Head to head
vs ME For light moderate heavy
Verdict · measured basis
Llama 3.2 3B Instruct costs $0.0001 per solved task against
gpt-oss-safeguard-20b at $0.0002 — ▼ 1.4x cheaper .
Over 200 tasks that is $0.01 a month.
On output token price alone gpt-oss-safeguard-20b looks 1.1x cheaper. The ranking reverses once you divide by pass rate — the cheaper tokens do not win the task.
Copy link Save scenario Export compare table Planned
Verified pricing, pass rate and cost per solved task for gpt-oss-safeguard-20b and Llama 3.2 3B Instruct Figure gpt-oss-safeguard-20b Llama 3.2 3B Instruct Input $/Mtok $0.075 $0.05 Output $/Mtok $0.3 $0.33 Cached input $/Mtok $0.037 $0.05 Pass rate 97% 33% Cost per solved task $0.0002 measured $0.0001 measured Cost per month $0.04 $0.03
Cost per month against tasks per month · drag the marker
Monthly cost against tasks per month Log-log lines for Llama 3.2 3B Instruct (measured, $0.0001 per solved task), gpt-oss-safeguard-20b (measured, $0.0002 per solved task) from 10 to 100,000 tasks a month. Marker at 200 tasks. 10 100 1k 10k 100k $0.001 $0.01 $0.1 $1 $10 TASKS / MONTH (log) → 200 tasks Llama 3.2 3B Instruct · MEASURED · $0.0001 per solved task $0.03 Llama 3.2 3B Instruct MEASURED gpt-oss-safeguard-20b · MEASURED · $0.0002 per solved task $0.04 gpt-oss-safeguard-20b MEASURED Monthly cost against tasks per month Log-log lines for Llama 3.2 3B Instruct (measured, $0.0001 per solved task), gpt-oss-safeguard-20b (measured, $0.0002 per solved task) from 10 to 100,000 tasks a month. Marker at 200 tasks. 10 100 1k 10k 100k $0.001 $0.01 $0.1 $1 $10 TASKS / MONTH (log) → 200 tasks Llama 3.2 3B Instruct · MEASURED · $0.0001 per solved task $0.03 gpt-oss-safeguard-20b · MEASURED · $0.0002 per solved task $0.04 — Llama 3.2 3B Instruct · MEASURED — gpt-oss-safeguard-20b · MEASURED Solid measured · dashed modeled · dotted stale. The gap is the verdict.
Solvency Bench (solvency.dev) — first-party measurement · verified 2026-08-26 · prices from each provider's pricing page; measured rows ignore the tier · Methodology · All models at these settings