The Leaderboard (Coding Score per $1 of Output)
Sorted by value = coding composite ÷ output $/1M. Composites verified 2026-08-28 (Artificial Analysis via OpenRouter); DeepSeek off-peak is 50% of peak:
| Rank | Model | Coding index | Output $/1M | Coding per $ |
|---|---|---|---|---|
| 1 | DeepSeek V4 Flash (off-peak) | 56.2 | $0.66 | 85.2 |
| 2 | GPT-5.6 Luna | 71.4 | $1.20 | 59.5 |
| 3 | DeepSeek V4 Flash | 56.2 | $1.32 | 42.6 |
| 4 | DeepSeek V4 Pro (off-peak) | 68.8 | $1.98 | 34.7 |
| 5 | Gemini 3.5 Flash-Lite | 49.3 | $2.50 | 19.7 |
| 6 | DeepSeek V4 Pro | 68.8 | $3.96 | 17.4 |
| 7 | GLM 5.3 | 74.8 | $4.40 | 17.0 |
| 8 | Grok 4.6 | 76.8 | $6.00 | 12.8 |
| 9 | Gemini 3.7 Flash | 76.1 | $6.00 | 12.7 |
| 10 | Claude Sonnet 5 | 71.5 | $10.00 | 7.2 |
| 11 | GPT-5.6 Terra | 76.7 | $12.00 | 6.4 |
| 12 | Claude Opus 5 | 78.0 | $25.00 | 3.1 |
| 13 | GPT-5.6 Sol | 77.4 | $30.00 | 2.6 |
| 14 | Claude Fable 5 | 76.5 | $50.00 | 1.5 |
Read it as a staircase, not a ranking: the budget tier delivers 20-85 value units, the balanced tier 6-17, the flagship tier 1-3. The quality delta between tiers (71 vs 77 coding) is small; the price delta (30×) is not. That asymmetry is the entire business case for model routing.
Intelligence and Agentic Leaderboards
Coding isn't every workload. The same math on the other two axes:
Intelligence per $1 output
Agentic per $1 output
Same staircase, same shape: budget tiers win every value axis. GLM 5.3 is the one balanced-tier exception worth noting — its agentic score (59.1) matches flagships, making it the best value for agent workloads specifically. See agent model selection for that deep dive.
How to Actually Use This Leaderboard
Route routine traffic to the top of the list
Classification, extraction, FAQ, summaries — 50-70% of most SaaS traffic belongs on the budget staircase. The routing playbook has the task matrix.
Keep flagships for the visible 10%
Complex reasoning, long refactors, client-facing writing. The value metric correctly ignores these — they justify their price differently.
Check the denominators quarterly
Output prices move. When DeepSeek or Google cut output rates, the leaderboard reshuffles — re-run the benchmark hub before your next pricing review.
Value is not a substitute for fit
A 85/$ model that can't do agentic loops is worthless for agents. Use fit scores per use case, then value to break ties within a tier.
Frequently Asked Questions
Which AI model has the best quality per dollar?
By coding score per $1 of output at verified 2026-08-28 rates: GPT-5.6 Luna leads (71.4 coding ÷ $1.20 = 59.5/$), then DeepSeek V4 Flash off-peak (56.2 ÷ $0.66 = 85/$), and Gemini 3.5 Flash-Lite (49.3 ÷ $2.50 = 19.7/$). Flagships score 1-3 per dollar — the value gap is an order of magnitude.
Is quality-per-dollar the right metric to choose a model?
It's the right metric for volume traffic. For latency-critical or quality-critical paths, use fit scores and speed instead. Value rankings tell you where to route the 70% of traffic that's routine; they don't tell you what to do with the 10% that's hard.
How is quality-per-dollar calculated?
Composite index (0-100, Artificial Analysis via OpenRouter) divided by the model's output price per 1M tokens. Output price is the correct denominator because output tokens drive 60-90% of real spend.
Why does DeepSeek's value look so extreme?
Its ~97% cache-hit discount and 50% off-peak pricing compound: $0.22 input / $0.66 output off-peak with $0.014 cached input makes it the cheapest serious production model on the market, and its coding composite (56.2) is respectable, not weak.
Where can I see the full leaderboard?
The benchmark hub ranks every scored model by coding-per-dollar with a quality-vs-price scatter chart, and each model page shows its own score-per-dollar line.