ModelRank AI Coding Model Tiers
Compare coding capabilities by tier, with supporting evidence and API quotes when you need them.
Coding capability tiers
Tiers are editorial assessments. Models within each tier are alphabetical, with no implied capability order.
DeepSeek V4 Flash 0731 = 100. The index reflects two shared tests, not capability multiples or statistically established gaps.
| Coding tier | Model | Coding IndexDifference from baseline | Evidence | Price (1M tokens) | Context | Details |
|---|---|---|---|---|---|---|
| S+Top-tier candidate · 1 | ||||||
| S+ | Claude Fable 5.1 anthropic | 120.65 +20.65 | Cross-checked | 1,000,000 | Details → | |
| SFrontier coding group · 5 | ||||||
| S | Claude Fable 5 anthropic | 114.18 +14.18 | Cross-checked | 1,000,000 | Details → | |
| S | Kimi K3 moonshotai | 113.02 +13.02 | Cross-checked | 1,048,576 | Details → | |
| S | Claude Opus 5 anthropic | 112.64 +12.64 | Cross-checked | 1,000,000 | Details → | |
| S | GPT-5.6 Sol openai | 112.62 +12.62 | Cross-checked | 1,050,000 | Details → | |
| S | GPT-6 Astra openai | 112.28 +12.28 | Cross-checked | 1,050,000 | Details → | |
| A+Near-frontier group · 9 | ||||||
| A+ | Gemini 3.8 Flash google | 111.92 +11.92 | Cross-checked | 1,048,576 | Details → | |
| A+ | GLM-5.3 z-ai | 111.83 +11.83 | Cross-checked | 1,048,576 | Details → | |
| A+ | Gemini 3.7 Flash google | 111.28 +11.28 | Cross-checked | 1,048,576 | Details → | |
| A+ | GPT-5.6 Terra openai | 110.54 +10.54 | Cross-checked | 1,050,000 | Details → | |
| A+ | Grok 4.6 x-ai | 108.55 +8.55 | Cross-checked | 500,000 | Details → | |
| A+ | Claude Opus 4.8 anthropic | 107.83 +7.83 | Cross-checked | 1,000,000 | Details → | |
| A+ | Grok 4.5 x-ai | 106.47 +6.47 | Cross-checked | 500,000 | Details → | |
| A+? | DeepSeek V4.1 Flash deepseek | — | Provisional · source review pending | 1,048,576 | Details → | |
| A+? | Qwen3.8 Max 0902 qwen | — | Provisional · source review pending | 1,000,000 | Details → | |
| AStrong coding group · 4 | ||||||
| A | Claude Sonnet 5 anthropic | 105.06 +5.06 | Cross-checked | 1,000,000 | Details → | |
| A | Qwen3.8-Next-Flash qwen | 104.89 +4.89 | Cross-checked | 1,000,000 | Details → | |
| A | GLM-5.3-Flash z-ai | 104.81 +4.81 | Cross-checked | 1,048,576 | Details → | |
| A | GPT-5.6 Luna openai | 104.63 +4.63 | Cross-checked | 1,050,000 | Details → | |
| B+Established reference group · 3 | ||||||
| B+ | DeepSeek V4 Pro 0813 deepseek | 100.69 +0.69 | Cross-checked | 1,048,576 | Details → | |
| B+ | GLM-5.2 z-ai | 100.32 +0.32 | Cross-checked | 1,048,576 | Details → | |
| B+ | DeepSeek V4 Flash 0731 deepseek Reference baseline = 100 | 100.00 0.00 | Cross-checked | 1,310,720 | Details → | |
| BOther capable candidates · 2 | ||||||
| B | MiMo V2.5 Pro xiaomi | 91.23 -8.77 | Single-source measurements | 1,050,000 | Details → | |
| B | HY3 tencent | 88.93 -11.07 | Single-source measurements | 262,144 | Details → | |
| B-More basic within this selection · 1 | ||||||
| B- | MiMo V2.5 xiaomi | 83.98 -16.02 | Single-source measurements | 1,050,000 | Details → | |
No matching models.
Results reflect particular model configurations. Agent harnesses, reasoning effort, tools, budgets and tasks may differ. Close indices do not establish superiority. B describes a relative position in this selected group, not an inability to code.
Awaiting comparable evaluation (20)
Models awaiting tier review remain in the catalog. An unassigned tier does not mean weaker capability.
Compare current models
USD / API
API cost estimate
Estimated token cost
Estimates use the displayed channel and time band; DeepSeek uses peak rates. Cache discounts, tool fees, long-context price tiers and uncounted additional tokens are excluded.
Ranking methodology
Only coding evidence determines the assessment. Measurements, known configurations and source dates remain available in the expanded notes; prices are separate purchasing information.
ModelRank assigns editorial tiers using public coding evidence, not automatic index thresholds or statistically proven classes. Price, speed, context size and sponsorship do not affect tiers.
Known model and agent configurations are retained with their sources. Unpublished evaluation dates, budgets, sample sizes and uncertainty remain unknown. Recent publication alone cannot rule out contamination.
Coding Index = 100 × the geometric mean of model/baseline ratios on AA Terminal-Bench 2.1 and SciCode, equally weighted. The baseline is fixed and displayed indices are integers. LiveBench and Terminal-Bench 4.0 provide supplementary checks; incomplete common coverage excludes them from this index.
100 × √[(TB / TB₀) × (SciCode / SciCode₀)]
Results reflect particular model configurations. Agent harnesses, reasoning effort, tools, budgets and tasks may differ. Close indices do not establish superiority. B describes a relative position in this selected group, not an inability to code.
ModelRank assigns editorial tiers using public coding evidence, not automatic index thresholds or statistically proven classes. Price, speed, context size and sponsorship do not affect tiers.
Qwen3.8 Max 0902 has a provisional A+ tier pending original-source review, with no numeric rank or comparable index. A transcribed AA index is not independent verification.
DeepSeek V4.1 Flash is provisionally A+ by editorial judgment, with low confidence. Early community feedback is mixed; the concrete coding comparison concerns GLM-5.3 Flash, not GLM-5.3. No comparable independent Coding Index is available.
Prefer recently collected tasks after a fixed model snapshot or well-governed private sets. A new benchmark name or a recent data fetch does not establish freshness. Require original sources, sample counts and test settings. Unknown task or evaluation dates remain unknown; recency alone cannot guarantee freedom from contamination.