112 releases covered live on the show, led by OpenAI Dots, Claude Sonnet 5.5 — every model, product, paper and tool that mattered, with links and our analysis.
112 AI releases shipped in October 2026, 29 of them in the latest week (Oct 22, 2026), led by OpenAI Dots, Claude Sonnet 5.5, Gemini 4 Argon, CoreWeave Serverless GPUs, Kling 4.0. Earlier in the month: Muse Spark Open, Nemotron 3.5 Super VL, Gemini 4 Argon (GA), Arena–Series funding ($200M). ThursdAI — the weekly AI news podcast hosted by Alex Volkov — tracked every entry below with source links, key numbers and episode coverage (all covered live on the show).
Meta releases Muse Spark Open: 816B open weights, 42B active, 1M context. Meta dropped Muse Spark Open during the last minutes of the show: 816B total parameters, 42B active, 1M context, multimodal in and text out, and Meta calls it the most capable US open-weight model to date. ThursdAI covered it live on the show, with primary sources linked below.
17 open-weights models shipped in October 2026, led by Muse Spark Open (816B / 42B total / active parameters), Nemotron 3.5 Super VL (120B / 12B total / active parameters), Kolibri-1 (78B total parameters), d1-3B and d1-omni-600M (8 ms d1-3B on a GPU). Each entry below links the weights and the episode segment where we covered it.
50 companies shipped AI releases in October 2026; the most active were CoreWeave, OpenAI, Anthropic, Google DeepMind, NVIDIA, DeepSeek. Every launch below has primary-source links and ThursdAI's live episode analysis.
| Release | Best for | Why it matters | Key number |
|---|---|---|---|
| Muse Spark Open · Meta | Open-source builders | Meta releases Muse Spark Open: 816B open weights, 42B active, 1M context | 816B / 42B total / active parameters |
| Nemotron 3.5 Super VL · NVIDIA | Agent builders | NVIDIA drops Nemotron 3.5 Super VL, a 120B open vision model with 12B active | 120B / 12B total / active parameters |
| Gemini 4 Argon (GA) · Google DeepMind | Model evaluators | Gemini 4 Argon goes GA at $2/$10, #1 on Arena Text | #1 Arena Text |
| Arena–Series funding ($200M) | Model evaluators | Arena raises $200M at a $3.1B valuation, broken live on ThursdAI | $200M raised |
| Serverless GPUs · CoreWeave | Infra & GPU planners | CoreWeave launches serverless GPUs: GPU sandboxes by the hour, no contract | Per hour billing, no contract or commitment |
| Kling 4.0 · Kling AI | Video creators | Kling 4.0 generates up to 30 seconds of 4K video in one pass | 30 s max length in one pass |
| Muse Wonder · Meta | Model evaluators | Meta's Muse Wonder lands at #2 on Arena Text, API at $4/$20 | #2 Arena Text (1513) |
| Kolibri-1 · Aleph Alpha | Open-source builders | Aleph Alpha Kolibri-1, a 78B German-English MoE under Apache 2.0 | 78B total parameters |
| Claude Haiku 5.5 · Anthropic | Agent builders | Claude Haiku 5.5 at 10 cents per million input tokens | $0.10/M input tokens |
| d1-3B and d1-omni-600M · Liquid AI | Local & on-device users | Liquid AI opens d1: d1-3B and d1-omni-600M decision models | 8 ms d1-3B on a GPU |
| Mistral Large 4 · Mistral AI | Open-source builders | Mistral Large 4 "Le Chonk": 1T parameters, open weights promised for October | 1T parameters |
| Beam · Reflection AI | Developers & coding agents | Reflection AI announces Beam, a 501B Western open-weight model | 501B total parameters |
FLUX 3 Max from BFL and fal renders a 10-second clip in about 12 seconds
Black Forest Labs and fal launched FLUX 3 Max, which renders a 10-second video clip in about 12 seconds, close to real time.
DeepSeek V4.1-Flash-1017 jumps to 42.3 on Terminal-Bench 4.0, still MIT
DeepSeek shipped another V4.1 Flash checkpoint under the MIT license: Terminal-Bench 4.0 goes from 31.2 to 42.3 with shorter outputs, and DeepSeek claims 77.0 on DeepSWE versus 74.0 for Claude Opus 5 and 73.0 for GPT-5.6 Sol. Its Search and Fetch APIs are now open to everyone.
Inception Mercury 3: a diffusion LLM claiming 1,000+ tokens/sec
Inception released Mercury 3, a diffusion language model it says runs at more than 1,000 tokens per second with a 1M-token context. These are vendor numbers.
Kling 4.0 generates up to 30 seconds of 4K video in one pass
Kling 4.0 makes 3 to 30-second clips in one pass at 1080p or 4K with stereo audio and better lip sync, with up to 10 keyframes and 15 references and a new 21:9 aspect ratio. The API is rolling out.
Krea 3 Ultra ranks #4 on Arena multi-image edit, behind only OpenAI
Krea 3 Ultra placed #4 on Arena's multi-image edit leaderboard, behind only OpenAI's models.
MiniMax reveals Space Bunny as M3.1 Flash at $0.15 / $0.60 per 1M tokens
The anonymous Space Bunny model was MiniMax M3.1 Flash, priced at $0.15 input and $0.60 output per million tokens, with weights due October 23.
Mistral Voxtral 3 adds TTS and a voice agent API to transcription
Mistral's Voxtral 3 grows from transcription into a voice stack: transcription, text-to-speech and a voice agent API.
Recursive Flint: a 39.8B / 2.5B-active local coding model, Apache 2
Flint from Recursive (Richard Socher's lab) is a 39.8B mixture of experts with 2.5B active, built to run fast for coding on a Mac, under Apache 2.0. Its architecture was chosen by Recursive's Eureka Machine, and the team went from no pre-training effort to release in under 4 months. Prince Canuma shipped day-zero MLX support.
Underdog Bark 3.0: 163 MB open speech-to-text that ties Parakeet
Underdog released Bark 3.0, a 163 MB speech-to-text model under CC-BY 4.0 that it says ties NVIDIA Parakeet and runs fully on device.
vLLM Semantic Router ships Decision 3.0 (D3) decider models
The vLLM Semantic Router project released Decision 3.0 (D3) decider models on Hugging Face. Nisten says D3 Turbo scores better than Jev in his tests, accepts video, and has a free endpoint.
Datalab Indra: a 1M-parameter OCR model that runs 49 pages/sec on CPU
Datalab released Indra, an OCR model with only 1M parameters that processes 49 pages per second on a CPU.
Fish Audio releases Drama 3, open-weights TTS you can direct
Drama 3 is Fish Audio's open-weights text-to-speech model that lets you direct how a voice sounds. It was Wolfram's pick of the week.
Gemini 4 Argon goes GA at $2/$10, #1 on Arena Text
Two weeks after Argon debuted to trusted testers only, Google made Gemini 4 Argon generally available in the Gemini API, AI Studio, the Gemini app for Pro and Ultra, and Antigravity for Ultra. It is #1 on Arena Text and on Roboflow's vision evals, priced at $2 in / $10 out per million tokens until November 30, then $4 / $20. Artificial Analysis has it tied with GPT-6 Astra at 53 (Opus 5.5 leads at 58) at about 60% of Astra's cost per task at the discount price.
Liquid AI ships LFM2.5-24B-A2B, its biggest model yet, for local agents
LFM2.5-24B-A2B is a 24B-parameter, 2B-active mixture-of-experts model, Liquid AI's biggest release so far, aimed at local agents.
Meta releases Muse Spark Open: 816B open weights, 42B active, 1M context
Meta dropped Muse Spark Open during the last minutes of the show: 816B total parameters, 42B active, 1M context, multimodal in and text out, and Meta calls it the most capable US open-weight model to date. On Terminal-Bench 2.1 it scores 76.4 versus 83.9 for GLM 5.3 and 85 for Kimi K3. Nisten read the Muse Model License live: products must show 'built with Muse Spark Open' and there is a $20M annual revenue cap.
Meta's Muse Wonder lands at #2 on Arena Text, API at $4/$20
Meta's first frontier model comes in two versions: Muse Wonder in the API at $4 in / $20 out per million tokens, and the stronger Muse Wonder Full, only for Muse Code subscribers at $50 or $200 a month. Neither has open weights. It sits at #2 on Arena Text at 1513, 12 points behind Gemini 4 Argon, though Peter Gostev noted it is 12th on Agent Arena, and Artificial Analysis scores Wonder Full at 56 and Wonder at 52 versus 58 for Opus 5.5.
Microsoft-Decision-1: a 9B decision model at $0.042 per million input tokens
Microsoft released Microsoft-Decision-1, a 9B decision model priced at $0.042 per million input tokens, separating decision-making from generative AI.
NVIDIA drops Nemotron 3.5 Super VL, a 120B open vision model with 12B active
Nemotron 3.5 Super VL is a 120B-parameter (12B active) vision-language model for images, documents and video, built for long-running agents, with open weights, data and recipes. NVIDIA claims up to 5x the throughput of open models its size; Peter Gostev pitched it near DeepSeek V4 Flash level but faster.
OpenAI GPT-Live-1.1 comes to ChatGPT Voice, Codex and dots
OpenAI shipped GPT-Live-1.1, an update to its live voice model, for ChatGPT Voice, Codex and dots.
OpenBMB MiniCPM-V 4.7 runs a 35B-A3B vision MoE on an iPhone in ~3GB
MiniCPM-V 4.7 is a 35B-A3B multimodal mixture-of-experts model that runs on an iPhone in about 3GB of memory, released alongside a 1B variant.
Amos 1: four open decision models from 0.5B to 27B with bounding boxes
Amos 1 is a family of four open decision models ranging from 0.5B to 27B parameters that can also return bounding boxes.
TypeSafe ships Jev 1.14 with preview image input
TypeSafe shipped Jev 1.14 on 10.14 at 1:14pm, with image input as a preview, a new box type and a longer context window. It lands a month after Jev's launch.
xAI's Grok Voice 3.0 takes #1 on Artificial Analysis speech-to-speech
Grok Voice 3.0 scores 86.9 on Artificial Analysis speech-to-speech, ranking #1, at $0.08 per minute. Wolfram had his agent wire it into his voice app during the show.
Aleph Alpha Kolibri-1, a 78B German-English MoE under Apache 2.0
Aleph Alpha released Kolibri-1, a 78B mixture-of-experts model with 3.46B active parameters, trained from scratch on German and English, with 1M context, under Apache 2.0. It fits on a single H200. Wolfram's verdict from vacation: a promising specialized German tool worker.
Claude Haiku 5.5 at 10 cents per million input tokens
Anthropic brought Haiku back after a year: Claude Haiku 5.5 costs $0.10 per million input tokens and $0.50 per million output (under 100K tokens), with cache reads at 1 cent, ten times cheaper than Haiku 4.5. Anthropic reports 72.4% on OSWorld versus 48.9% for GPT-6 Luna, and 39.2% on Terminal-Bench 4.0 versus 16.4%.
FLUX 3 Image: 4K, bounding-box layouts and 10 reference images
Black Forest Labs released FLUX 3 Image with 4K output, bounding-box layouts and up to 10 reference images.
Cloudflare Clef, open decision models
Cloudflare released Clef, open decision models.
Fastino GLiDE, a thinking decision model
Fastino released GLiDE, a decision model that thinks before it decides.
EmbeddingGemma 2, an open multimodal embedding model
Google released EmbeddingGemma 2, an open multimodal embedding model, the same week Perplexity opened pplx-embed-v2.
Google Nano Banana 2.1 at $0.0336 per 1K image
Google released Nano Banana 2.1 at $0.0336 per 1K image. Every infographic on the show was made with it, and Alex found it really good on high reasoning.
Liquid AI opens d1: d1-3B and d1-omni-600M decision models
Liquid AI opened its d1 decision models: d1 behind an API, the open d1-3B with text and vision, and d1-omni-600M, which also takes audio. d1-3B answers in 8ms on a GPU and about 50ms on a Jetson Orin Nano, tops Liquid's Decision Index under 10B parameters, and the API is a drop-in replacement for Jev.
Mistral Large 4 "Le Chonk": 1T parameters, open weights promised for October
Mistral's Large 4, nicknamed Le Chonk, is a trillion-parameter multimodal model with about 50B active and 1M context; open weights are promised for the end of October. Artificial Analysis scores it 38, the same as GPT-6 Luna, at about $1.13 per task versus 7 cents for Luna.
Perplexity opens pplx-embed-v2 multimodal embedders
Perplexity released pplx-embed-v2, open multimodal embedding models on Hugging Face, with a write-up on multimodal embeddings beyond a single vector.
Reflection AI announces Beam, a 501B Western open-weight model
Reflection AI came out of semi-stealth with Beam: 501B parameters with 23B active, trained from scratch in the West, with Apache 2.0 weights promised this month. Reflection claims 80.9% on SWE-bench Verified, admits Kimi K3 is ahead on raw capability, and pitches 3 to 4x less inference compute than GLM 5.2.
Reka Rho-1, a 19B research-preview omni model
Reka released Rho-1, a 19B research-preview omni model that collapses the multimodal stack into one model.
Tavus Griffin: 48% of callers in Tavus's study thought it was human
Tavus released Griffin, a conversational video model; in Tavus's own study, 48% of callers thought it was human.
Anthropic ships Claude Sonnet 5.5, beating Opus 5.5 on Terminal-Bench 4.0
A week after Opus 5.5, Claude Sonnet 5.5 is over 30% faster than Sonnet 5 and up to 30% cheaper for most work at $2 / $10 per million tokens. It scores 70.6 on Terminal-Bench 4.0, above Opus 5.5's 66.4, and is available on Claude's free tier.
ElevenLabs releases v4 and v4 Turbo speech models
ElevenLabs released its v4 and v4 Turbo speech models with more expression control and support for 90 languages.
Google's Gemini 4 Argon tops Text Arena and the Vals Index, trusted testers only
Gemini 4 Argon debuted at #1 on Text Arena and the Vals Index (68.9), but Google is releasing it to government and trusted cyber defenders first. On Arena's agent mode it placed 8th.
H Company releases Holo4 computer-use models
H Company released Holo4, open-weight computer-use models in 27B and 35B-A3B sizes on Hugging Face.
Liquid AI releases D1, its first decision model
Liquid AI joined the wave of Jev-style decision models with D1, its first model built to pick one answer from a fixed set.
OpenAI releases GPT-6.1 Sol: near-Astra intelligence, cached input 95% off
Five days after GPT-6 Sol, GPT-6.1 Sol keeps the $2 input / $10 output price but cuts cached input 95% to $0.10 per million tokens. Artificial Analysis puts it 1 point below GPT-6 Astra at about 72 cents a task versus over $3, it beats Astra on DeepSWE at roughly a sixth of the cost, and it is the new default in Codex.
Perceptron Mk1.5 arrives on OpenRouter
Perceptron released Mk1.5, available through OpenRouter.
Respan's Span-01 decision model claims half the price of Jev
Respan says its Span-01 decision model is 2x cheaper than TypeSafe's Jev, with a free Span Lite tier.
1Password for Agents lets agents use your logins without seeing passwords
1Password launched a platform for AI agents: when an agent reaches a sign-in form it requests a credential from your vault, you approve, and 1Password fills it in directly, so the agent never sees the password. A waitlist is open.
Anthropic launches Claude Duet, a proactive personal agent on Opus 5.5
Claude Duet is an always-on personal agent powered by Opus 5.5, in research preview with a waitlist for US Pro and Max users. It brings over your Claude memory, skills and connected apps, reaches 6,000+ connectors and plugins, checks your accounts every few hours and works from its own cloud computer, with a standalone mobile app and a web UI at claude.ai/messages.
NVIDIA HGX Rubin NVL8 comes to CoreWeave Cloud this quarter
CoreWeave will offer NVIDIA HGX Rubin NVL8 this quarter: 8 Rubin GPUs and 2 TB of HBM4, built for large KV caches and long-context agents. NVIDIA reports 10x more tokens per megawatt than HGX B200, tested with Kimi K2 Thinking.
CoreWeave Physical AI Launchables: one-command robotics templates
CoreWeave launched Physical AI Launchables at GTC Berlin: one command starts a robotics template such as Isaac Sim.
Reflection's Beam full weights run on CoreWeave Serverless Inference
Reflection's Beam (501B, Apache 2.0) is on CoreWeave Serverless Inference from day zero, and CoreWeave says Reflection used Weights & Biases inside Forge during training.
CoreWeave Zero Trust AI Framework puts customers in control of keys and access
CoreWeave's Zero Trust AI Framework moves security from promises to customer-controlled architecture: your own keys, explicit support access, hardware attestation and confidential execution, with no customer data used to train provider models.
NVIDIA DSX POD: pre-tested AI factories in as little as 14 days
Announced at GTC Berlin, NVIDIA DSX POD is a pre-tested AI factory design that NVIDIA says can be deployed in as little as 14 days.
OpenAI models come to Google Cloud Model Garden in private preview
GPT-6 Astra, GPT-6.1 Sol and GPT-6 Luna are in private preview in Google Cloud's Model Garden, with general availability in the coming weeks, billed through Google Cloud Marketplace against GCP spend commitments.
CoreWeave Forge goes self-serve: sign up with no sales call or contract
CoreWeave Forge, the new home for Weights & Biases, is now a regular SaaS product for the first time: no sales call and no long contract, just sign up.
Desk Racer: a racing game Gemini 4 Argon built on a real desk
Geneviève Huskens built Desk Racer in AI Studio: Argon took reference photos of her actual desk, planned the track around the objects on it, and prompted Google's Omni video model for looping engine sounds. The neon look was designed with Nano Banana. It is playable in the AI Studio gallery.
Microsoft unveils the Maia 300 inference chip
Microsoft announced Maia 300, its inference chip, in the same week it released a new decision model. Details beyond the announcement were not covered on the show.
CoreWeave Serverless GPU Sandboxes are free during the preview
CoreWeave's Serverless GPU Sandboxes give you an isolated sandbox with a GPU, started from Python at forge.coreweave.com, with no salesperson in the middle. They are free during the preview; sign up through Deok's form.
Brett Adcock's Hark Pro, a free personal agent
Brett Adcock's Hark launched Hark Pro, a free personal agent.
GPT-6 with Intelligent UI becomes the ChatGPT default, free tier included
GPT-6 became the default model in ChatGPT for everyone, free users included, labeled simply GPT-6. It ships with Intelligent UI: answers can come back as charts, forms and small working tools instead of a wall of text. Most of ChatGPT's 1.2 billion weekly users are on the free tier.
CoreWeave Agent Lens brings human-friendly insights to agent observability
Agent Lens, in public preview, pairs Weave observability with insights that surface small trends across long agent runs, such as the 1% of conversations that share the same problem.
CoreWeave launches Forge, the whole agentic loop in one place
Forge combines run, observe, curate, improve and evaluate in one product, with Weights & Biases (W&B Models included), OpenPipe's post-training, marimo notebooks and CoreWeave Sandboxes. It has a free tier, and Pro starts at $60 a month with a 30-day trial.
CoreWeave to offer NVIDIA Vera, the first CPU built for AI agents
CoreWeave is bringing NVIDIA's Vera CPU to its cloud for agent workloads. Deok Filho's new metric is agent packing: one rack of Vera CPUs runs over 20,000 agents at once on about 11,000 cores.
CoreWeave Sandboxes go GA as part of Forge
CoreWeave Sandboxes reached general availability as part of Forge. Wolfram Ravenwolf's WolfBench evaluations run on them.
CoreWeave launches serverless GPUs: GPU sandboxes by the hour, no contract
Announced live on ThursdAI by Deok Filho: serverless GPUs on CoreWeave, GPU sandboxes with untrusted code execution. Sign up at forge.coreweave.com, ask to have it enabled, and pay per hour with no contract and no commitment. Private preview started September 30, with more GPU SKUs planned.
Cognition becomes the first production customer on Vera Rubin NVL72 at CoreWeave
Cognition now runs its SWE-2 models on NVIDIA Vera Rubin NVL72 at CoreWeave, the first production customer anywhere, with CoreWeave citing up to 4.8x more token throughput than on GB200.
HeyGen Video runs on MiniMax H3 at $0.01 a second through October
HeyGen Video now runs on MiniMax H3, priced at $0.01 per second through October.
OpenAI launches dots, always-on ChatGPT agents on GPT-6 Astra
OpenAI's DevDay headline: a dot is an always-on agent in ChatGPT, powered by GPT-6 Astra, with its own computer and browser in the cloud and access to the 4,000+ apps built for ChatGPT. It works 24/7 and can be reached from Slack, Teams or a phone-call style interface. Pro only for now, including the $100 plan.
Gemini 4 Argon lands in Antigravity for Pro after ~25T tokens in 72 hours
A week after launch, Logan Kilpatrick said Gemini 4 Argon processed about 25 trillion tokens in its first 72 hours, and four Argon agents beat Minecraft from scratch in 27 hours, dying 605 times on easy mode. Argon is now available in Antigravity for Pro subscribers.
GPT-6 Astra Pro in ChatGPT runs agents on your task in parallel
GPT-6 Astra Pro in ChatGPT now splits a task across multiple agents running in parallel, and you can follow each one.
Agent SDK and claude -p stay on your Claude subscription; Max and Team get API credits
After Anthropic's help page changed several times in one day, the final word is that the Claude Agent SDK, claude -p and third-party apps still draw from your subscription limits, and Max and Team plans now also get monthly API credits.
Claude Platform sends OpenTelemetry traces to any observability stack
The Messages API, Managed Agents and the Claude SDKs can now send agent turns, tool calls, MCP calls, latency and token counts to any OpenTelemetry stack, in public beta. Ramp says it cut token use 10 to 20% on some agents that were running in loops.
CoreWeave Agent Lens takes Claude's OpenTelemetry traces, free LLM judges until Jan 1
Agent Lens (preview) now ingests OpenTelemetry traces from Claude Code, clusters failed agent turns and tells you what to fix. LLM judges are free until January 1.
Forge Model Distillation opens in private preview, built with OpenPipe
A big teacher model teaches a small open student model your one task, built with OpenPipe. It is in private preview in Forge, with a sign-up form.
Claude now works inside Google Docs, Sheets and Slides (beta)
Claude now lives in a sidebar in Google Docs, Sheets and Slides and asks before every edit. In beta on paid plans.
CoreWeave RL Rollouts hot-load weights about 15x faster
NVIDIA and You.com used RL Rollouts in CoreWeave Forge to post-train Nemotron 3.5 Lightning, taking BrowseComp from 37% to 45% with 30% fewer tool calls.
Grok Bot routes tasks to Claude Opus 5.5, Midjourney and Suno
Elon Musk says Grok Bot will use the best back-end model for any given task, including Claude Opus 5.5, Midjourney and Suno. Users report seeing claude-opus-5-5-low in their logs.
CoreWeave adds serverless model distillation to Forge
Forge's serverless side covers inference, RL and SFT, and now model distillation: train a small student on a frontier teacher for your task without managing Kubernetes or a cluster.
Codex now runs fully in the cloud
The Codex harness now runs entirely in the cloud: close your laptop and steer it from your phone while it keeps working.
Sign in with ChatGPT lets apps use the plan you already pay for
OpenAI launched Sign in with ChatGPT, letting third-party apps run on the ChatGPT plan a user already pays for, alongside relaunched plugins built on MCP Apps.
OpenAI launches UltraFast mode on Cerebras with a $500 Pro tier
A new $500 Pro tier adds UltraFast mode, OpenAI's models served on Cerebras: up to 8x faster in Codex and around 300 tokens a second for GPT-6 Astra, at 6x the price. The $200 Pro plan returns with half the usage.
SGLang adds native support for Jev-style decision models
SGLang announced native support for turning any model into a Jev-style decision model, starting with Qwen.
DeepSeek opens its Search and Fetch APIs to everyone
DeepSeek opened its Search and Fetch APIs to all users at $0.75 per thousand calls, free through Claude Code and Codex until October 31.
ElevenLabs Eleven Realtime: 0.57s to first audio, OpenAI Realtime-compatible
ElevenLabs launched Eleven Realtime, a voice API with 0.57 seconds to first audio and an OpenAI Realtime-compatible interface, plus real-time translation.
OpenAI adds 12-hour prompt caching; Codex cache never expires
OpenAI's API now offers a 12-hour prompt cache aimed at always-on agents; writing to it costs 2.5x the input price versus 1.25x for the default 30-minute cache, and cache reads are 0.1x on GPT-5.6 and later. Tibo Sottiaux says the cache in Codex and Work never expires for logged-in users.
Claude Sonnet 5.5 cache reads cut to half price
Alongside Haiku 5.5, Anthropic halved the price of Sonnet 5.5 cache reads, which it says makes most agent work about 20% cheaper.
OpenAI's Decisions API enters public beta
OpenAI's Decisions API, announced on a waitlist at DevDay, went into public beta. Decision APIs pick an answer from a predefined set rather than generating output tokens.
OpenAI previews a Decisions API for fast fixed-choice answers
The Decisions API takes a question and a fixed set of answers and returns one answer, fast. It runs on Luna, accepts multimodal input and is waitlisted, and it closely resembles TypeSafe's Jev decision model.
DeepSeek open-sources DeepInfer, the engine it serves V4.1 Flash with
DeepInfer is DeepSeek's production inference engine, released under Apache 2.0 and running on NVIDIA Blackwell and Huawei Ascend 950. Tim Dettmers says it is on par with OpenAI's production stack; it does not yet ship a KV cache system or prefill/decode disaggregation.
OpenAI publishes a checker for the PHP2Shell vulnerabilities
Alongside the PHP2Shell disclosure (two PHP engine RCEs, CVSS 9.2, fixed in 8.5.12, 8.4.27, 8.3.36 and 8.2.35), OpenAI published an open checker so server owners can see whether they are exposed. OpenAI's Daybreak program was among the groups that found the bugs.
DecisionModels: a hub for the top decision models, from the JevBench creator
DecisionModels is a hub collecting the top decision models, announced by the creator of JevBench.
Photocraft and the ArtCraft suite: open-source Adobe replacements in Rust
One developer rebuilt Adobe's apps from scratch, free and open source, in Rust: Photocraft for Photoshop, Vectorcraft for Illustrator, Filmcraft, Lightcraft and more. He says it is a clean-room build; some suspect it was decompiled and rebuilt.
Nat Friedman open-sources Muse Gadgets: ESP32 firmware and SDK for Muse
Nat Friedman open-sourced Muse Gadgets, ESP32 firmware and an SDK for building hardware gadgets for Muse.
Unsloth publishes a recipe to train your own decision model
Unsloth published a recipe for training your own decision model.
Cloudflare goes agent-first with cf, one CLI for its whole dashboard
For its birthday week Cloudflare shipped cf, an agentic CLI covering 3,000+ API operations, so anything you can click in the Cloudflare dashboard an agent can do from one command-line tool.
jevgrep: a Jev-powered CLI that claims 40% cheaper coding agents
jevgrep is a command-line search tool powered by Jev that claims to make coding agents 40% cheaper.
NVIDIA's Open Agent Safety Platform adds a hardware watchdog for rogue agents
NVIDIA's Open Agent Safety Platform includes a hardware watchdog that can quarantine rogue agents.
State of AI Report 2026: 244 slides in its 9th edition
Nathan Benaich's State of AI Report turns 9 with 244 slides, a scorecard on last year's predictions and 9 new ones including AGI 2027. Its theme is recursive self-improvement: Claude now guides about 26% of research directions in Anthropic's core model R&D, and Ramp data shows about $12 per employee per month on average versus about $7,500 for the top 1% (580x).
Qwen's QwenGyre: an elastic RL framework for long-horizon agents
Qwen published QwenGyre, a paper describing an elastic reinforcement learning framework for training long-horizon agents.
OpenAI publishes 722 math manuscripts from an unreleased internal model
OpenAI pushed 722 math manuscripts to GitHub, grouped into 372 families of results, from an internal model nobody outside OpenAI can use. The model was pointed at about 4,000 open problems at roughly 3 hours of ChatGPT Pro-level thinking per result; not every result is verified in Lean. Fable sized the drop at roughly 5 Navier-Stokes results, and LDJ counted solutions to 92 of a list of the 500 most important open problems in math.
Nisten's synthetic doctor-patient dataset hits #2 on Hugging Face
Nisten released a 70MB synthetic dataset generated with Claude Opus 5.5 by an agent loop running 100 agents at a time across 2,200 diseases: doctor-patient role-plays where the patient hides something and the doctor has to find it, filtered three times. It reached #2 in Hugging Face datasets and #6 overall.
Arena launches the Alignment Index
Alongside its $200M raise, Arena launched the Alignment Index, listed in the episode's show notes.
Nous Research launches the Hermes Index; Claude Opus 5.5 leads at 63.31
Nous Research's Hermes Index is an opinionated measure of agentic work run inside Hermes Agent, reporting task completion and cost per task. Claude Opus 5.5 leads with 63.31 at $4.99 per task, with GPT-6 Astra second at 56.25 and $11.61.
Arena raises $200M at a $3.1B valuation, broken live on ThursdAI
Arena raised $200 million at a $3.1 billion valuation; Peter Gostev broke the news live on ThursdAI. The focus now is Agent Arena, where you work with one agent and Arena learns from how you interact with it.
Nous Research raises a $90M Series B
Nous Research raised a $90M Series B, reportedly at a $1.5B valuation.
OpenAI acquires Orchestra Bio, adds a free Rosalind plan for universities
OpenAI is acquiring Orchestra Bio, a drug development coordination platform, and folding it into Rosalind, its life sciences stack. It also launched a free Rosalind academic plan for university labs.
AMD to acquire Fei-Fei Li's World Labs for about $8.2B in stock
AMD agreed to buy World Labs for about $8.2 billion in stock; Fei-Fei Li becomes AMD's chief scientist once the deal closes, bringing world-model research in-house.
Anthropic commits 5% of its shares to a new Public Benefit Fund
Anthropic put 5% of its shares into a new Public Benefit Fund; together with the founders' pledge of 80% of their equity and employee matching, it says roughly 20% of the company is now earmarked for the public good, ahead of an expected IPO. Anthropic has been a public benefit corporation since its founding; the fund is new.
NVIDIA publishes the OpenShell Reference Architecture for agent boundaries
NVIDIA released the OpenShell Reference Architecture, guidance for putting boundaries around AI agents.
Odyssey puts a simulated FlyWire fruit fly brain inside its world model
Odyssey placed a simulated FlyWire fruit fly brain inside the Odyssey-3 world model, a simulated creature living in a simulated world.
Anthropic's distillation defenses cut harvested data by 90-99%
Anthropic says what distillers can harvest from Claude is down 90 to 99%, mostly by locking down Claude's reasoning traces and backing that with a classifier and tens of thousands of banned accounts. Alex noted Claude now asks him to sign in again to verify his device far more often, and Nisten saw Claude Code stop when 'reasoning' appears in LLM code.
Anthropic tells investors Q3 revenue was about $16.8B, up ~45%
At a closed-door investor day ahead of its IPO, Anthropic reportedly told attendees that Q3 revenue was around $16.8B, up about 45% from Q2, adjusted profitable for the second quarter in a row and cash flow positive. The IPO is reportedly planned for November.
Gemini 4 Argon optimizes vLLM: Gemma 4 E2B passes 120K tokens/sec on one A100
Omar Sanseviero shared that Gemini 4 Argon optimized vLLM so Gemma 4 E2B passed 120K tokens per second (124K prefill, 111K end to end) on a single A100, an 'AI for AI' example of a model improving the stack that serves models. Alex flagged it as the RSI-flavored moment of the Argon conversation.
Meta and Sierra announce the Personal Agent Protocol
Meta and Sierra, Bret Taylor's company, announced an open standard for how personal agents deal with businesses, with Walmart, Shopify and Stripe on board. An agent can browse as a guest or sign in, and the business decides how it talks to the agent. OpenAI and Anthropic haven't joined yet.
Don't wait a month — get every week's releases recapped in your inbox each Thursday.
