Release Cadence & Lab Activity
Calculates cadence variances, elapsed release windows, and cycle pressure across frontier labs. Estimates use recorded release intervals, or a category preset when history is missing. They do not predict release probabilities or confirmed schedules.
Catalog Cadence Analysis for Imminent and Approaching Model Horizons
This analysis reviews uncalibrated cadence estimates across eight models currently in their imminent release horizon based on historical interval data. Sample sizes range from a single observed interval to four intervals, introducing significant uncertainty into the projections. These figures represent historical pacing comparisons rather than confirmed release probabilities or schedules.
AI briefings refresh at most once every 24 hours. Catalog estimates remain available when AI is unavailable or its daily allowance is used.
Udio v2 v2
Based on 0 observed intervals using a genre-baseline basis, with 342 days elapsed versus a 90-day average cadence, limiting the reliability of the projection.
Seedream 5 Pro 5 Pro
Derived from 1 observed interval using historical basis, showing 97 days elapsed against a 29-day average cadence with limited historical depth.
LTX 2.5 2.5
Calculated from 1 observed interval using historical basis, featuring 92 days elapsed against a 26-day average cadence.
Wan 2.2 2.2
Reflects 1 observed interval on a historical basis, with 471 days elapsed compared to a 153-day average cadence.
Fish Audio Drama 3.0 Drama 3.0
Based on 1 observed interval with a historical basis, showing 26 days elapsed against an 11-day average cadence.
Gemini Argon 4 Argon 4
Uses 1 observed interval on a historical basis, with 29 days elapsed versus a 13-day average cadence.
Qwen 0902 0902
Drawn from 4 observed intervals using a historical basis, with 69 days elapsed against a 32-day average cadence.
Muse Spark 1.3 Spark 1.3
Involves 3 observed intervals on a historical basis, displaying 70 days elapsed compared to a 37-day average cadence.
Udio
Musician-favoured output, especially on instrumental prompt-following.
Seedream
Region-precise image editing that changes one element and leaves the frame.
LTX
Fast video generation with audio, in Pro and Fast tiers.
Wan
Open-weight video foundation model supporting 1080P text-to-video and image-to-video.
Fish Audio Drama
35B MoE emotive speech model with 16-character multi-speaker synthesis.
Gemini Argon
Google DeepMinds flagship frontier reasoning engine with native audio-visual stream processing.
Qwen
Open-weight family shipping at a cadence nobody else matches.
Muse Spark
Sold on computer use and long-horizon task completion.
Reve
Prompt-driven image editing with strong layout and text handling.
KAT-Coder
Agentic coding model meant to take a whole issue end to end.
Hy3
295B mixture-of-experts built for reasoning and agentic workflows.
MAI Image
Microsoft's in-house image model, aimed at design-ready output.
Tripo
Turns a single photo into a 3D object mesh.
Gen-4
Pinnacle cinematic video generation engine with custom motion brushes and depth control.
PaddleOCR-VL
Compact 0.96B Apache-2.0 vision-language document parser for multilingual text, tables, formulas, and charts.
Sora
Still the most physically convincing motion. Removed from the API on 24 September 2026.
Qwen-Image
7.1B open-weight text-to-image and image editing foundation model from the Qwen team.
Sonilo
Text-to-music with licensing cleared for commercial use.
Hi3D
Single-image to 3D model generation.
Gemini Flash
The cheap, fast tier of Gemini.
Ray
Luma's video model, directed frame by frame with up to 16 keyframes.
GLM
753B open-weight mixture-of-experts reasoning and coding family from Z.ai.
Nano Banana
Google's image model — shipped under two names, one on each platform.
DeepSeek R1
Frontier open-weights reasoning model with emergent self-verification reasoning chains.
GPT Astra
Frontier model with native operating system computer-use and autonomous task execution.
Genie 2
Generative interactive 3D world model creating playable environments from text and images.
Laguna S
Purpose-built for software work rather than general chat.
Seedance
Long-form image-to-video. Currently top of the public leaderboards.
GPT-5.6
Shipped as a three-model lineup rather than one flagship.
Recraft
State-of-the-art graphic design, branding, and typography generative foundation model.
Ideogram
Image model whose distinguishing trait is that it renders text correctly.
Muse Wonder
Meta's frontier models for personal and professional agent work.
Claude Sonnet
The mid-tier Claude, sized for volume work rather than peak reasoning.
Claude Opus
Frontier reasoning model, tuned hard for long agentic coding runs.
Sonic 2
Ultra-low latency streaming voice engine with 85ms TTFB for conversational agents.
Claude Code
Autonomous agentic command-line software engineering and multi-repo refactoring system.
Hunyuan3D
Open-weights high-resolution 3D asset generation from single images and text prompts.
Hailuo
Video model taking image, video and text references at once.
Grok
Claims frontier coding at markedly lower token consumption.
Ling Flash
Small, fast open MoE aimed at cheap high-throughput serving.
FLUX
The open-weight image default, now expanding past images entirely.
Claude Haiku
Ultra-fast frontier efficiency tier with 1M token context window and adaptive thinking depth.
DeepSeek V4.1
Frontier Pareto-efficient flash model with integrated live web retrieval APIs.
ElevenMusic
The licence-clean option. Trained on licensed data by design.
Voxtral
Mistral's speech family: transcription, text-to-speech and a voice-agent SDK.
MiniMax M
Hybrid Lightning Attention multimodal mixture-of-experts with ultra-long context.
Veo
Native 4K with synchronised audio generated in a single pass.
Midjourney
Still the reference point for artistic style over literal accuracy.
Kimi
Kimi flagship for long-context coding, creation, and agentic work.
Suno
The consumer leader on audio fidelity. Licensing still unresolved.
Kling
The value pick — four entries in the current video top ten.
Matrix-Game
Open-weight interactive world foundation model for controllable game and environment simulation.
Unlimited-OCR
3.3B MIT-licensed open-weight vision-language OCR model for dense document, table, and formula recognition.
ElevenLabs TTS
The default for expressive speech synthesis and dubbing.
LongCat
Very sparse MoE aimed at coding and repository-scale work.
Lucy
Real-time video editing over WebRTC — restyle a live feed as it runs.
Oasis
First interactive, real-time playable Minecraft world model generated end-to-end by neural diffusion.
Claude Fable
Anthropic Mythos-tier frontier model tuned for long-horizon autonomous workflows.
Mistral Large
Mistral's natively multimodal flagship, served from European datacenters.
Moshi
First real-time native speech-to-speech multimodal dialogue model with 160ms latency.
Stable Audio
Latent diffusion audio model for full-length tracks with coherent musical structure.
AlphaFold
Revolutionary biomolecular model predicting structures of proteins, DNA, RNA, and ligands.
Big Sleep
Autonomous vulnerability research agent finding zero-day exploits in real-world software.
Claude Cyber Shield
Autonomous cyber defense initiative for code auditing, software integrity, and vulnerability triage.