Recommended Models by Phase

Snapshot date: 2026-07-14. Model lineups, prices, and benchmark standings decay fast. This page is a cache, and caches need invalidation (ADLC Principle 10). Re-verify against provider pricing pages before committing budget, run model-ratchet after any model swap, and let your own gate-manifest first-pass ledger override every static prior here once you have ≥3 samples per tier. The canonical, most-current version of this table lives in the repo at docs/models-by-phase.md.

The ADLC series argues that using the most expensive model everywhere is the wrong default: model tier is a function of the cost of detecting an error, not of task prestige. The toolkit makes that mechanical. model-router reads rail density and DAG float and emits an abstract tier (cheap / mid / frontier) per ticket. What neither the theory nor the router resolves is the binding from abstract tier to concrete model. This page is that binding.

The three tiers

Two doctrine points shape everything below:

1. "Frontier" means the best model you are allowed to run, not the most expensive model that exists. The Frontier-Free Doctrine sets an Opus-class ceiling as the design center: the lifecycle must hit its accuracy targets with stable, generally available models without depending on any provider's preview or special-access tier. Above-ceiling models are optional headroom, never a requirement. You never need a model smarter than the gate it must pass.
2. Measure the stack, not the model. A 3-pass mid-tier prosecution stack with 0.85 planted-bug recall is the more capable reviewer than a 1-pass frontier model at 0.6. review-calibration makes that exchange rate a number. Tier labels are cold-start estimates only.

Phase → tier map

The barbell falls out of this table: heavy spend at P1–P2 and P5, light in the middle where rails make building cheap.

Tier → model, by provider

Prices are USD per million tokens, input/output, standard API rates. Benchmark numbers are directional only. SWE-bench Verified and SWE-bench Pro are different benchmarks with non-comparable scores.

Anthropic

OpenAI

The ADLC Codex plugin does not select or pin a Codex model. Verify availability, context, and pricing in the official OpenAI model catalog before binding a router tier.

GPT-5.4 is this repository's operator default, not a claim that it is the newest model in OpenAI's catalog. At this snapshot the official catalog also documents the GPT-5.6 family; teams may calibrate an available newer model, but the plugin never silently retargets the repository policy.

Google

xAI

Mistral

Open-weight, hosted APIs

These are the cheapest capable tiers available, and they are distinct model families, exactly what the P5 multi-provider quorum needs.

Local / self-hosted

Local models are a legitimate cheap tier for rail-dense work where the gates do the verification, and on strong hardware a legitimate mid tier. The practical single-machine ceiling at this snapshot is ~71–72% SWE-bench Verified. Do not assign local models frontier duties (P1/P2 specs and contracts): those phases exist precisely because their errors escape gates.

Notes:

• The coldstart probe is a perfect local job: free, private, and the weaker the probe the more honest the gate.
• Skip Llama 4 for coding work (~24% Verified at snapshot).
• A local cheap tier changes the ladder math: when the ladder start costs ~$0, the router's --floor can be more permissive for float-rich tickets.

Suggested P5 cross-family quorums

Anchor on your primary family, then add one different family:

A single provider's clean approve on high-blast-radius paths is advisory, not a gate-pass.

Keeping this honest

• Your ledger beats this page. model-router prefers empirical per-repo priors from .adlc/manifest.jsonl over any static table. This page is the cold-start estimate.
• Re-verify on model churn. New model ships → model-ratchet schedules re-prosecution; review-calibration re-measures reviewer recall before you trust a swap.
• Benchmark hygiene. Vendor-reported numbers, promo pricing, and preview models are flagged inline; treat every number as directional and every price as perishable.