New on Ropewalk: 24 Models You Missed (Sept 2026)

By Ropewalk Team. Tested on 2026-09-03 against the live model catalog and the platform's own release log, covering everything that shipped between 2026-08-13 and 2026-09-03.

The Quick Answer

24 new models went live on Ropewalk across four weeks (Aug 13 – Sep 3, 2026) — 12 text models, 6 image, 3 video, and 3 audio. The headliner is Moonshot AI's Kimi K3, a 2.8T-parameter MoE flagship with a 1M-token context window. The other big story: OpenRouter arrived as a new routing provider in week 36, debuting Z.ai's GLM-5.3 and re-routing four already-live models onto cheaper per-token pricing overnight.

Four weekly "new models" posts stalled behind other approvals this cycle, so here's the single catch-up digest — grouped by the week each model actually went public, cheapest first inside each section.

Week 33 (Aug 13, 2026) — Text and Audio Take the Lead

Six models went live on 2026-08-13 alone — the single busiest release day of the four weeks covered here. OpenAI split its GPT-5.6 line into three public tiers by adding Terra and Luna alongside the existing flagship; Meta's first post-Llama open-weights model landed via Together AI; and the entire ElevenLabs provider came back online after a gap the team's own tracking had incorrectly marked as closed.

Kimi K3 — Moonshot AI's current flagship: 2.8T-parameter MoE (16-of-896 experts active), Kimi Delta Attention, native vision, 1M-token context. From 240 gems/generation.

GPT-5.6 Terra — OpenAI's balanced mid-tier, slotted between the Sol flagship and the budget Luna tier, 1M-token context. From 195 gems.

GPT-5.6 Luna — OpenAI's budget tier in the same family, 1.05M-token context, priced for high-volume use. From just 20 gems.

Meta Muse Glimmer 30B — Apache-2.0, open-weight, 30B dense agentic model built for long tool-call sequences — Meta's first release since retiring the Llama line. From 20 gems.

ElevenLabs Music — text-to-music generation: full tracks or instrumentals across orchestral, electronic, cinematic and more, 5–180 second duration control. From 320 gems.

ElevenLabs v3 (TTS) — the provider's most expressive text-to-speech tier yet, 29+ languages, voice cloning and emotional delivery control. From 300 gems.

Week 34 (Aug 20, 2026) — One Model, One Big Gap Closed

Just one release this week, but it closed a catalog gap the team had flagged as a priority: the entire Z.ai/GLM family was missing from Ropewalk until GLM-5.2 landed on 2026-08-20. It wouldn't stay alone for long — two more GLM tiers (5.3-Flash, then the 5.3 flagship) followed within the next two weeks, and 5.2 itself got re-routed onto cheaper infrastructure the moment OpenRouter arrived in week 36.

GLM-5.2 — Z.ai's open-weights flagship, beating GPT-5.5 on multiple long-horizon coding benchmarks per independent reporting, 1M-token context with 131K output tokens. From 50 gems.

Week 35 (Aug 27, 2026) — Video, Image and Audio Catch Up

Eight models shipped on 2026-08-27, the second-busiest day of the four weeks — almost entirely visual and audio work, closing gaps in video (Happy Horse, Seedance), image (Luma Uni-1, Qwen Image, Meta's own text tier hadn't landed yet), music, and a faster GLM text tier.

Happy Horse 1.0 — Alibaba's text-to-video and image-to-video model, 720p/1080p, 3–15 second clips across five aspect ratios. From 1,685 gems.

Happy Horse 1.1 — adds multi-reference-image video generation on top of 1.0's text/image-to-video, same 720p/1080p and 3–15 second range. From 1,685 gems.

MiniMax Music 2.6 — full songs or instrumentals from a text prompt, with optional auto-generated lyrics. From 600 gems.

Luma Uni-1 — Luma's autoregressive image model, matching or beating GPT Image and Nano Banana on quality at a lower cost, 2K resolution. From 175 gems.

Luma Uni-1 Max — a higher-quality tier above Uni-1, same 2K resolution. From 415 gems.

Seedance 2.5 — ByteDance's flagship multimodal video model, native audio generation and native 30-second clips from large reference sets. From 4,600 gems.

Qwen Image 3.0 Pro — dense, accurate text rendering inside generated images plus complex multi-element layouts and photographic detail. From 160 gems.

Pika 2.2 — fast, stylized text-to-video at 720p/1080p. From 800 gems.

Gemini 3.7 Flash — Google's coding and agent-focused Flash tier, 1.05M-token context, generally available since 2026-08-13. From just 15 gems.

GLM-5.3-Flash — Zhipu AI's 320B-A18B natively multimodal MoE, 1M-token context, superseding the GLM-5.2 base tier. From just 5 gems.

Week 36 (Sep 2–3, 2026) — OpenRouter Arrives

The newest batch, and the one with the biggest infrastructure story: OpenRouter became a live routing provider on Ropewalk this week. Z.ai's flagship GLM-5.3 debuted through it directly, and on 2026-09-02 the team re-routed three already-live models — GLM-5.2, GLM-5.3-Flash, and Meta Muse Glimmer 30B — from their original providers onto OpenRouter for lower per-token pricing. GLM-5.3-Flash's output price alone dropped from $0.50 to $0.25 per million tokens, exactly half, with the same test generation confirmed working post-switch.

GLM-5.3 — Z.ai's flagship: hybrid sparse plus linear attention, a 1.3M-token context window, the strongest GLM tier for long-horizon coding and agent work. From 75 gems.

Qwen3.8 Max — Alibaba's flagship: natively multimodal (text, image, video input), 1M-token context, deep reasoning. From 100 gems.

Qwen3.8 Flash — the fast, low-cost sibling in the same family, multimodal input, 1M-token context. From just 10 gems.

Meta Muse Spark 1.2 — Meta's omni-modal flagship (text, image, video, audio, files), 1M-token context, the successor to the retired Llama line. From 70 gems.

Meta Muse Image — faithful instruction-following with accurate rendering of fine details like text, plots, and QR codes; launched on Fal on 2026-09-01 after being app-only inside Meta AI, Instagram and WhatsApp. From just 40 gems.

Mistral Large 3 — Mistral's current flagship: 675B/41B-active MoE, Apache 2.0 open weights, 262K context, tool calling and structured outputs. From just 25 gems.

Prices across the full batch span from 5 gems (GLM-5.3-Flash, a single text generation) to 4,600 gems (Seedance 2.5, a 30-second native-audio video) — roughly a 920× spread between the cheapest and most expensive model released this cycle, which is exactly why the picker below sorts by job instead of by launch date.

Which Model for Which Job

Job	Model	Why	From
Flagship general reasoning	Kimi K3	2.8T-parameter MoE, 1M-token context	240 gems
Long-horizon coding	GLM-5.3	1.3M context, strongest GLM coding tier	75 gems
Cheapest capable text	GLM-5.3-Flash	1M context, $0.25/Mtok output on OpenRouter	5 gems
Open-weights agent work	Meta Muse Spark 1.2	Omni-modal, 1M context	70 gems
Text-accurate image gen	Qwen Image 3.0 Pro	Dense, accurate in-image text rendering	160 gems
Fast, cheap image gen	Meta Muse Image	Accurate fine-detail rendering (text, QR codes)	40 gems
Native-audio video	Seedance 2.5	30-second clips, native audio	4,600 gems
Full songs from a prompt	MiniMax Music 2.6	Auto-generated lyrics included	600 gems
Expressive voiceover	ElevenLabs v3 TTS	29+ languages, voice cloning	300 gems

Try One Now

24 models, four weeks, one catch-up post. If you only try one thing from this list, make it the headliner: Kimi K3's 1M-token context and native vision handle document-heavy work that trips up smaller models, and it's live in chat right now at 240 gems a generation. If you want the three digests this one had to catch up on, see week 31, week 30, and week 28.