OpenCode Go: bang for buck

Artificial Analysis quality × OpenCode Go usage limits  ·  $10/mo subscription  ·  generated 2026-09-24 17:41

The checkbox is on by default, showing versions without the flagged models — untick to show them.

BANG is this report's composite score: quality0.7 × quota-value0.3, where quality is the raw AA intelligence index with a smooth convex stretch — 57 vs 73 counts as much worse, not 22% worse, and weak scores shrink toward zero without ever zeroing out — and quota-value is 100·req/(req+K) with K=8k req/mo and K=1.6k req/5h, bottlenecked as the lower of the two windows, peaking around 60k req/mo (12k req/5h) and shedding a few points per 10× beyond that: absurd quotas score a touch below merely-generous ones instead of winning forever. A model only scores high if it is good first and sufficiently metered. It is a heuristic for this specific $10/mo subscription, not an official benchmark.

Quality vs. usage limit — the money graph

x-axis: your monthly request quota (log scale), y-axis: AA intelligence index. Bubble size = included usage value for that model (per-model $ tier). Blue dots lie on the Pareto front: no other model is both smarter and gives you more requests, so the dashed line marks the rational choices. Top-right is the sweet spot.

Bang-for-buck ranking

BANG = quality^0.7 × quota-value^0.3, where quality is the raw AA intelligence index with a smooth convex stretch (100·(q/100)², so 57 vs 73 counts as much worse, not 22% worse, and nothing ever zeroes out) and quota-value = 100·req/(req+K) peaking around 60k/mo (12k/5h, bottleneck of both windows), then shedding a few points per 10× beyond the peak — absurd quotas score a touch below merely-generous ones instead of winning forever. Blue bars are also Pareto-optimal in the graph above. Use it as a shortlist generator, not gospel — quality deliberately outweighs quota; if you never hit quotas, just read the raw score column.

Full leaderboard — ranked by BANG

Rows without a BANG score have no Artificial Analysis benchmark yet. The pairing between Go model names and AA names is automated, so give the Match column a glance before trusting a score. Click any column header to sort; click again to reverse.

Column glossary

req/5h · req/mo

Requests allowed per 5-hour window / per month. Go limits are defined in dollar value ($12 / 5h, $30 / week, $60 / month) and the docs translate that into per-model request counts using a typical request profile.

$/req

Estimated cost of one typical request at Go token prices (docs' input / cached / output token mix per model).

value/mo

Included usage value at the monthly limit (the per-model $ tier, e.g. $15/$30/$60/$100) and its multiplier over the subscription price.

tokens/mo

Roughly how many tokens/month the request quota allows (req/mo × typical tokens per request). Cached-read tokens dominate this figure.

AA intell. / tok/s

Artificial Analysis intelligence index and median output speed (tokens/s). Higher is better; the index goes 0–100 and measures general reasoning ability.

BANG

The bang-for-buck score explained above (0–100, higher is better value). Quality enters as the raw AA intelligence index, quota as the saturating quota-value (bottleneck of both windows).

Match

Confidence the AA scores shown belong to this exact model: exact name match, fuzzy (name variant, e.g. publisher prefix), none (no AA data yet, or only an unverified guess which is never scored).

Privacy

From the Go docs Privacy table: whether the provider trains models on your prompts, plus data retention. The checkbox above (on by default) swaps charts, cards and leaderboard rows to versions without the flagged trains models.