Agent model setup & recommendations

[mhenke]

Findings: Two threads about which models to use per agent. One is a long community Q&A on quality-first configs (rough consensus: GPT-5.4 orchestrator, Opus oracle, Gemini designer, fast or cheap models for explorer/librarian/fixer; expensive models everywhere slowed output around 10x for little gain; open asks on council fallbacks and missing modes like ULW). The other is an extensive May-2026 model-ranking guide (tiers, tables, pricing) per agent role, shared as a reference.

What was said:

• @someoneinjd (Question / Best Practice: highest-quality per-agent model setup when cost is not a concern #275): proposed a speculative quality-first config (GPT-5.4 orchestrator, Opus oracle, Gemini designer, fast models for the rest), noting it was unverified.
• @alvinunreal (Question / Best Practice: highest-quality per-agent model setup when cost is not a concern #275): orchestrator on a top coding model, oracle on Opus, designer on Gemini, and fast or cheap models (kimi-k2.5-turbo) for explorer/librarian/fixer; pointed to his README preset.
• @rustanacexd (Question / Best Practice: highest-quality per-agent model setup when cost is not a concern #275): confirmed fast models work well for the worker agents and that expensive models everywhere slowed output around 10x for little quality gain.
• @mrjared16 (Question / Best Practice: highest-quality per-agent model setup when cost is not a concern #275): asked whether the council can fall back when a councillor's quota runs out.
• @tlerbao (Question / Best Practice: highest-quality per-agent model setup when cost is not a concern #275): asked about missing modes like ULW and whether the council triggers automatically.
• @FirebirdRender ([INFORMATIONAL] Model ranking/recommendations - May 2026 #498): posted tiered rankings (reasoning, complex, standard, simple) with top-5 closed and open-weight picks per role, plus specialist picks and a quick-reference table; noted rankings shift monthly and are approximate.

Open questions:

1. For quality-first work, which model should each agent use when cost is no object, and do the May-2026 rankings still hold?
2. Is mixing a strong orchestrator with a specialized oracle and designer better than one model everywhere?
3. Can the council add model fallbacks so a quota outage on one councillor doesn't break the run, and should curated rankings live in the docs or community discussions?

Original issues: #275 and #498

[mhenke]

Findings (update): A third thread joins the agent-model-selection discussion, this one scoped to the limited opencode-go plan and focused on cost-vs-benefit plus per-agent fallbacks.

What was said:

• @mhenke (Guidance on choosing agent-specific models (opencode-go plan) — cost vs. benefit and fallbacks #565): asked the community and maintainers for recommended models per agent on the opencode-go plan, where the plan is capped and cost/benefit must be balanced. Wants concrete fallback examples per agent when the primary model fails or hits quota.
• Shared their current config as a reference point: orchestrator = kimi-k2.7-code, oracle = deepseek-v4-pro, council = deepseek-v4-pro, librarian = minimax-m2.7, explorer = minimax-m2.7, designer = kimi-k2.7-code, fixer = deepseek-v4-flash, observer = kimi-k2.7-code.
• Stated latency is not a concern, but good reasoning, solid role performance, and low price matter most.

Open questions (added):
4. On a capped plan like opencode-go, which model per agent gives the best cost/benefit, and what is a safe fallback chain when the primary hits quota?
5. Do the quality-first configs from #275 / the #498 rankings translate to opencode-go, or does the plan roster force different trade-offs?

Original issue: #565

[mhenke]

Findings (update): @MichaelLi423 submitted a community preset (<gpt&go>) that splits models across three providers based on what each agent actually needs.

What was said:

• @MichaelLi423 ([Preset] <gpt&go> #890): posted a preset config using OpenAI, OpenCode Go, and Zen paygo. The idea: reasoning-heavy roles get GPT-5.6 from OpenAI paygo, oracle gets DeepSeek-v4-pro from the Go subscription where deep reasoning justifies the cost, and librarian/fixer run on DeepSeek-v4-flash-free from Zen where cheap and fast is fine.

The preset (<gpt&go>):

Agent	Model	Provider	Rationale
orchestrator	gpt-5.6-terra	openai (paygo)	Strong reasoning for planning/coordination
council	gpt-5.6-sol	openai (paygo)	Multi-model consensus needs a capable synthesizer
oracle	deepseek-v4-pro	opencode-go (subscription)	Deep reasoning for architecture decisions
designer	gpt-5.6-sol	openai (paygo)	UI/UX judgment benefits from a strong model
explorer	gpt-5.3-codex-spark	openai (paygo)	Fast code search, lower-tier GPT is fine
librarian	deepseek-v4-flash-free	opencode zen (free tier)	Web/docs research, cheap model suffices
fixer	deepseek-v4-flash-free	opencode zen (free tier)	Mechanical implementation, no need to spend

This is a different take from the opencode-go-only configs in #565. Instead of picking one provider and working within its roster, the submitter is mixing providers to get the best model per role without paying subscription prices for every agent.

Open questions (added):
6. Does mixing providers create setup friction that outweighs the cost savings, or is it worth the extra config?
7. Should community presets list required providers upfront so people know what accounts they need before trying one?

Original issue: #890

[tlerbao]

I am currently testing the following setup:

Role	Primary model	Fallback(s)	Variant
Orchestrator	openai/gpt-5.6-sol	-	high
Oracle	openai/gpt-5.6-sol	-	max
Librarian	xai/grok-4.5	grok-composer-2.5-fast -> openai/gpt-5.4-mini	-
Explorer	xai/grok-4.5	grok-composer-2.5-fast -> openai/gpt-5.4-mini	-
Designer	gemini-3.6-flash-high	gemini-3.1-pro-high	-
Fixer	xai/grok-4.5	openai/gpt-5.6-sol-fast -> grok-composer-2.5-fast	-
Council synthesizer	openai/gpt-5.6-sol-fast	-	high
Councillor Alpha	anthropic/claude-opus-4-8	-	-
Councillor Beta	openai/gpt-5.6-sol	-	-
Councillor Gamma	xai/grok-4.5	-	-
Fast Generic	grok-composer-2.5-fast	openai/gpt-5.4-mini-fast	-

I am still testing this configuration, so these are only my initial observations.

So far, Composer has been extremely fast, which makes it particularly useful for routine mechanical work and command execution. Grok offers a good balance of capability and speed, so I am currently trying it as the primary model for Explorer, Librarian, and Fixer.

The council intentionally uses three different model families to provide more diverse perspectives, while GPT-5.6 Sol handles orchestration and synthesis.

Does anyone have suggestions for improving this setup, particularly the model assignments, fallback order, or overall cost/performance balance?

[MichaelLi423]

How is work going so far? Maybe I will subscribe ChatGPT Plus & SuperGrok and follow your configuration.

[mhenke]

@ricatix opened #918 with some real numbers on the Luna patch reliability problem, figured it fits in this thread too.

They pulled apply_patch stats from their local OpenCode session DB. Luna is sitting at a 44% failure rate per 100 calls. Terra is under 10%. The gap shows up almost entirely in the fixer lane, which tracks given how patch-heavy that role is.

The current OpenAI preset ships Luna as the default for both fixer and designer. Rico is suggesting either documenting the patch-reliability caveat, switching the fixer default to Terra, or adding telemetry so we can compare properly across models and effort levels.

Anyone else seeing the Luna patch failures on their end? Even rough numbers from session history would help confirm whether this is widespread or local to one environment.

[mhenke]

Related: PR #938 from @binyangzhu000-sudo adds an Atlas Cloud preset backed by deepseek-v4-pro.

It registers Atlas as an OpenAI-compatible provider (key from ATLASCLOUD_API_KEY, endpoint https://api.atlascloud.ai/v1), generates the preset on install, and merges provider defaults without overwriting your existing config. All six lanes (orchestrator, oracle, librarian, explorer, designer, fixer) default to the same model.

'atlas-cloud': {
    orchestrator: { model: 'atlas-cloud/deepseek-ai/deepseek-v4-pro' },
    oracle:     { model: 'atlas-cloud/deepseek-ai/deepseek-v4-pro' },
    librarian:  { model: 'atlas-cloud/deepseek-ai/deepseek-v4-pro' },
    explorer:   { model: 'atlas-cloud/deepseek-ai/deepseek-v4-pro' },
    designer:   { model: 'atlas-cloud/deepseek-ai/deepseek-v4-pro' },
    fixer:      { model: 'atlas-cloud/deepseek-ai/deepseek-v4-pro' },
},

Compare that to the OpenAI preset which differentiates (e.g. gpt-5.6-terra for orchestrator, gpt-5.6-sol for oracle, gpt-5.6-luna for fixer). The Atlas preset is flat, one model everywhere. That is accurate to what I posted in the discussion comment.

Anyone using this provider? Any feedback?