opencode-go Value: best model per agent role

Picks by role

Score is the average of AA intel and LMArena, each scaled to 0 to 100. A ~ marks an Arena entry mapped to the closest match.

Orchestrator

n/a

n/a score

req/5h: n/a
$/month: n/a
AA intel: n/a
LMArena: n/a

Many small turns over long sessions. So it gets the largest req/5h of all models with AA intel 38 or better.

Dev

n/a

n/a score

req/5h: n/a
$/month: n/a
AA intel: n/a
LMArena: n/a

Runs the most calls per task. Frequency beats peak, so the pick maximises AA speed times req/5h at AA intel 34 or better.

Reviewer

n/a

n/a score

req/5h: n/a
$/month: n/a
AA intel: n/a
LMArena: n/a

About one call per task, then a long report. Highest AA intel at req/5h 1000 or better, from a family no other role uses.

Researcher

n/a

n/a score

req/5h: n/a
$/month: n/a
AA intel: n/a
LMArena: n/a

Reads a lot, writes little. Cheapest AA cost per task at req/5h 3000 or better and AA intel 34 or better.

All models

Click a header to sort. Score averages AA intel and LMArena. Grey rows have no AA data. Free models stay listed but never get picked.

JavaScript builds this table from the local data.

Methodology and notes

Role matrix

Orchestrator: largest req/5h among models with AA intel at least 38.
Dev: highest AA speed times req/5h, at AA intel at least 34 and a monthly limit of at least $30.
Reviewer: highest AA intel at req/5h at least 1000, in a family no other role uses.
Researcher: cheapest AA cost per task at req/5h at least 3000 and AA intel at least 34, never the dev model.

Score

The Score column blends both benchmark sources. Each source is scaled to 0 to 100 between the lowest and the highest value among the listed models, so both weigh the same. The scaled values are then averaged. A missing source is skipped, never guessed.

Notes

A family is the letter prefix of a model id. req/5h is the estimate from the docs, not a hard limit. Where a model has price tiers, the base tier counts.

Daily prompt

Copy it, run it once a day in any opencode setup, get fresh picks.

refresh-prompt

Show prompt

Re-evaluate the opencode Go models, $10 plan, NOT Go Plus, and recommend the best
model per agent role.

Sources, primary only, never guess:
- https://opencode.ai/docs/go/  -> the Go column of the pricing and requests tables
  only, not Go Plus. Read req/5h, monthly limit in USD, in/out/cached prices in $/1M.
- https://opencode.ai/inference/go/v1/models  -> full model ids.
- https://artificialanalysis.ai/models/<slug>  -> Intelligence Index, Cost per Task in USD,
  output speed in t/s. If the AA page is missing with a 404, mark it "no AA data", do not guess.
- https://lmarena.ai/leaderboard/text  -> LMArena Overall score, latest ranking. If the
  entry is missing, mark it "missing" and do not guess.

Role matrix:
- orchestrator: largest req/5h among models with AA intel >= 38.
- dev: maximise AA speed times req/5h at AA intel >= 34 and monthly limit >= $30.
- reviewer: highest AA intel at req/5h >= 1000 in a model family not used by
  orchestrator, dev or researcher, where a family is the leading letters of the model id.
- researcher: lowest AA $/Task at req/5h >= 3000, AA intel >= 34, and not the dev model.

Models flagged as free are excluded: limited-time only, unstable and
timeout-prone, no own AA slug.

Score: compute a combined score per model as the mean of the scaled AA and
LMArena values. Scale each source to 0 to 100 between the lowest and highest
value among the listed models. Mark mapped and missing Arena values in the table.

Output:
1. Role table: role | model id | rationale | key numbers, meaning req/5h, $/month, AA intel, AA speed, AA $/Task.
2. Full model table with every Go model; mark free models as "excluded".
3. Call out missing AA data explicitly as "no AA data".