2026 OPENROUTER
TEN_DIM_
WEEKLY_
SNAPSHOT.

Open openrouter.ai/rankings and most Mac teams scan Top Models #1 once, then rewrite Cursor defaults — but early-June 2026 usage is already sliced: the overall chart may show MiMo-V2.5-Pro while the updated Programming Collections (June 2026 refresh) is led by DeepSeek V4 Flash and Tencent Hy3 preview. Platform weekly throughput is publicly cited around 25T–31T+ tokens; Chinese providers collectively hold roughly 60%+ of token share. This article is a full-page ten-dimension weekly snapshot readout for Mac developers: when to follow Top Models, when to follow Collections and slice charts, and how to update OpenClaw / Cursor routes on a Monday cadence. Bottom line: treat rankings as a routing config surface — open the dimension that matches the task; never let overall #1 overwrite programming and Agent chains. Below: pain points — ten-dimension lookup — June Programming mismatch — six rollout steps — matrix — case study — acceptance checklist.

1. Pain points: four ways Top Models #1 breaks your stack

(1) IDE follows the wrong board: Cursor defaults to overall #1 while programming traffic lives on the Programming slice — completion quality and cost both lose. (2) Collections vs Leaderboard conflation: Collections curate by scenario (programming / Agent); the main board counts all-site tokens. Hy3 preview ranks high in programming collections but may sit outside the overall top ten. (3) Ignoring the Provider layer: SiliconFlow-class relays can move 4T+ weekly tokens yet account for <2% revenue (see the dual-track article); changing model id without locking provider slug still drifts. (4) No weekly diff: June chart velocity outpaces release cycles; openclaw.json frozen on last month's model triggers bill spikes when previews end (the Trending article covers the time axis — this piece adds full-page alignment).

2. Ten-dimension panel: what each chart answers

Dimension (rankings page)	Metric	Mac use
Top Models / LLM Leaderboard	Cumulative tokens this week	OpenClaw default, general Agent
Market Share	Share by model author	Vendor concentration, compliance
Languages	By natural language	Separate EN/ZH defaults
Programming	By programming language (e.g. Python)	Cursor / Cline / Continue
Context Length	By prompt length bucket	Long-document RAG routing
Tool Calls	Tool-calling traffic	OpenClaw / Hermes stack
Images / Image Output	Multimodal in/out	Vision Agent vs ComfyUI
Audio Input	Speech input tokens	Podcast / meeting STT API
Trending	Week-over-week gain	Gray pool (not production default)
Programming Collections	Scenario curation + June 2026 usage	IDE-first reference table

3. Early-June platform scale: reading ~31T weekly tokens and 60%+ Chinese share

Third-party snapshots of OpenRouter public endpoints in late May–early June report roughly 31.34T tokens/week; official Series B materials often cite ~25T/week. The gap is statistical window and routing coverage — ops only needs relative rank and WoW delta. Chinese vendors (Xiaomi, Alibaba, DeepSeek, MiniMax, Moonshot, and others) collectively account for 60%+ of token share in multiple analyses — that is not "global API market share" but developers on OpenRouter choosing cost-performance. Mac teams should translate this into: default chains favor low-cost open / preview lines; Dollar track keeps Claude Opus / GPT-5.5 for hard-task fallback.

4. Programming Collections vs Top Models: June 2026 mismatch

OpenRouter refreshed Programming collection copy in June 2026: rankings reflect real programming-scenario usage routed through the platform. Public reporting and the collections page show a leading cluster that diverges from overall #1:

• DeepSeek V4 Flash: MoE 284B / 13B active, 1M context — high-throughput coding;
• Tencent Hy3 preview: high on programming collections — IDE gray candidate;
• MiMo-V2.5-Pro / DeepSeek V4 Pro: steady on overall chart and often co-present on programming;
• Claude Opus / Sonnet 4.6–4.7: not always #1 by token volume, still common on Dollar track for complex refactors.

Mismatch rule: when Top Models #1 ≠ Programming #1, Cursor should follow Programming + Collections; OpenClaw follows Top Models + Tool Calls. Do not let one overall-chart article overwrite IDE and Agent configs. Cross-reference the May Programming decode for SWE-bench vs real-usage contrast.

5. Six weekly rollout steps (Monday, 30 minutes)

Step 1 — Capture the ten-dimension one-pager

Log Top Models top five, Programming top five, Tool Calls top three, Market Share Chinese-vendor combined share.

Step 2 — Open Collections and diff

Visit the programming collection; flag models >5 ranks away from the main board (e.g. Hy3 preview).

Step 3 — Update Cursor routes

Programming default → Programming #1–#2; hard bugs → Dollar track Sonnet / Opus.

Step 4 — Update OpenClaw fallback

Primary follows Top Models; Tool Calls top three enter fallback; Collections newcomers only at gray tail.

Step 5 — Provider and preview calendar

Lock provider slug; calendar free-preview end dates; run openclaw doctor on expiry day.

Step 6 — Mac three-tier acceptance

Steady-state MLX-quantizable models on-box; experimental models on remote Mac regression; hard tasks on API Dollar track.

6. Three-tier decision matrix

Task	Which chart	Mac path
Cursor daily completion	Programming + Collections	Local MLX small model or OpenRouter low-cost line
OpenClaw 7×24	Top Models + Tool Calls	Remote Mac Gateway persistent
Long-document RAG	Context Length >32K bucket	On-box 64GB+ or remote Mac MLX
Image assets	Image Output	ComfyUI on-box + API fallback
Architecture / security review	Not token boards	Dollar track Claude / GPT

7. Case study: wiring Top Models #1 into Cursor, programming bill +38%

Takeaway: the overall chart measures 'hottest on the network'; Collections / Programming measure 'best fit for the scenario'. Complements Trending and dual-track articles without repeating them.

8. Industry insight: rankings become weekly routing PR

At ~31T-class weekly throughput, model listing cycles are shorter than traditional software release trains. OpenRouter's rankings + collections combination is effectively public telemetry for a multi-tenant router. Mac teams gain an edge: Apple Silicon on-box MLX baselines for steady models; remote Mac runs Collections newcomers for regression while the laptop keeps unified memory for review and Dollar-track fallback. Windows or Linux can call OpenRouter too, but launchd-persistent Gateway, parallel Xcode / FCP / ComfyUI pipelines, Metal sidecar inference remain smoother on macOS. Pure cloud Linux hosts excel at raw API relay but lack the local graphics and IDE toolchain co-location Mac provides. When you need physical isolation between overall-chart steady-state and programming-chart experiments, MACGPU remote Mac nodes carry OpenClaw and IDE gray traffic; the laptop keeps review workloads and expensive fallback only — no experiment traffic saturating 36GB unified memory during a FCP export.

9. Citable numbers and acceptance checklist

(1) Weekly tokens (third-party snapshot): ~31.34T; official cite ~25T. (2) Chinese-vendor token share: 60%+ (multiple analyses). (3) Programming leaders (June collections): DeepSeek V4 Flash, Hy3 preview. (4) Post-split routing case: $3,400 → $2,108 (-38%).

Acceptance checklist: Ten-dimension panel captured | Programming vs Top Models mismatch logged | Cursor default aligned to Programming | OpenClaw fallback aligned to Tool Calls | Provider slug locked | Preview-end calendar set | Remote Mac 50-prompt regression |

Q: Does this conflict with Trending? No — Trending governs "trial this week"; this piece governs "which grid cell to read." Q: Still read Programming slice separately? Yes, and cross-check the June Collections refresh. Q: What does MACGPU add? Remote Mac runs gray and steady Agent load; laptop keeps MLX baseline and Dollar-track fallback.