Today’s updates · 2026-10-22

New releases

Newly listed modelsPrices →

• MiniMax-M3.1-FlashMiniMax · 1M · 10-19¥1/4 Official
• DeepSeek V4.1 Flash 1017DeepSeek · 1.05M · 10-17$0.07/0.5 OpenRouter
• Beam 1.0 501B A23BReflection · 1.05M · 10-16$0.25/1 OpenRouter
• Muse Spark OpenMeta · 1M · 10-16$1.25/4.25 OpenRouter
• Moondream 3 PreviewMoondream · 32K · 10-16$0.3/2.5 OpenRouter
• Mercury 3Inception · 1M · 10-16$0.1/0.35 OpenRouter
• Mercury 3Inception · 1M · 10-16$0.1/0.35 Official
• Nemotron 3.5 Super VLNVIDIA · 262K · 10-15$0.1/0.4 OpenRouter
• Gemini 4 ArgonGoogle · 1.05M · 10-13$2/10 Official
• ExtractScaleDown · 1M · 10-13$0.04/0 OpenRouter Unverified
• SummarizeScaleDown · 1M · 10-13$0.04/0 OpenRouter Unverified
• Muse WonderMeta · 1.05M · 10-12$4/20 OpenRouter Unverified
• Step-5-PreviewStepFun (阶跃星辰) · 1M · 10-08¥7/20 Official
• Claude Haiku 5.5Anthropic · 1M · 10-07$0.1/0.5 Official
• Mistral Large 4Mistral AI · 1.05M · 10-06$0.68/2.09 Official
• GPT-6.1 SolOpenAI · 1.05M · 09-29$2/10 Official
• Claude Sonnet 5.5Anthropic · 1M · 09-28$2/10 Official
• Qwen3.8-Max-PrimeAlibaba (Qwen) · 1M · 09-23¥24/72 Official
• GPT-6 SolOpenAI · 1.05M · 09-22$2/10 Official
• GPT-6 LunaOpenAI · 1.05M · 09-22$0.1/0.5 Official
• Claude Opus 5.5Anthropic · 1M · 09-22$4/20 Official
• Command A+Cohere · 192K · 09-22$0.3/1.5 OpenRouter Unverified

Dates are often OpenRouter listing dates rather than official release dates.

Latest newsAll →

New milestones · AI × Disciplinesdata as of 2026-10-22All →

• Pinal: a 16B-parameter foundation model that designs functional proteins from natural-language descriptions, validated in four experimental tasks (Nature)Life sciences (structural & molecular biology) · 10-21 · Nature · source
• hEVO: a human genomic foundation model trained on 4.784B haplotype sequences from 63,809 genomes, ranking first in 18 of 19 primary comparisons (bioRxiv)Genomics & genetics · 10-21 · bioRxiv · source
• Prix Pascal de Littérature IA launches for French-language fiction co-written with AILiterature · 10-21 · The Guardian · source
• AI-assisted proof of the double-cap conjecture, with a separate Lean 4 proof formalised almost entirely using AI tools (arXiv 2610.27523)Mathematics · 10-20 · arXiv · source
• Radical Numerics: Omnii-1-designed phages administered to a patient under compassionate usePharmacy & drug discovery · 10-20 · Radical Numerics · source
• Sophont launches Holon-S1, a pan-cancer pathology model predicting spatial gene expression from H&E (research use)Medical imaging & pathology · 10-20 · PR Newswire · source
• Biodiversity foundation model transfers to field moth monitoring of 2,727 species (Sci. Rep.)Ecology & biodiversity · 10-20 · Scientific Reports · source
• KAIST GeoCore-9B: 9B-parameter geo-aware generative EO foundation model (NeurIPS 2026)Geography & remote sensing · 10-20 · KAIST / EurekAlert! · source
• Spotify: 50M+ AI slop/spam tracks removed from recommendations this year, 75M+ removed outright; AI Personas excluded from recommendations by defaultMusic · 10-20 · Spotify Newsroom · source
• OpenAI analysis: ChatGPT receives ~70M religion-related requests a week, 45% about Islam and 34% about ChristianityReligious studies · 10-20 · The Christian Post · source
• MR-MAS (Bioinformatics): an LLM multi-agent system that runs end-to-end Mendelian randomization — data retrieval, causal estimation, interpretation and write-upStatistics · 10-20 · Bioinformatics · source
• Petit Palais opens “Visionaries”: Google Arts & Culture uses Gemini-based tools to reconstruct the lost studios of 19th-century women sculptors from archival photos and lettersVisual arts & design · 10-20 · Google Blog · source
• Quanta: number theorists largely accept OpenAI’s proof of a quasi-Riemann hypothesis (zeros between 1/12 and 11/12)Mathematics · 10-19 · Quanta Magazine · source
• CMA’s Fengqing AI global ensemble forecast system V1.0 passes operational review and goes live (11-day skill)Meteorology & climate science · 10-19 · 中国新闻网 · source
• Seth, Block & Tononi in Nature: three reasons why AI is not conscious and perhaps never will bePhilosophy · 10-19 · Nature · source
• Appeals court upholds Thomson Reuters’ win over Ross Intelligence on AI-training fair use (opinion sealed)Law · 10-18 · The Verge · source
• Large RCT of 8,647 rural Chinese pupils: AI-personalised learning no significant maths gainEducation · 10-17 · Economics of Education Review · source

New papersdata as of 2026-10-21All →

• Has GPT-6-Astra Achieved Embodied Intelligence?Evaluates GPT-6-Astra on 12 robotics benchmarks and discusses what a general model's embodied competence means for the field.10-15
• OpenGCUA: Toward a Fully Open Stack for Generalist Computer UseOpenGCUA aims to open the full stack for generalist computer-use agents, with a first release spanning GUI and code execution.10-15
• FDE-Gym: The Last Mile of AI DeliveryFDE-Gym offers 95 real delivery tasks testing whether agents can clarify vague requirements and ship working solutions like forward-deployed engineers.10-15
• PremiseCheck: Efficient Fact-Checking of Human and AI Peer ReviewsReviewFact labels factual errors in peer reviews; PremiseCheck combines an LLM with cheap decision models to verify them efficiently.10-15
• MJev-Spatial: System One Spatial Intelligence via Native Attention MatchingMJev-Spatial casts spatial questions as classification over candidates, answering relation, distance and point queries faster than text generation.10-15
• HexaShell: A Thin Harness for Exploring the Limits of VLM Agents as Robot Coding AgentsHexaShell keeps the harness thin so a coding agent drives a two-arm robot via a command line and writes its own tools after failures.10-16
• MOSS-Image Technical ReportMOSS-Image is an open 5B text-to-image system strong at bilingual long-text rendering and layouts, with intent-aware prompt enhancement.10-16
• MemT-1.0: Towards Memory-Native EvolutionMemT-1.0 makes memory intrinsic to the model using Gated DeltaNet-2 and Mixture-of-Storage to scale native memory.10-19
• OpenVLHarness: Real World Visual Agentic Tool-Use with Multimodal MemoryOpenVLHarness gives MLLMs a tool capability layer and persistent multimodal memory for visual tasks needing external evidence.10-19
• OR-License: Evaluating LLM Agents as Real-World Operations Research PractitionersOR-License benchmarks whether LLM agents can turn original client materials from 31 paid operations-research engagements into expert-grade solutions.10-16
• Waypoint-OPD: Transferring RL-Induced Policy Shifts via Gated On-Policy DistillationWaypoint-OPD uses an intermediate teacher checkpoint to transfer only consistent RL-induced policy shifts to the student.10-15
• Focused Omni-Modal GUI AgentsFocusGUI is a lightweight adapter that picks the step-relevant omni-modal evidence for a frozen GUI agent, curbing compounding distraction.10-19
• Ropedia W0: Scaling Loco-Manipulation World Models across Embodiments with Human ExperienceRopedia W0 is an egocentric loco-manipulation world foundation model spanning humans and robots, trained on 10k hours of human and 3k hours of robot data.10-19
• 3DCodeVerse: Exploring 3D Worlds as Code3DCodeVerse offers 2M executable 3D programs and a coding harness so agents can create editable, verifiable 3D content.10-15
• EgoXperience: Learning World-Grounded Egocentric Representations from Human ExperienceEgoXperience jointly reconstructs observer motion, scene geometry and hand-object interaction from egocentric video, with an auto-annotation engine and benchmark.10-19
• Sage-A1.5: From Long-Horizon Execution to Scientific Exploration with Capability CompositionSage-A1.5 is a 35B MoE agentic model adding subagent orchestration, self-verification and open-ended scientific exploration.10-20
• Holi-World: Building Holistic World Intelligence from Human-Ego ExperienceHoli-World is an ecosystem for learning world intelligence from human egocentric experience, from capture to evaluation.10-19
• Scaling Parameter-Efficient Fine-Tuning for Brain EncodingUses parameter-efficient fine-tuning to adapt multimodal backbones of up to 104B for brain encoding, better predicting neural responses.10-19
• Behavioral Compositionality of Sparse Autoencoder RepresentationsIntroduces SAR to measure whether sparse-autoencoder features compose additively in downstream behaviour, aiding interpretability.10-15
• SolarRenderer: Few-Step Generative Reference RenderingSolarRenderer adds two LoRAs to a frozen video model to render coarse structural video into realistic footage in four steps, including real-time streaming.10-20
• ScaleNav: Scalable Internet Video Pretraining for Language-Conditioned Local NavigationScaleNav learns goal-directed visual dynamics from internet egocentric videos for language-conditioned robot navigation.10-19
• Speculative Forcing for Efficient Autoregressive Video GenerationSpeculative Forcing treats intermediate estimates as drafts and accepts converged tokens early to speed up autoregressive video generation.10-15
• Matrix-Game 4.0: Harnessing Video Models for Real-Time Playable Game WorldsMatrix-Game 4.0 separates game logic (engine plus coding agent) from visuals (autoregressive video renderer), reaching playable 60 FPS on one RTX 5090.10-20
• S-Agent 2.0: Spatial Reasoning with a General Coding AgentS-Agent 2.0 equips a general coding agent with command-line 3D perception tools and scene memory, lifting spatial-reasoning benchmarks.10-20
• Learning to Think and Explore in 3D WorldTrains VLMs with RL to reason and actively move their viewpoint in 3D, seeing one image at a time like a human.10-15
• Self-Evolving Coding Agents via Harness-Policy OptimizationMicrosoft's Self-HPO alternately evolves a coding agent's harness and policy model for recursive self-improvement.10-16
• Mosaic: Multi-Session Agentic Reinforcement LearningMosaic generalises agentic RL to multi-session trajectories, matching inference-time context compaction and sub-agent forking.10-15
• Scene-Designer: Learning Design Intelligence for Agentic 3D Scene GenerationScene-Designer learns design intelligence from design and critique experience to generate coherent 3D indoor scenes from text.10-16
• PAR: Progress-Aligned Recurrence for Length Generalization in State TrackingPAR aligns each loop iteration of a looped LM with task progress and anchors state norms, extrapolating state tracking to 16x training length.10-20
• VLA-$Δ$: Learning Physical State Transition for Vision-Language-Action ModelsVLA-Δ explicitly supervises physical state transitions to reduce VLA overfitting to observation-action pairs.10-15