New AI Models — September 2026
Every model release from our AI news in September 2026, newest first.
Model changes only: new models, open weights, API access, first platform listings and leaderboard ranks. No product features, pricing or single benchmark scores.
85 releasesLast updated: September 30, 2026
- Mercury VoiceInceptionNew modelInception's Mercury Voice diffusion model posts a 320ms median time to first token
- d1Liquid AINew modelLiquid AI's d1 returns probabilities with zero output tokens
- Kumo TabularNVIDIAOpen weightsNVIDIA opens Kumo Tabular for training-free table prediction
- Qwen-Audio-3.1-RealtimeAlibabaNew modelQwen-Audio-3.1-Realtime launches with an ~85% price cutQwen timeline →
- Claude Sonnet 5.5AnthropicNew modelClaude Sonnet 5.5 lands with up to 30% lower per-task costClaude timeline →
- Qwen3-TTSAlibabaPlatformQwen3-TTS on SageMaker, streaming speech before generation endsQwen timeline →
- Qwen-Image-2512AlibabaOpen weightsQwen-Image-2512 Released, Billed as Top Open Image ModelQwen timeline →
- Holo4H CompanyOpen weightsH Ships Holo4 in 27B Dense and 35B-A3B MoE
- Holotron4 NanoH CompanyNew modelH Ships Holo4 in 27B Dense and 35B-A3B MoE
- Ember-1FireworksNew modelEmber-1 trims Kimi K3 reasoning tokens by about 40%
- Space BunnyUndisclosedLeaderboardAnonymous model Space Bunny tops OpenRouter's daily call chart
- FermiQLLM 1.0费米宇宙New modelFermiQLLM 1.0 adds quantum methods to a Qwen base
- Saaras V4SarvamNew modelSarvam's Saaras V4 covers all 22 official Indian languages
- MiniMax M3.1-Flash-PreviewMiniMaxNew modelMiniMax ships M3.1-Flash-Preview on MiniMax CodeMiniMax timeline →
- LFM2.5-VL-3B-DSparkLiquid AIOpen weightsLiquid releases a VLM drafter with up to 3.13x decoding
- AWETARS RoboticsNew model它石智航 (Tashi Zhihang) details an AWE embodied base model and three hardware lines
- Opus 5.5AnthropicNew modelOpus 5.5 lands at $20 per million output tokens
- GPT-6 LunaOpenAINew modelGPT-6 Sol and Luna arrive at half the previous token priceOpenAI timeline →
- Claude Opus 5.5AnthropicPlatformClaude Opus 5.5 lands on Amazon Bedrock at a claimed 40% lower costClaude timeline →
- MiMo-V2.6-ProXiaomiLeaderboardMiMo-V2.6-Pro tops open model rankings at 46 points
- Qwen-Image-2.1AlibabaOpen weightsQwen-Image-2.1 outputs transparent images natively, commercial use needs a separate licenseQwen timeline →
- Qwen3.8-LiveTranslateAlibabaNew modelQwen3.8-LiveTranslate cuts average interpretation lag to 2.3 secondsQwen timeline →
- Step 5 PreviewStepFunLeaderboardStepFun's Step 5 Preview enters the open-weight top three at one-eighth of Opus 5's task cost
- Qwen3.8-Omni-FlashAlibabaNew modelQwen3.8-Omni-Flash prices inputs at a fifth of Gemini FlashQwen timeline →
- Xing4.0-29B-A4BChina TelecomOpen weightsChina Telecom open-sources Xing4.0-29B-A4B
- jina-ocr-v1Jina AIOpen weightsjina-ocr-v1 open weights: 3.4B total, 570M active per token
- Grok Voice Transcribe 2.0xAINew modelGrok Voice Transcribe 2.0 ships, Loom transcribes every video with itGrok timeline →
- LimiX-2Tsinghua & Stable AINew modelTsinghua and Stable AI release LimiX-2 for structured data
- ZDTaichu5.0-9BZidong TaichuOpen weightsZDTaichu5.0-9B opens weights, claims 8 firsts on spatial tests
- JevTypeSafe AINew modelTypeSafe exits stealth with $40M and Jev, a model that outputs decisions
- Mistral Small 4Mistral AIPlatformFirefox's AI assistant now runs on Mistral Small 4
- KoaSalesforceNew modelSalesforce's first CRM reasoning model is a post-trained Nemotron 3 Super
- Doubao 2.1 ProByteDanceNew modelDoubao 2.1 Pro ships an 0915 build, wired into Doubao Work
- GeminiGooglePlatformApple's rebuilt Siri runs on Google's Gemini, but not in the EUGemini timeline →
- StepAudio 3StepFunNew modelStepFun ships five StepAudio 3 models, claiming top rankings
- OM-1Reward AINew modelReward AI releases OM-1, trained only on human glove demonstrations
- Claude Fable 5.1 MAnthropicLeaderboardArena's two leaderboards have different winnersClaude timeline →
- Kimi K2.8 PreviewMoonshot AIPlatformKimi K2.8 Preview is fully rolled out to Kimi Code and WorkKimi timeline →
- GPT-6 AstraOpenAILeaderboardGPT-6 Astra tops two agent benchmarks and sweeps all five drone tasksOpenAI timeline →
- SWE-2CognitionNew modelCognition releases SWE-2, post-trained from Kimi K3 at 64% lower cost
- Fugu MaxSakana AINew modelSakana AI launches Fugu Max and Fugu Ultra v2 for learned multi-model orchestration
- Fugu Ultra v2Sakana AINew modelSakana AI launches Fugu Max and Fugu Ultra v2 for learned multi-model orchestration
- Marengo Embed 3.0TwelveLabsPlatformBedrock Knowledge Bases adds Marengo Embed 3.0 for semantic video search
- DeepSeek-V4.1-FlashDeepSeekOpen weightsDeepSeek-V4.1-Flash reaches GA with an 890-byte-per-token KV cacheDeepSeek timeline →
- Qwen 2.4TAlibabaOpen weightsQwen opens a 2.4T model to self-hosting with 95B active paramsQwen timeline →
- Suno v6SunoNew modelSuno rebuilds its training data and ships v6 in three tiers
- PatchTST-FM-r2IBMNew modelIBM ships a new Granite time-series foundation model
- V4.1 FlashDeepSeekNew modelV4.1 Flash enters a time-boxed internal beta on a natively multimodal architecture
- Ling-3.0-flash-FinAnt GroupOpen weightsAnt's Ling team open-sources a finance-tuned model and a matching research benchmark
- GPT Image 2.5OpenAINew modelGPT Image 2.5 ships, with Flare and Sunburst splitting the APIOpenAI timeline →
- Ling-3.0-flash-VLAnt GroupOpen weightsAnt open-sources Ling-3.0-flash-VL: 124B total, 5.5B active
- Mercury 2.5InceptionNew modelMercury 2.5 lands on OpenRouter at a claimed 1,107 tokens per second
- Mercury 2.5InceptionPlatformMercury 2.5 lands on OpenRouter at a claimed 1,107 tokens per second
- NeoHorse-1基元律动New modelWang Yunhe's team ships NeoHorse-1 in 4B and 9B sizes (Chinese-language source)
- N2.5Nex-AGINew modelNex-AGI ships N2.5 agent family, Pro weights still pending
- Qwen-Drive 1.0AlibabaNew modelQwen-Drive 1.0 merges driving system and cockpit assistantQwen timeline →
- Hy4 previewTencentNew modelTencent Hy4 preview update cuts token use in agent tasks
- HUMAIN-M3HUMAINNew modelHUMAIN-M3 builds Saudi Arabic model on MiniMax M3
- Lyria 3.5GoogleNew modelLyria 3.5 lands in Gemini with multi-minute songs
- WeatherNext 3GoogleNew modelWeatherNext 3 adds hourly refresh and ships into Search and Gemini
- NeoMMENeoMMEOpen weightsNeoMME releases 260M and 800M multimodal multilingual encoders
- WorldGenDeemosNew modelHyper3D's WorldGen builds editable 3D scenes, not single assets
- Gemini 3.8 FlashGoogleNew modelGemini 3.8 Flash arrives: third Flash in six weeks, same price as 3.7Gemini timeline →
- Gemini 3.8 Flash CyberGoogleNew modelGemini 3.8 Flash Cyber opens to defenders via the Fairwind ProgramGemini timeline →
- Muse Spark 1.3MetaNew modelMeta releases Muse Spark 1.3 for long-running agents and coding (Chinese-language source)
- Quasar 438BMultiverse ComputingNew modelMultiverse's Quasar 438B tops European models on the Artificial Analysis index (Chinese-language source)
- Quasar 438BMultiverse ComputingLeaderboardMultiverse's Quasar 438B tops European models on the Artificial Analysis index (Chinese-language source)
- Claude Fable 5.1AnthropicPlatformClaude Fable 5.1 lands on AWS as a "Covered Model"Claude timeline →
- OrbisViskoNew modelVisko launches Orbis, a "Live Model" for continuously running video
- ZevaTsinghua AIRNew modelTsinghua AIR's self-evolving Zeva lifts success rate from 26% to 73%
- DeepSeek-V4-Flash-Vision-ExpDeepSeekOpen weightsDeepSeek open-sources first V4 vision model, 305B under MITDeepSeek timeline →
- GPT-5.6 SolOpenAIPlatformCerebras serves GPT-5.6 Sol at 750 tokens/second, unchanged weightsOpenAI timeline →
- TimesFM-3GoogleNew modelGoogle releases TimesFM-3: 330M-parameter zero-shot multivariate forecaster