How the Model Picker decides
Every recommendation traces back to two plain data tables — no hidden scoring, no fabricated winners. This page is generated directly from those tables, so it can never drift from what the tool actually does.
Back to Model Picker1. Task/stakes floors
For each combination of task type and quality stakes, the matrix names a minimum capability tier (a "floor") and a plain-English reason. Realtime latency and Batch API eligibility add rationale bullets on top of the floor — they never silently change which tier is required.
| Stakes | Required family (this generation's names) | Why |
|---|---|---|
| Internal draft | HAIKU | Classification/extraction is a narrow, well-specified task — Haiku's accuracy is sufficient and it's the cheapest option. |
| Customer facing | HAIKU | Still a narrow task even when customer-facing — Sonnet would be overkill for straightforward classification/extraction. |
| High stakes | SONNET | When errors are costly (legal, financial, medical), the extra accuracy margin from Sonnet is worth the cost even for a narrow task. |
| Stakes | Required family (this generation's names) | Why |
|---|---|---|
| Internal draft | HAIKU | Internal draft replies tolerate occasional rough edges — Haiku is fast and cheap for this. |
| Customer facing | SONNET | Customer-facing conversation needs steadier tone and reasoning than Haiku reliably provides. |
| High stakes | OPUS | High-stakes support (guidance with legal, financial, or medical consequences) warrants Opus's stronger reasoning over Sonnet's baseline. |
| Stakes | Required family (this generation's names) | Why |
|---|---|---|
| Internal draft | HAIKU | Internal drafts (notes, rough copy) are fine with Haiku — a human edits before it goes anywhere external. |
| Customer facing | SONNET | External-facing content needs the baseline quality Sonnet provides — Haiku content tends to read as generic or rough. |
| High stakes | OPUS | High-stakes content (legal, financial, or medical copy) needs Opus's stronger reasoning to avoid costly errors. |
| Stakes | Required family (this generation's names) | Why |
|---|---|---|
| Internal draft | SONNET | Coding needs Sonnet-level reasoning even for internal/draft work — Haiku is not reliable enough to produce correct code unsupervised. |
| Customer facing | SONNET | Sonnet remains the baseline for coding — customer-facing does not by itself raise the bar further. |
| High stakes | OPUS | High-stakes coding (production systems where bugs are costly, or complex agentic coding) needs Opus's stronger reasoning. |
| Stakes | Required family (this generation's names) | Why |
|---|---|---|
| Internal draft | SONNET | Sonnet is the minimum for multi-step, tool-using workflows — Haiku is not reliable enough to plan and execute multi-step tool use. |
| Customer facing | SONNET | Sonnet remains the minimum for agentic work — customer-facing does not by itself raise the bar further. |
| High stakes | FABLE | High-stakes agentic workflows are one of the two cases where the largest model is justified — long-horizon autonomy with costly failure modes needs the strongest available reasoning. |
| Stakes | Required family (this generation's names) | Why |
|---|---|---|
| Internal draft | SONNET | Deep analysis needs Sonnet-level reasoning even for internal work — Haiku is not reliable enough for multi-step analysis. |
| Customer facing | SONNET | Sonnet remains the baseline for deep analysis — customer-facing does not by itself raise the bar further. |
| High stakes | FABLE | High-stakes deep analysis is the other case where the largest model is justified — costly wrong conclusions (legal, financial, medical) need the strongest available reasoning. |
2. Per-model capability tiers
Every model's tier is comparable across providers on the same 0-3 scale (0 = Small / Fast, 1 = Balanced / Mid, 2 = Flagship, 3 = Frontier-Extended) — and every single assignment below is cited. Nothing here is a guess.
3. How ties are handled
Within a required tier, the Picker always recommends the CURRENT cheapest model — never a pricier same-tier alternative for an unstated reason. When two or more same-tier models are within 5% of each other's estimated cost for your workload, neither is declared a winner: both render as a co-recommendation ("Either fits — {Model A} is ~N% cheaper"), with the price difference stated plainly so you can judge for yourself.
Source for the model config files themselves:
config/tool_matrices/model_picker.php and
config/tool_matrices/model_capability_tiers.php.