Choosing models
Snapshot: September 2026. Plans change; confirm on the vendor’s site before you buy or reconfigure.
Start here if you still need to choose a channel. Copy-paste wiring is in Model configuration recipes and Examples. Those snippets use a current flagship so every field is filled in; swap the model IDs after you decide.
Will what you already pay for reach Chord?
Section titled “Will what you already pay for reach Chord?”Chord accepts an API key, or Codex OAuth (chord auth codex). A subscription that only signs you into that vendor’s own app does not.
| You already have | In Chord? |
|---|---|
| A coding plan that supports third-party clients (for example ChatGPT Plus / Pro (Codex), Command Code GOAT, OpenCode Go) | Yes |
| An API key for any provider Chord speaks (Chat Completions, Responses, Messages, Generate Content) | Yes |
| Claude Pro / Max, SuperGrok, Muse Code, Google AI Pro / Ultra, or another vendor-only coding plan | No |
| GLM Coding Plan | No (only in the tools and products GLM officially supports) |
Four ways to decide
Section titled “Four ways to decide”1. You already subscribe to something
Section titled “1. You already subscribe to something”If the table says Yes, use it. If it says No, leave that app alone and continue with case 2, 3, or 4. On Codex, one plan fills all five roles: orchestrator, explorer, and coder on the cheapest model your account lists; expert and reviewer on the strongest model the plan can sustain. For a cheap plan or an API key, follow case 2.
2. Light use, and you want value
Section titled “2. Light use, and you want value”Get a cheap official API key (for example Gemini 3.8 Flash, DeepSeek V4.1 Flash, or GPT-6 Luna), or a cheap plan that exposes a standard endpoint (for example Command Code GOAT or OpenCode Go). Use it for routing, search, most edits, and all five roles. Upgrade expert and reviewer only when hard decisions become common; if the cheap plan does not carry a strong enough model, add a separate official key for them.
3. Heavy use
Section titled “3. Heavy use”A flat subscription beats per-token billing when you use it every day. Subscribe to Codex and sign in: orchestrator, explorer, and coder on the cheapest model your account lists; expert and reviewer on the strongest model the plan can sustain.
4. Bill size is not the issue
Section titled “4. Bill size is not the issue”Give the strongest API models you can get (GPT-6.1 Sol and Claude Opus 5.5 as of this snapshot) to expert and reviewer, and keep GPT-6 Astra and Claude Fable 5.1 for architecture design and for work that is unusually hard or has already failed. Keep a cheap fast model on orchestrator, explorer, and coder; they do not need a flagship.
Which model goes to which role?
Section titled “Which model goes to which role?”Routing and mechanical edits do not need a flagship; a cheap model handles them. Split pools by job: one for cheap, fast models, one for the flagship you will pay for; the team example calls them fast and deep.
The names below are not built-in. Chord ships builder and planner. The five-role split is the optional team example. Copy it if you want that layout.
| In the team example | Suggested model | Because |
|---|---|---|
| orchestrator | Gemini 3.8 Flash, DeepSeek V4.1 Flash, GPT-6 Luna | It classifies, dispatches, and synthesizes every turn. |
| explorer | DeepSeek V4.1 Flash, Gemini 3.8 Flash, GPT-6 Luna | Read-only scouting; it reports where files are and makes no judgment calls. DeepSeek has the cheapest cache reads, Gemini reads material better. |
| coder | DeepSeek V4.1 Flash, Gemini 3.8 Flash, GPT-6 Luna | What to change and how is already written down; mechanical edits are within reach of any of them. |
| expert | Claude Opus 5.5, GPT-6.1 Sol | Root cause, architecture, concurrency, and hot paths; a wrong call becomes hidden debt. |
| reviewer | Claude Opus 5.5, GPT-6.1 Sol | Catches regressions and invariant breaks; it does not redesign. |
Architecture design and root-cause work are the judgment-heavy end of expert. Fable 5.1 (Anthropic) and GPT-6 Astra (OpenAI) still have the strongest record there, so treat them as the first choice for design-level calls. The two defaults above carry the rest of expert work; escalate to Fable 5.1 or Astra when they stall.
If you only have a subscription and no separate API keys, pick the same two levels from that plan’s catalog: the cheapest model for orchestrator, explorer, and coder, and the strongest the plan can sustain for expert and reviewer. On Codex that is GPT-6 Luna and GPT-6 Astra (use GPT-6.1 Sol when the account does not include Astra).
In practice, split pools by job: deep holds the expert and reviewer models and fast holds the explorer and coder models. The team example keeps orchestrator on deep; in real use it needs no flagship, so fast or its own cheap pool works too.
Which model for retrieval?
Section titled “Which model for retrieval?”- Files and code inside the repo: DeepSeek V4.1 Flash. Read-only scouting does not need closed-book knowledge, and its cache reads are the cheapest, which suits re-reading the same files.
- Web material, PDFs, charts: Gemini 3.8 Flash. Long PDFs and charts are where it is strongest; DeepSeek is weak closed-book, so retrieval has to come from a search tool rather than its memory.
- Codex subscription only: use GPT-6 Luna for repo scouting; narrow the range first on very large repos.
- Need both and want a single model: use Gemini 3.8 Flash.
Can coder use a cheap model?
Section titled “Can coder use a cheap model?”Yes, and it should by default. Coder is for changes that are already decided: renames, mechanical refactors, format and config updates, small local fixes, tests behind a fixed interface. Judgment stays with expert, and for this kind of work DeepSeek V4.1 Flash, Gemini 3.8 Flash, and GPT-6 Luna are all enough, for far less money than a flagship.
It is not for work that still needs judgment: an open “why” or “which approach”, a change to protocol, data models, concurrency or lifetimes, permissions, or recovery, or a task that has already failed twice. System-level work in an unfamiliar environment (new language, new build system, inside a container) needs its path and acceptance criteria pinned down first, and of the three, DeepSeek V4.1 Flash is the weakest there.
When the cheap tier keeps coming back with execution mistakes (multi-file edits against a fixed spec, bug fixes spread across files, fast iteration on an existing feature), Claude Sonnet 5.5 is the middle option. It is the faster complement to Opus 5.5, priced far below the Opus-class flagships, with agentic coding near their level at higher effort and a much lower cost per task at lower effort. Open-ended judgment still belongs to expert.
No GPT or Claude subscription — what should expert use?
Section titled “No GPT or Claude subscription — what should expert use?”Neither vendor requires a subscription: both sell API keys, so Opus 5.5 and Sol stay on the table without a ChatGPT or Claude plan. If you want to stay off both vendors entirely, start at item 3.
- Claude Opus 5.5: pay as you go with an Anthropic API key. It lands near Fable 5.1 on most work for much less, and its cache reads suit re-reading the same files; for architecture design and other calls you would not want a cheaper model to get wrong, step up to Fable 5.1.
- GPT-6.1 Sol: buy an OpenAI API key on its own. It is the GPT-6 generation’s coding and agentic model at a fraction of Astra’s price, with the same 1.05M window and cheap cache reads; use it for coding, debugging, and agentic implementation work, and step up to GPT-6 Astra for architecture design and root-cause judgment; stay on Astra where a relay does not list Sol yet.
- Muse Spark 1.3 (Meta Model API): the strongest model outside the GPT and Claude channels. Long-horizon implementation and large-repo work are its strengths; its root-cause and architecture judgment is a notch lower, so split expert work smaller and verify more.
- GLM-5.3 or Kimi K3: the strongest open models, carried by open-model plans such as OpenCode Go and Command Code GOAT. They can hold expert work, but not the final word on architecture or concurrency.
- None of these: keep the expert question small (have a cheap model reproduce it and narrow the range) and revisit the paid options above when a wrong call would become hidden debt. Gemini 3.8 Flash can hold a first discussion; do not let it be the final reviewer.
Reviewer runs the same model as expert. With only one flagship budget, give it to expert first and open reviewer after a substantial change.
What does orchestrator need?
Section titled “What does orchestrator need?”It runs every turn: read the task, classify it, dispatch, collect results, decide between correcting and escalating. It needs to:
- call tools reliably and read a worker’s report well enough to restate the conclusion;
- tell whether a task still needs a product decision: if yes, send expert; if the path and the replacement are already written down, send coder; if it is only about where files are, send explorer;
- stay cheap and fast. Flagship reasoning and taste are wasted here; the only risk worth paying to avoid is misrouting.
Default to Gemini 3.8 Flash, DeepSeek V4.1 Flash, or GPT-6 Luna; DeepSeek has the cheapest cache reads, and Luna’s uncached input and output are lower. Upgrade only if you observe frequent misrouting; do not start on a flagship.
After you decide
Section titled “After you decide”- Copy the matching snippet from Model configuration recipes.
- For a full file layout, start from Examples and replace the flagship IDs with what you actually picked.
- Confirm with
chord doctor models.