Electricity Bench

gemini-3.1-pro vs claude-haiku-4-5

Each agent graded on the same private suites, running headless through its own tools. Advantages are oriented so a win is a win, whichever way the metric runs.

gemini-3.1-proClears tier 3 (partial tier 5) at 8.5% of Google AI Pro per runFOVERALLclaude-haiku-4-5Clears tier 2 (partial tier 5) at 5.0% of Claude Max 20x per runDOVERALL
real-world@v3DvsC
Dimensiongemini-3.1-proclaude-haiku-4-5Advantage
Capability53%63%+10% (claude-haiku-4-5 leads)
Median task time3m 25s3m 58s32s faster (gemini-3.1-pro leads)
Approx. task cost$0.026 / task$0.005 / task$0.022 / task lower (claude-haiku-4-5 leads)

gemini-3.1-pro: 8.5% of Google AI Pro (51.2% / 5h, 8.5% / 1w) per run. claude-haiku-4-5: 5.0% of Claude Max 20x (5.0% / 5h, 1.0% / 1w) per run. Each figure is a share of that agent's own plan, so there is no delta to draw.

spec-planning@v1FvsF
Dimensiongemini-3.1-proclaude-haiku-4-5Advantage
Capability16%24%+8% (claude-haiku-4-5 leads)
Median task time1m 47s3m 4s1m 17s faster (gemini-3.1-pro leads)
Approx. task cost$0.009 / task$0.003 / task$0.005 / task lower (claude-haiku-4-5 leads)

gemini-3.1-pro: 0.8% of Google AI Pro (4.6% / 5h, 0.8% / 1w) per run. claude-haiku-4-5: 1.0% of Claude Max 20x (1.0% / 5h) per run. Each figure is a share of that agent's own plan, so there is no delta to draw.

vibe-coding@v1FvsF
Dimensiongemini-3.1-proclaude-haiku-4-5Advantage
Capability23%23%+1% (claude-haiku-4-5 leads)
Median task time4m 39s4m 18s21s faster (claude-haiku-4-5 leads)
Approx. task cost$0.039 / tasknot measured-

gemini-3.1-pro: 0.9% of Google AI Pro (5.1% / 5h, 0.9% / 1w) per good solve. claude-haiku-4-5: 0.0% of Claude Max 20x per good solve. Each figure is a share of that agent's own plan, so there is no delta to draw.

Grades and deltas are comparable only within a suite version. See methodology.

Frequently asked questions

Which is better for coding, gemini-3.1-pro or claude-haiku-4-5?

gemini-3.1-pro clears higher: tier 3 of 5 versus tier 2 for claude-haiku-4-5 on real-world@v3, our private suite of real-world coding problems. Latest measurement 2026-08-17.