2026 OPENROUTER
IMAGE_
OUTPUT_
LANGUAGES_
MAC.
On openrouter.ai/rankings, post–Series B (~25T tokens/week as of May 26), the board splits into parallel slices. Our overall chart, Programming, Tool Calls, and Images input trio already cover text, agents, and multimodal understanding — but image generation (Image Output) and Languages still lacked a dedicated close-out. Late May, Image Output reshuffled around Flux.2 Pro, xAI Grok Imagine, and Recraft V4.1, while Market Share shows Chinese vendors above 60% of weekly tokens despite the Languages UI defaulting to English. This article delivers: chart reading — Image Output snapshot — Languages vs Market Share — vendor week data — six steps — decision matrix — case study — acceptance checklist.
1. Pain points: overall and Images-input charts do not pick image generators
(1) Wrong dimension: #1 MiMo-V2-Pro is text-token volume, not “posters generated per week.” The 0528 Images slice measures image understanding; Image Output tracks models with output_modalities including image, often billed per image (e.g. Grok Imagine from ~$0.05/image). (2) Languages trap: the Languages filter defaults to English — easy to assume “English chart leaders = best for my users,” while Market Share is dominated by Xiaomi, Alibaba, MiniMax, etc. (3) Billing split: mixing ComfyUI local “electricity only” with unchecked OpenRouter image APIs blows monthly spend. (4) Mac unified memory: ComfyUI Flux plus API batch jobs on 36GB M3 often OOM. (5) Close-out week releases (Step 3.7 Flash, Claude Opus 4.8 Fast) trend on the homepage but do not displace Flux/Grok at the top of Image Output.
2. Reading Image Output × Languages × Market Share (late May)
| Slice | Measures | Common mistake | Mac use |
|---|---|---|---|
| Image Output | Generated image traffic | Same as Images input | Posters, ecommerce, thumbnails |
| Languages | Prompt language distribution | Equals best model per language | Split UI locale vs model vendor |
| Market Share | Token share by author | Equals legal/compliance verdict | Cost and fallback vendor order |
Weekly ops: align Image Output + Market Share; for bilingual products, compare English Languages slice to real user locales. Image requests need modalities: ["image"] or ["image","text"] per OpenRouter docs.
3. Image Output snapshot (week of 2026-05-29)
| Tier | Models | Use case | Mac path |
|---|---|---|---|
| T1 | black-forest-labs/flux.2-pro, flux.2-max | Product, brand KV | API; local ComfyUI for A/B |
| T2 | x-ai/grok-imagine-image-quality | Multilingual in-image text | API; per-day $ cap |
| T3 | recraft/* V4.1 | SVG, gradients | API |
| T4 | google/gemini-3.1-flash-image-preview | Text + image one shot | API |
| T5 | ComfyUI + Flux/SDXL | Nightly batch, auditable assets | Local or remote Mac 128GB |
Overlap between Image Output and Images-input leaders is often <35%. Split OpenRouter budgets: “review agent” (Images input) vs “asset generation” (Image Output).
4. Languages: English UI vs Chinese-dominated Market Share
Public May 2026 aggregates: six Chinese providers ~60%+ weekly tokens; MiMo-V2-Pro ~21.1% vs OpenAI ~7.5% platform share (third-party summaries). Many English prompts route to Chinese base models on cost. Mac teams should run 50 English + 50 Chinese poster prompts per candidate model. MiMo + Qwen reportedly ~49% of programming tokens — relevant for “English UI, Chinese ops” dual routing.
5. Market Share close-out and Mac vendor strategy
| Vendor | Flagship | ~Share | Mac note |
|---|---|---|---|
| Xiaomi | MiMo-V2-Pro | ~21.1% | Text default; image gen still Flux/Grok |
| Alibaba | Qwen 3.6 / 3.7 Max | ~13.9%+ | Long context, bilingual agents |
| OpenAI | GPT-5.5 | ~7.5% | Compliance fallback |
One analyzed week (Apr 27–May 3): Chinese models 7.942T (+81.7% WoW), US models -34.6%. Use Market Share for text fallback order, not as default image generator.
6. Six steps: three slices to Mac image routing
Step 1 — Pull Image Output + Market Share weekly
Step 2 — Filter models API by output_modalities image
Step 3 — Separate understanding vs generation budgets
Step 4 — A/B ComfyUI vs API (20 images each)
Step 5 — Bilingual poster probe (30 EN + 30 ZH)
Step 6 — Move >200 images/night to remote Mac queue
7. Three-lane matrix: ComfyUI / OpenRouter API / remote Mac
| Scenario | Lane | Acceptance |
|---|---|---|
| Single poster iteration | API Grok / Flux Flex | <25s, <$0.08/image |
| Brand asset library | Local ComfyUI | 20 images no OOM |
| 500+ images/night | Remote Mac queue | <2% failure in 8h |
8. Case study: cross-border DTC team cut image spend 41%
Six-person DTC on Mac Studio M2 Ultra 64GB used Claude for all “image” work ($4,100/mo OpenRouter). After Image Output routing: Flux.2 Pro + Grok Imagine per-image, Claude for copy only, Qwen for Chinese prompts, 800/week batch to MACGPU remote M4 Max ComfyUI nights — image-related spend $2,420 (-41%), no daytime swap from ComfyUI + Claude overlap.
9. Insight: image chart and vendor chart will diverge
Text tokens are shifting east; Image Output stays with BFL, xAI, Google, Recraft. Windows/Linux cloud GPUs run ComfyUI, but macOS wins for ColorSync, Final Cut handoff, and launchd night queues. When 64GB is not enough for daytime edits plus 500 API images, rent remote Apple Silicon: MACGPU M4 Max 128GB with ComfyUI Flux workflows and shared OpenRouter keys — local Cursor for review only.
10. Citable numbers and FAQ
① ~25T tokens/week (May 26). ② Chinese vendors >60% share. ③ MiMo ~21.1%. ④ Grok Imagine ~$0.05/image. ⑤ Case $4,100 → $2,420. ⑥ Image Output vs Images-input overlap <35% (validate weekly).