Shortlist models for software development using the reported Coding Index. Compare LiveCodeBench, Terminal-Bench, and SciCode for the tasks that matter to you.
The answer in 60 seconds
Claude Fable 5.1 (Adaptive Reasoning, Max Effort, Default Fallback) leads the reported Coding Index at 81.6. Start with it for a capability-first shortlist, then compare alternatives against your workload, budget, and deployment requirements.
Using Artificial Analysis API benchmark data, cached for up to 24 hours. The source measurement date is not supplied. Selection rules reviewed .
Coding Index: 81.6
Why this pick: Highest reported Coding Index in this dataset; start here when benchmark capability is your priority.
Before choosing: Repository repair, review quality, and cost need an evaluation on your own code and agent harness.
Also compare: GPT-5.6 Sol (xhigh).
Why this pick: Highest reported Coding Index among the open-weight models in this dataset.
Before choosing: Open weights do not imply a permissive license or a fit for your hardware. Check the model license and memory requirements.
Also compare: GLM-5.3 (max).
Why this pick: Highest reported Terminal-Bench Hard score among models with a result for that exact benchmark.
Before choosing: Coverage is incomplete. Terminal-Bench Hard results do not establish a winner for other tasks or benchmark versions.
Also compare: Claude Fable 5 (Adaptive Reasoning, Max Effort, Opus 4.8 Fallback).
Source: Artificial Analysis model data. These are WhatLLM selection rules applied to published scores, not WhatLLM-run evaluations. Missing results are excluded from each comparison. Tied scores do not imply a performance difference. Read the methodology.
| Rank | Model | Coding Index | LiveCodeBench | Terminal-Bench | SciCode | License |
|---|---|---|---|---|---|---|
| 1 | Claude Fable 5.1 (Adaptive Reasoning, Max Effort, Default Fallback) Anthropic | 81.6 | - | - | 63% | Proprietary |
| 2 | GPT-5.6 Sol (xhigh) OpenAI | 78.3 | - | 61%Hard | 57% | Proprietary |
| 3 | Claude Opus 5 (Adaptive Reasoning, Max Effort) Anthropic | 78 | - | - | 56% | Proprietary |
| 4 | GPT-6 Astra (high) OpenAI | 77.1 | - | - | 55% | Proprietary |
| 5 | Grok 4.6 (high) SpaceXAI | 76.8 | - | - | 56% | Proprietary |
| 6 | GPT-5.6 Terra (max) OpenAI | 76.7 | - | 58%Hard | 55% | Proprietary |
| 7 | Claude Fable 5 (Adaptive Reasoning, Max Effort, Opus 4.8 Fallback) Anthropic | 76.5 | - | 63%Hard | 61% | Proprietary |
| 8 | Muse Spark 1.3 (xhigh) Meta | 76.5 | - | - | 60% | Proprietary |
| 9 | Gemini 3.8 Flash (high) | 76.3 | - | - | 57% | Proprietary |
| 10 | Kimi K3 (max) Kimi | 76.2 | - | - | 60% | Open |
The table sorts by reported Artificial Analysis Coding Index and excludes models without that index. Equal scores are ordered by name. Task-level scores below add context without changing the sort:
Fresh code-generation problems across languages. Useful for raw coding ability, but not a complete repository-agent test.
Tests complex terminal operations, shell scripting, DevOps work, and recovery over longer trajectories.
Scientific computing and research programming. Tests ability to implement algorithms from papers and numerical methods correctly.
Coding Index from Artificial Analysis. Terminal-Bench versions are labeled separately and should not be compared as the same test. Read which coding benchmark matters before treating one score as a universal ranking.
See exact pricing, latency, and benchmark scores for all 10 coding models in our interactive comparison tool.
Claude Fable 5.1 (Adaptive Reasoning, Max Effort, Default Fallback) leads the reported Coding Index at 81.6. Start with it for a capability-first shortlist, then compare alternatives against your workload, budget, and deployment requirements.
Start with the top Coding Index and Terminal-Bench models above, then run a private evaluation on recent issues from your repositories. Public rankings do not measure your agent harness, review standards, or failure cost.
Kimi K3 (max) leads the open-weight models in this dataset with a Coding Index of 76.2.Verify its license, total parameter count, and provider or hardware fit before deployment.
Qwen3-Coder 30B is the practical 24GB-class option, while Qwen3-Coder-Next targets systems with at least 64GB of usable memory. See the hardware-tier guide for artifact sizes and commands.
Compare the current OpenAI and Anthropic variants in the live table rather than relying on a fixed brand verdict. Use the same harness for repository work and include accepted-result cost, latency, and review quality.
Using Artificial Analysis API benchmark data, cached for up to 24 hours. The source measurement date is not supplied. Selection rules were reviewed on September 6, 2026. Data refreshes and editorial reviews are separate; a page load is not a new benchmark evaluation.
Self-hostable models
🧮AIME 2025 rankings
🤖Tool use & agents
⚡Any 2–4 models
Data sources: Rankings based on the Artificial Analysis Coding Index. Explore all models in our interactive leaderboard or compare models side by side.