AI Letters #05 · Evidence Dashboard

The Agent Performance Numbers

Benchmarks, cost curves, and the data behind the agentic AI transition.

SWE-Bench: Agents Solving Real GitHub Issues
SWE-Bench presents real GitHub issues and evaluates if the model's code fix passes the actual test suite. No simulations.
Jul '23
1.96% · GPT-4 (no scaffolding)
Jan '24
13.9%
13.9% · SWE-agent + GPT-4
May '24
22.0%
22.0% · Devin (Cognition)
Aug '24
41.0%
41.0% · Claude 3.5 Sonnet + scaffolding
Oct '24
49.0%
49.0% · Claude 3.5 Sonnet (Oct release)
Early '25
50%+
50%+ · Multiple frontier systems
📊 Key Insight
From 2% to 49% in 14 months is not a linear improvement — it is a phase transition. The same trajectory that moved SWE-Bench from 2% to 49% is now beginning on harder benchmarks (WebArena, OSWorld). Engineers who understand the architecture behind these numbers are 18 months ahead of the curve.
14mo
2% → 49%
SWE-Bench solve rate
12K+
Real GitHub issues
in SWE-Bench
100%
Real test suites
used for evaluation

Sources: SWE-bench.com leaderboard · Princeton NLP · Cognition AI · Anthropic

Model Comparison: Agentic Task Performance
Success rate, steps to completion, and true cost for complex multi-step agent tasks. Hover bars for details.
⚠️ The Cheap Model Trap
GPT-4o-mini costs ~4× less per token than GPT-4o. But on complex agentic tasks, it fails 49% of the time (vs ~28% for GPT-4o) and averages 6.1 steps when it doesn't fail (vs 4.2). Total cost per completed task is higher with the cheaper model — before accounting for cost of failure in production.
78%
Claude 3.5 Sonnet
complex task success
More retries with
cheap models
3.8
Avg steps (Claude)
vs 6.1 (mini)

Sources: Berkeley Function Calling Leaderboard · Internal benchmarks · HELM

Cost vs Performance: The Frontier Model Case
True cost per completed task (including failures and retries) plotted against task success rate. Hover points for model details.
💡 The Counter-Intuitive Finding
The models that appear cheapest at the token level are often the most expensive at the task level. A model with 50% success that costs $0.04/task actually costs $0.08/task when you account for retries — before adding the cost of failed tasks that require human remediation (typically $5-50 in engineering time). The frontier model that costs $0.20/task but succeeds 78% of the time is the better economic choice for most production agents.
$0.20
True cost/task
Claude 3.5 Sonnet
$5-50
Human remediation cost
per failed agent task
78%
Success rate threshold
for production viability

Sources: OpenAI pricing · Anthropic pricing · Berkeley evals · Practitioner estimates

www.engineersofai.com · AI Letters #05 · Agentic AI A-Z Series