Expanded capability & pricing table (AA Intelligence Index v4.3.2 composite; Terminal-Bench 2.1 and DeepSWE v1.1 are harness- and effort-dependent—ranges or representative high/independent figures shown; pricing is standard or introductory list rates per 1M tokens, input/output):
Model | AA Index | Terminal-Bench 2.1 | DeepSWE | Pricing (in/out) |
|---|---|---|---|---|
Claude Opus 5.5 | 57.6 | ~87.6 | ~74 | $4 / $20 |
Claude Sonnet 5.5 | 56 | ~83 | — | $2 / $10 |
Kimi K3.1 Pro | 55.1 | — | — | ~$2–4 / $10 |
Claude Fable 5.1 | 53.4 | 85–91 | ~67.4 | $10 / $50 |
GPT-6 Astra | 52.7 | ~87.3 | 74.1 | $10 / $50 |
Gemini 4 Argon | 52.6 | ~86 (est.) | ~77.9 | $2 / $10 (intro; $4/$20 from Dec 1) |
GPT-6.1 Sol | 51.8 | ~75.2 | — | $2 / $10 |
| 48 | TB 4.0: 50.3% max | - | $1 / $5 |
Grok 4.7 | ~46.5 | ~73.4 | — | $2 / $6 |
GLM-5.3 | 44.8 | 71.5–88.2 | ~69 | $1.40 / $4.40 |
Gemini 3.8 Flash | ~40.9 | 81–89 | 73.8 | $0.75 / $3.75 (intro; $1.50/$7.50 from Jan 1) |
Muse Spark Open | 38.9–39 | ~76.4 | ~60 | $1.25 / $4.25 |
DeepSeek V4.1 Flash | ~39.5 | 74–90.6 | ~77 (self) | $0.30 / $1.20 |
GPT-6 Luna | ~38 | 66.6 | - | $0.10 / $0.50 |
Reflection Beam | 30 | ~80.1 (vendor) | ~43–44 | ~$0.25–0.40 / $1–1.20 |
Notes
AA Intelligence Index is the most consistent composite across models (Artificial Analysis).
Terminal-Bench 2.1 numbers differ a lot by harness (Terminus, Claude Code, mini-swe-agent, etc.) and effort. DeepSeek and some others have high self-reported scores that independent runs often put lower.
DeepSWE has fewer independent numbers for the newest models; many top scores are self-reported.
Scores vary significantly by reasoning effort, harness (e.g. Terminus 2, Claude Code, Codex), and source (vendor self-reports often higher).
AA Index is independent composite.
Pricing is approximate list rates; long-context surcharges, cache discounts, and batch rates apply. Intro prices for Gemini models are temporary.
Pricing is standard list rates (cached input is usually much cheaper, especially on OpenAI models).
Self-hostable models table (open weights; “best fit” is a practical Q4-class / recommended quant VRAM or unified-memory figure for usable context, not absolute minimum; hosted API prices where available):
(includes AA Intelligence Index, Terminal-Bench 2.1, and DeepSWE for comparison). Scores are representative/recent figures (late Oct 2026); many are vendor-reported or harness-dependent and can vary significantly. “Best-fit memory” is a practical Q4-class / recommended quant figure for usable context (not absolute minimum).
Model | AA Index | Terminal-Bench 2.1 | DeepSWE | Best-fit memory | Hosted pricing (in/out) | Notes |
|---|---|---|---|---|---|---|
Muse Glimmer 30B | ~17 (AA high; earlier launch ~35) | ~51.7% (Meta, Terminus 2) | Low / not competitive on long-horizon SWE | ~19–20 GB (Q4, fits 24 GB GPU) | ~$0.30 / $1.20 | Apache 2.0; strong tool-use/agentic for size; weaker terminal work |
Gemma 4 31B | ~14–17 | ~43% (comparisons) | Low / limited strong reports | ~18–22 GB (Q4) | ~$0.09–0.14 / $0.34–0.40 | Apache 2.0; dense; solid overall local option |
Gemma 4 26B A4B | ~16–17 | ~39% (AA / comparisons) | Low | ~15–17 GB (Q4) | ~$0.04–0.13 / $0.22–0.40 | Apache 2.0; MoE (~3.8B active); efficient |
Nemotron 3.5 Super VL (120B A12B) | 26 | ~60.7% (vendor); up to ~72% with scaling | ~10–34% (setup-dependent) | ~69–90 GB (Q4) or ~65–75 GB (NVFP4) | — (self-host / partners) | OpenMDW; high-end / multi-GPU or DGX Spark class; vision |
Reflection Beam (501B A23B) | 30 | ~80.1% (vendor BF16) | ~43–44% | High (multi-GPU / 48–96+ GB class for practical Q4) | ~$0.25–0.40 / $1–1.20 | Apache 2.0; MoE (23B active); coding/agent focused; weaker on some terminal hard sets |
Muse Spark Open (816B A42B) | ~39 | ~76.4% | ~60 | High (server / multi-GPU) | $1.25 / $4.25 | Muse Model License (open weights with commercial limits); strong agentic |
Notes: Actual VRAM depends on quant, context length, KV cache, vision projector, and runtime overhead. Muse Glimmer and Gemma variants are the most practical single-consumer-GPU options. Nemotron Super and Beam are server/high-end or multi-GPU territory. Weights are free; hosted prices are typical API rates where offered. Always verify current quants (Unsloth, official GGUF/NVFP4) and your hardware.
Sources include Artificial Analysis, BenchLeader, provider cards, OpenRouter, and VRAM calculators (as of late Oct 2026). Numbers are directional—test on your workloads.