The Agentic Leaderboard
Agentic composite (0-100), verified 2026-08-28 via OpenRouter's Artificial Analysis feed, with loop economics for a 20-turn session (2K in / 500 out per turn):
| Model | Agentic | Output $/1M | 20-turn session | Agentic per $ | Best for |
|---|---|---|---|---|---|
| Claude Opus 5 | 59.2 | $25.00 | $0.45 | 2.4 | Hardest agent loops |
| GLM 5.3 | 59.1 | $4.40 | $0.10 | 13.4 | Best balanced-tier agent |
| Grok 4.6 | 58.7 | $6.00 | $0.14 | 9.8 | Frontier loop value |
| GPT-5.6 Sol | 57.8 | $30.00 | $0.50 | 1.9 | Reasoning-heavy orchestration |
| Qwen 3.8 Max | 58.4 | $6.00 | $0.14 | 9.7 | Agent + coding combo |
| DeepSeek V4 Pro (off-peak) | 49.6 | $1.98 | $0.05 | 25.1 | Budget production loops |
| Claude Sonnet 5 | 49.7 | $10.00 | $0.18 | 5.0 | Reliable mid-tier default |
| DeepSeek V4 Flash (off-peak) | 33.7 | $0.66 | $0.02 | 51.1 | High-volume simple loops |
Two columns tell the story. Quality: the agentic index is remarkably flat at the top — Opus 5, GLM 5.3, Grok 4.6 and Qwen 3.8 Max all sit between 58.4 and 59.2. Economics: that 1-point spread costs between $0.10 (GLM) and $0.55 (Opus) per session. GLM 5.3 is the rational default for most agent stacks; Opus earns its premium only on loops where a failed turn costs real money.
Why Loops Amplify Model Choice
The loop multiplies every per-turn difference. Three effects to model before you pick:
Output compounds
Each turn bills output; a 20-turn loop bills 20 completions plus reasoning tokens on reasoning models. The output rate matters 20× more than in chat.
Failures re-bill context
A 10% tool-failure rate re-sends full context on every failure — the hidden tax that makes agentic quality (not just price) a cost variable.
Caching is the loop's best friend
System prompt + tool schemas are byte-stable across turns — 90% cached input on most providers makes the input line almost disappear and the output rate the entire decision.
The full compounding math — runaway loops, turn anatomy, ceiling tables — is in the agent loop cost guide.
The 2026 Agent Stack
Orchestrator: GLM 5.3 or Grok 4.6
Agentic 59 at $4.40-6.00 output. The planning loop runs here — high reliability, sane bill.
Execution: DeepSeek V4 Flash off-peak
Mechanical steps, tool calls, data transforms — 33.7 agentic is plenty for well-specified subtasks at $0.66 output.
Escalation: Opus 5 or Sol
When the orchestrator hits its confidence floor — the 10% of turns worth $0.55 instead of $0.02.
Reasoning islands: Sol / o3-style only where needed
Planning steps that genuinely need deduction. Everything else runs thinking-off to avoid the reasoning-token tax.
Frequently Asked Questions
What makes a model good for agentic workflows?
Three things: the agentic composite (tool-use reliability and multi-step planning), cheap output (loops generate output every turn), and cache-friendly input (the system prompt and tool schemas repeat). Coding score matters less than loop discipline.
Which model is best for agents in 2026?
GLM 5.3 leads the agentic index among balanced models (59.1) at $4.40 output — the best agentic-per-dollar on the market. GPT-5.6 Sol (57.8) and Claude Opus 5 (59.2) lead the frontier, and DeepSeek V4 Pro off-peak is the budget loop champion.
How much does an agent session cost per model?
A 20-turn loop (2K in/500 out per turn): $0.50 on GPT-5.6 Sol, $0.45 on Claude Opus 5, $0.18 on Claude Sonnet 5, $0.10 on GLM 5.3, $0.05 on DeepSeek V4 Pro off-peak, and $0.02 on DeepSeek V4 Flash off-peak — before caching, which cuts the input share ~90%.
Should I use a reasoning model for my agent?
Only for the sub-tasks that need it. Reasoning tokens bill at output rates and compound over loop turns — a reasoning-heavy agent can burn 3-5x the output of a standard one. Route: reason for planning steps, fast models for execution steps.
What is the cheapest viable agent stack?
DeepSeek V4 Flash off-peak with a cached system prompt, strict turn caps, and a GLM 5.3 or Sonnet 5 fallback for the hard turns. In our 20-turn example that's about $0.02-0.10 per session — $200-1,000/month at 10K sessions.