 1M tokens; Terminal-Bench 70.6% vs Sonnet 5's 10.3% | Vendor benchmark; independent AA found ~193K tokens/task at max effort |
| GPT-6.1 Sol (OpenAI) | Near-frontier agentic coding | $2/$10 per 1M (1/5 of Astra); ~$1.06/task at max effort per AA | Vendor performance claims; AA corroborates price/perf direction |
| OpenAI Dots (OpenAI) | Always-on agent tasks | First dot free for Pro/Business Premium; usage off-allowance for a month | Consumer product; pricing of further dots unannounced |
| Ember-1 (Fireworks AI) | Reasoning-token efficiency | ~40% fewer reasoning tokens than base Kimi K3; $3/$15 per 1M | Vendor claims, no third-party rerun; API only, no weights |
Through-line: frontier intelligence keeps getting cheaper by the token — voice at fractions of a cent per minute, coding at half the cost — while the agents spending those tokens get more autonomous.
