Mesh API Value Ranking: capability per dollar.
A price-performance value ranking of 122 models published through Mesh API, to help you pick the one that delivers the most capability for the least cost, instead of overpaying out of habit. Mesh API is AI Fiesta's routing gateway: one OpenAI-compatible endpoint to 1,000+ models from every major lab, so you can compare and switch models without touching your code.
← All researchThe priciest model is rarely the smart buy.
DeepSeek V4 Flash scores 47 on the Artificial Analysis Intelligence Index. GPT‑5.5 scores 60, but costs $11.25 vs $0.18 per blended million tokens. You pay 64x more for those thirteen points. Finding trades like that, the most capability per dollar, is exactly what the ranking below is for.
Ranked by value, not vibes.
Pick a job, the table re-ranks on the benchmark that actually drives it. Click any column to sort. Click a model row to see the full benchmark breakdown.
| # | Model | Value | Capability | $/M in | $/M out | Blended $ | tok/s | TTFT | Context |
|---|---|---|---|---|---|---|---|---|---|
| 1 | Qwen3.5-397B-A17B Qwen | 184.8 | 45idx | $0.17 | $1.03 | $0.39 | 53.2 | 2.54s | 262K |
| 2 | Qwen3.5-122B-A10B Qwen | 175.9 | 42idx | $0.12 | $0.92 | $0.32 | 140.3 | 2.44s | 262K |
| 3 | Qwen3.5-35B-A3B Qwen | 163.8 | 37idx | $0.06 | $0.46 | $0.16 | 150.1 | 2.39s | 262K |
| 4 | MiniMax M2.5 MiniMax | 159.5 | 42idx | $0.30 | $1.20 | $0.52 | 188.1 | 3.14s | 197K |
| 5 | Qwen3.5-27B Qwen | 158.7 | 42idx | $0.20 | $1.56 | $0.54 | 77.5 | 5.8s | 262K |
| 6 | Nemotron 3 Super 120B NVIDIA | 154.3 | 36idx | $0.10 | $0.50 | $0.20 | 154.9 | 1.4s | 262K |
| 7 | DeepSeek V4 Flash DeepSeek | 147 | 34.3idx | $0.14 | $0.28 | $0.18 | | | | | 1M |
| 8 | Kimi K2.5 MoonshotAI | 145.7 | 47idx | $0.60 | $3.00 | $1.20 | 45.1 | 3s | 262K |
| 9 | Kimi K2 Thinking MoonshotAI | 127.6 | 41idx | $0.60 | $2.50 | $1.07 | 122.1 | 1.57s | 131K |
| 10 | Gemma 4 26B Google | 126.9 | 31idx | $0.13 | $0.40 | $0.20 | | | | | 262K |
| 11 | Gemini 3.1 Flash-Lite Google | 120 | 34idx | $0.25 | $1.50 | $0.56 | 333 | 6.12s | 1M |
| 12 | Qwen3-Max Thinking Qwen | 109.5 | 40idx | $0.78 | $3.90 | $1.56 | | | | | 262K |
| 13 | Qwen3-Coder-Next Qwen | 105.6 | 28idx | $0.12 | $0.75 | $0.28 | 75.4 | 1.75s | 262K |
| 14 | Qwen3-235B Thinking 2507 Qwen | 105 | 30idx | $0.15 | $1.50 | $0.49 | 52.7 | 2.89s | 131K |
| 15 | DeepSeek V3.1 DeepSeek | 104.3 | 28idx | $0.15 | $0.75 | $0.30 | 24.9 | 2.04s | 33K |
| 16 | DeepSeek V3.2 Exp DeepSeek | 104 | 28idx | $0.27 | $0.41 | $0.30 | | | | | 164K |
| 17 | Qwen3-Max! Qwen | 103.4 | 31idx | $0.36 | $1.43 | $0.63 | 57.8 | 2.51s | 262K |
| 18 | Gemini 3 Flash Google | 102.9 | 35idx | $0.50 | $3.00 | $1.13 | 175.6 | 42.5s | 1M |
| 19 | Qwen3-235B-A22B (2507) Qwen | 101 | 25idx | $0.07 | $0.10 | $0.08 | 48.2 | 2.36s | 262K |
| 20 | Qwen3-Next-80B Thinking Qwen | 100.8 | 27idx | $0.10 | $0.78 | $0.27 | 143.5 | 2.2s | 131K |
| 21 | DeepSeek V4 Pro DeepSeek | 85.1 | 36idx | $1.74 | $3.48 | $2.17 | 79.3 | 1.15s | 1.1M |
| 22 | Qwen3-235B Instruct 2507 Qwen | 84.3 | 25idx | $0.23 | $0.92 | $0.40 | 48.2 | 2.36s | 131K |
| 23 | Claude Haiku 4.5 Anthropic | 78 | 31idx | $0.80 | $4.00 | $1.60 | 90.5 | 0.89s | 200K |
| 24 | Kimi K2 MoonshotAI | 71.1 | 26idx | $0.57 | $2.30 | $1.00 | 24.6 | 2.51s | 131K |
| 25 | Gemini 2.5 Pro Google | 69.3 | 35idx | $1.25 | $10.00 | $3.44 | 142.2 | 23.39s | 1M |
The cost vs capability frontier.
Every dot is a routable model on Mesh. The line traces the Pareto frontier, the models that offer the most capability you can get at any given price. Anything well above the line is overpriced for what it does. Anything on the line is the smart buy at that budget.
Methodology and disclosure.
Fully auditable: every number is public and the formula is right here.
Capability
A 10-eval composite, rescaled 0|100 (floor 8, ceiling 61, the current frontier). Built to resist saturation: single benchmarks like GPQA and AIME are now tied within margin of error at the frontier, so a composite is the honest capability term. Every model's Index was re-checked against its AA page on 6 Jun 2026.
Price
Blended per million tokens. Production traffic is input-heavy (long context, system prompts, RAG), so input is weighted 3:1. A user-adjustable slider ships next.
The log
The log dampens price so a cheap-but-weak model can't top the board on denominator alone, while still rewarding cheap and capable.
Honest gaps
Coverage, not fabrication. Of 1,000+ routable text models, 122 have a citable AA Intelligence Index and are ranked; 176 don't and sit on the data-pending shelf. We show | for any missing field and never guess. LMArena Elo is display-only.
Where each number comes from
We never run the models to score them. The Mesh API supplies prices, context and a live routability check; everything about capability and speed is harvested from public benchmarks and cited. So roughly: price and context = Mesh API; intelligence and benchmarks = Artificial Analysis + model cards; Value Score = a formula over the two.
| Column on the board | Source | How we get it |
|---|---|---|
| $/M in . $/M out . blended | Mesh /v1/models price catalogue | Mesh API |
| Context window | Mesh /v1/models | Mesh API |
| Routable today | Mesh /v1/models + live test calls | Mesh API |
| Intelligence Index (the capability axis) | Artificial Analysis (v4.0) | Cited public data |
| GPQA . SWE-bench . AIME | Official model cards / AA / vals.ai | Cited |
| LMArena Elo | arena.ai | Display-only, cited |
| Output tok/s . TTFT | Artificial Analysis median , a few measured first-party through Mesh where AA had none | Mostly cited . a few Mesh-measured |
| Value Score | computed: capability / log10(blended + 2) | Computed (a formula) |
/v1/models API, which also verifies every listed model is genuinely routable today. Last verified 6 Jun 2026.Asia's largestsovereign AI gateway.
1,000+ AI models. One endpoint. Your data stays inside your border.