LLM Reference

The agents leaderboard · for developers

Best for agents

6 editor picks · 19 eligible models · Tool-use loops that don't crash on turn 12.

Editorial pick plus benchmark and API pricing context.

See raw /best
EDITOR'S CHOICEResearched 6d ago

Claude Sonnet 5

Anthropic · 1m context
Excellent

The most agentic Sonnet yet — strong computer use and multi-step recovery at Sonnet pricing.

Anthropic's new default agent: 85.2% SWE-bench Verified, 81.2% OSWorld-Verified, 86.6% BrowseComp multi-agent, and 1M context at the same durable $3/$15 as Sonnet 4.6 ($2/$10 intro through 2026-08-31).

Open model
The numbers
$/1M out
$10.00
$2.00 input
Context
1m
max window
Pros
  • +81.2% OSWorld-Verified
  • +86.6% BrowseComp multi-agent
  • +1M context at Sonnet list price
Cons
  • −No τ-bench or MultiChallenge row yet
  • −$15 / 1M out at list

Also worth picking

The runners-up

ranked by editorial pick orderEditorial tiersExcellentStrongSolid
#2
Claude Sonnet 4.6
Anthropic · 1m
$15.00 / 1M out
Best generally-available τ-bench (87.5) among predecessors — solid fallback until Sonnet 5 publishes comparable agent-bench rows.
#3
Claude Fable 5
Anthropic · 1m
$50.00 / 1M out
OSWorld-Verified 85.0% and 80.3% SWE-bench Pro make it a top-tier computer-use and coding-agent candidate, but no Fable-specific τ-bench score is published yet.
#4
GLM-5
Zhipu AI · 200k
$2.08 / 1M out
τ-bench 82.1, open weights at $2.08 out — the best price-per-step agent we'd run always-on.
#5
GPT-5.4
OpenAI · 1.05m
$15.00 / 1M out
Picked over GPT-5.5 here: it carries the measured τ-bench score (78.3) and runs at half the output price ($15 vs $30) — decisive for always-on agents.
#6
Gemini 3 Pro
Google DeepMind · 1m
$5.00 / 1M out
Best for chatty agents with many read-only tools and large contexts.

Eligibility

19 models are eligible for this board

Eligibility means tagged with useCases: [agents]. Pins must come from this pool.

All picks