SB SubscriptionBench
Real-World Quota Benchmark Updated Nov 19, 2026 · Intelligence scores via Artificial Analysis v5

How Much Coding Value Does Google AI Pro Antigravity Deliver?

We measured real coding sessions on a $19.99/mo Google AI Pro subscription to see how many tasks each model delivers per month—and which model gives you the best intelligence for your quota.

Best Overall Strategy
Neon First
Neon Medium (54.9 intelligence, ~154 tasks/mo) beats Hydrogen xHigh on both intelligence and volume (2.3× more tasks).
Equivalent Monthly API Value
$210–$293 /mo
Roughly 10–15× the $19.99/mo Pro price on Gemini 4 models, plus ~$20/mo in bonus Claude quota.
2-Week Value Jump (Nov 7 → 19)
+11 pts · 3× tasks
Peak intelligence jumped from 50.4 (Argon, ~15 tasks/mo) to 61.4 (Neon xHigh, ~45 tasks/mo) on the same subscription.

Intelligence vs. Monthly Task Capacity

Higher is smarter; further right means more tasks per month from your subscription. Models on the solid line give the best trade-off.

Directly measured Effort-scaled estimate Best-value frontier (Today) Nov 7 frontier
Hover any point for details · Drag to zoom, double-click to reset
Advanced chart options ▾
Key takeaway for Antigravity users: Start with Gemini 4 Neon. Even Neon Low (44.1 intelligence, ~352 tasks/mo) beats Hydrogen Medium (41.2 intelligence, ~263 tasks/mo). Only switch to Hydrogen Low (~718 tasks/mo) if you need maximum volume on simple tasks. Read the full write-up →

Monthly Task Capacity by Model

Estimated coding tasks you can complete per month from each model's included subscription quota (at full weekly usage).

Tasks are standardized to Artificial Analysis v5 difficulty. Separate Claude bonus quota (~$20/mo API value) is included in Google AI Pro alongside the main Gemini quota.

Monthly Quota & API Savings Planner

Most people don't use 100% of their weekly quota due to 5-hour limits. Adjust your realistic usage and model mix to estimate your monthly tasks and what the same work would cost via direct API.

Weekly Quota Usage 75%

Gemini 4 Neon (Medium) · 54.9 intel 50%
Gemini 4 Neon (Low) · 44.1 intel 30%
Gemini 4 Neon (High) · 59.6 intel 20%
Gemini 4 Hydrogen (Medium) · 41.2 intel 0%
Gemini 4 Hydrogen (Low) · 26.9 intel 0%

Your Estimated Monthly Output

Tasks / Month
—
Avg. Intelligence
—
Equivalent API Cost
—
Value vs. $19.99 Pro
—
Want to see Claude Pro/Max and ChatGPT Plus/Pro (Codex) tested next? This project is independent and self-funded. Optional tips on Ko-fi go directly toward paying for the next round of subscription benchmarks.
Full Benchmark Data Table
Inspect all 28 model/effort projections, 11 empirical runs, confidence intervals, and download CSV
Methodology, Calibration & Caveats
How real quota burn is measured, API-equivalent formulas, statistical uncertainty, and known limitations

1. How We Measure Monthly Capacity

We run autonomous multi-turn coding tasks inside isolated containers using the actual Antigravity and Codex harnesses, recording exact token usage and net quota burn. Because task difficulty cancels out of the API-cost-to-quota-burn ratio, we scale each pool's implied monthly API value onto Artificial Analysis Intelligence Index v5 tasks.

Implied Pool $/mo = (API Cost ÷ Quota Burn Fraction) × 4.33 Tasks / Month = Implied Pool $/mo ÷ AA Cost per Task

2. Real Sessions vs. Theoretical Token-Only

We also run 3 synthetic probes (input-heavy, output-heavy, cache-heavy) per model to fit per-token quota weights. On Gemini 4 Neon, real multi-turn Antigravity sessions burn ~15–35% more quota than synthetic weights alone predict—so we default to the real-session numbers as the realistic user experience.

Token-Weighted Tasks/mo = (100 × 4.33) ÷ (w_in·T_unc + w_cache·T_cache + w_out·T_out)

3. Effort Scaling & Confidence Intervals

  • Measured vs. Scaled: Solid dots are directly measured runs; hollow dots are scaled from each family's measured anchor using Artificial Analysis per-effort cost ratios.
  • Uncertainty: 95% confidence intervals use leave-one-out Student-t jackknife on continuous quota meters and exact ±1% quantization bounds on Codex Free.

4. Important Caveats

  • Full Weekly Usage: Chart baselines assume 100% weekly quota utilization without 5-hour cap waste. Use the Quota Planner above to model realistic 50–80% usage.
  • Claude Cache Writes: Antigravity does not separate cache-write tokens from uncached input (~10–13% upper-bound shift on Claude models where cache write = 1.25× input).
  • Open Source: All experiment manifests, calibration scripts, and raw CSVs are available in the GitHub repository.