Frontier Model & Agent Harness Cost-Effectiveness Curves
Interactive Nov 2026 Pareto analyzer comparing 32 frontier LLMs (Claude 5.5 / Fable 5.2, GPT-6 / 6.1, Gemini 4, Qwen4 / 3.8, DeepSeek-V4.1, Kimi-K3.1, GLM-5.3) and 6 coding agent harnesses.
Models positioned higher and further left deliver more intelligence per dollar. Green dashed line connects non-dominated Pareto optimal models.
Frontier Model Pareto Leaderboard
| Model & Rates | Score | Blended $/1M | Tools |
|---|---|---|---|
GPT-6 Astra★ Pareto OpenAI • $10/$50 per 1M | 95.6 | $32.30 | |
Claude Fable 5.2★ Pareto Anthropic • $8/$40 per 1M | 95.4 | $24.66 | |
Claude Opus 5.5★ Pareto Anthropic • $4/$20 per 1M | 94.2 | $10.86 | |
Gemini 4 Carbon★ Pareto Google • $2/$10 per 1M | 94 | $4.80 | |
Qwen4-Max★ Pareto Alibaba • $1.5/$4.5 per 1M | 91.5 | $2.27 | |
DeepSeek-V4-Pro★ ParetoOpen-Weight DeepSeek • $1.32/$3.96 per 1M | 90.6 | $2.04 | |
Mistral Large 4★ ParetoOpen-Weight Mistral • $0.7/$2.1 per 1M | 88.8 | $1.04 | |
Mistral Large 3★ ParetoOpen-Weight Mistral • $0.5/$1.5 per 1M | 84.8 | $0.69 | |
DeepSeek-V4.1-Flash★ ParetoOpen-Weight DeepSeek • $0.18/$0.72 per 1M | 84.4 | $0.32 | |
MiniMax-M3.1-Flash★ ParetoOpen-Weight MiniMax • $0.15/$0.6 per 1M | 82.2 | $0.25 | |
GLM-5.3-Flash★ ParetoOpen-Weight Zhipu • $0.15/$0.5 per 1M | 80.4 | $0.22 | |
Claude Haiku 5.5★ Pareto Anthropic • $0.1/$0.5 per 1M | 78.4 | $0.19 | |
GPT-6 Luna★ Pareto OpenAI • $0.1/$0.5 per 1M | 77.8 | $0.19 | |
Gemma 4 31B★ ParetoOpen-Weight Google • $0.12/$0.36 per 1M | 76.2 | $0.17 | |
GPT-6.1 Sol OpenAI • $2/$10 per 1M | 91.2 | $4.68 | |
Kimi-K3.1-ProOpen-Weight Moonshot • $2/$10 per 1M | 91 | $4.71 | |
Claude Sonnet 5.5 Anthropic • $2/$10 per 1M | 90.8 | $4.55 | |
Gemini 4 Argon Google • $2/$10 per 1M | 90.4 | $4.30 | |
Grok 4.7 xAI • $2/$6 per 1M | 90.2 | $3.30 | |
Qwen3.8-Max (2.4T)Open-Weight Alibaba • $2/$6 per 1M | 89.6 | $3.15 | |
GLM-5.3Open-Weight Zhipu • $1.4/$4.4 per 1M | 88.4 | $2.19 | |
GPT-5.6 Sol OpenAI • $4/$20 per 1M | 88.2 | $9.92 | |
Gemini 4 Hydrogen Google • $1.5/$7.5 per 1M | 88.1 | $3.13 | |
GPT-6.1 Terra OpenAI • $1/$5 per 1M | 86.8 | $2.15 | |
Claude Sonnet 4.6 Anthropic • $3/$15 per 1M | 84.6 | $6.69 | |
Gemini 3.8 Flash Google • $0.75/$3.75 per 1M | 83.5 | $1.49 | |
Qwen3.8-Flash-NextOpen-Weight Alibaba • $0.2/$0.8 per 1M | 81.8 | $0.34 | |
DeepSeek-V4-FlashOpen-Weight DeepSeek • $0.3/$1.2 per 1M | 80.8 | $0.48 | |
Llama 4 MaverickOpen-Weight Meta • $0.22/$0.75 per 1M | 76.8 | $0.30 | |
GPT-OSS-120BOpen-Weight OpenAI • $0.25/$0.69 per 1M | 75.2 | $0.36 | |
Gemini 3.5 Flash-Lite Google • $0.3/$2.5 per 1M | 74.5 | $0.80 | |
DeepSeek-V3Open-Weight DeepSeek • $0.27/$1.1 per 1M | 73.5 | $0.42 |
Reading Nov 2026 Frontier Model & Agent Harness Pareto Curves
Headline benchmark scores alone hide how much reasoning tokens and multi-file agent exploration multiply real engineering bills.
1. The Late-2026 Cost-Intelligence Shift
Ultralight $0.10-$0.18/1M tiers (Claude Haiku 5.5, GPT-6 Luna, GLM-5.3-Flash, DeepSeek-V4.1-Flash) and $2/1M flagships (Claude Sonnet 5.5, GPT-6.1 Sol, Gemini 4 Carbon, Qwen4-Max) pushed frontier coding above 90 pts at 10x lower cost than 2025.
2. Harness Choice Changes Cost 3x-5x
On identical base models, Claude Code, Codex CLI, Antigravity CLI, OpenHands, Cursor Agent, and Aider make very different context-indexing vs retry trade-offs on Terminal-Bench 4.1.
3. Watch Cost per Solved Task, Not Token Price
A cheap stack resolving 38% of bugs can cost more per merged PR than Sonnet 5.5, GPT-6.1 Sol, or Gemini 4 Carbon resolving 61%-65% cleanly on the first pass.