LLM Models Reference

This doc summarizes current and legacy models from official provider docs. Use it when choosing models or updating context limits in shared/llm/model-context-limits.ts (re-exported from tools/prr/llm/model-context-limits.ts).

Sources (check for latest):

• Claude: Models overview
• OpenAI: Models — full ID list for scraping lives on All models

Machine-readable catalog (for PRR / automation)

Vendor doc pages change often; review bots may lag and suggest wrong renames (e.g. “use gpt-4-mini” when gpt-5-mini exists).

Refresh: npm run update-model-catalog. Weekly: GitHub Action refresh-model-catalog.yml. Override file path: PRR_MODEL_CATALOG_PATH.

PRR behavior: Outdated bot comments that call a catalog-valid id a “typo” and suggest another id are dismissed (assessSolvability, check 0a6) and optionally auto-healed in the workdir before issue analysis.

Environment: PRR_MODEL_CATALOG_PATH (override JSON path). Disable pieces: PRR_DISABLE_MODEL_CATALOG_SOLVABILITY=1 (no 0a6 dismissal), PRR_DISABLE_MODEL_CATALOG_AUTOHEAL=1 (no file rewrite; dismissal still applies if solvability is on). Full narrative: DEVELOPMENT.md — Commit gate and catalog model auto-heal.

Limitations (intentional): Detection requires framing (“typo”, “invalid model”, etc.) so neutral suggestions (“prefer X for cost”) are not auto-dismissed. Parsing needs a confident pair of ids (change A to B, use B instead of A, replace … with …, quoted A → B, or a heading like ### Model name typo \gpt-5-mini`with a lateruse `gpt-4o-mini`/recommended `…`in the same body). Summary tables that only say “FIXED →gpt-4o-mini” without an explicit wrong id may **not** match — extend parseModelRenameAdvice` if another stable pattern appears.

---

Claude (Anthropic)

All current Claude models support text + image input, text output, multilingual, and vision. Available via Claude API, AWS Bedrock, and Google Vertex AI.

Latest models

• Extended thinking: All three. Adaptive thinking: Opus 4.6, Sonnet 4.6 (not Haiku 4.5).
• 1M context (beta): Opus 4.6, Sonnet 4.6 — use context-1m-2025-08-07 header; long-context pricing above 200K.
• Reliable knowledge cutoff: Opus 4.6 May 2025, Sonnet 4.6 Aug 2025, Haiku 4.5 Feb 2025.

Legacy Claude models (still available)

---

OpenAI

Featured / frontier models

Codex-optimized (agentic coding)

Other widely used

Specialized

• Image: gpt-image-1.5, chatgpt-image-latest, gpt-image-1, gpt-image-1-mini
• Video: sora-2, sora-2-pro
• Deep research: o3-deep-research, o4-mini-deep-research
• Realtime / audio: gpt-realtime, gpt-realtime-1.5, gpt-audio, gpt-audio-1.5, gpt-realtime-mini, gpt-audio-mini
• Speech: gpt-4o-transcribe, gpt-4o-mini-transcribe, gpt-4o-mini-tts, etc.
• Embeddings: text-embedding-3-large, text-embedding-3-small

Open-weight (Apache 2.0)

• gpt-oss-120b — most powerful, fits H100
• gpt-oss-20b — lower latency

For full list, deprecations, and pricing see OpenAI Models.

---

Using this in PRR

• ElizaCloud / context limits: Edit ELIZACLOUD_MODEL_CONTEXT in shared/llm/model-context-limits.ts. Each entry sets maxContextTokens (total context window for that API model ID). PRR derives fix-prompt char caps from that (small contexts use a denser tokenization estimate). Optional maxFixPromptCharsCap tightens the derived value when the gateway still times out. Unknown gateway models use a conservative default until you add a row. Use ELIZACLOUD_MODEL_ID_ALIASES and pattern aliases in that file when the same physical model appears under multiple strings (e.g. Qwen/Qwen3-14B → alibaba/qwen-3-14b).
• Model IDs: Prefer stable snapshot IDs (e.g. claude-sonnet-4-5-20250929) when you need reproducible behavior; use aliases (e.g. claude-sonnet-4-6) for "latest" behavior.
• Claude 4.6: See Migrating to Claude 4.6 when moving from older Claude versions.

Rotation order and skip list

• llm-api / ElizaCloud: Fallback rotation order is DEFAULT_MODEL_ROTATIONS in shared/runners/types.ts; at runtime the list usually comes from the runner’s supportedModels (gateway/API discovery) and is filtered in tools/prr/models/rotation.ts using getEffectiveElizacloudSkipModelIds() from shared/constants.ts. Do not assume the static table in types.ts is the exact live order.
• Skip list (authoritative): ELIZACLOUD_SKIP_MODEL_IDS in shared/constants.ts. The table below is a snapshot for operators; if it disagrees with the source array, trust the source file and update this table when you change skips.

Last reviewed (skip table): 2026-03-28 — pill-output / audit follow-up (Qwen 14B default churn, empty-response logging).

Overrides and maintenance

• PRR_ELIZACLOUD_INCLUDE_MODELS: comma-separated — removes matching ids from the effective skip set (retry a timeout-skipped model after infra improves). Hyphenless suffix match is supported (see getEffectiveElizacloudSkipModelIds).

• PRR_ELIZACLOUD_EXTRA_SKIP_MODELS: comma-separated — adds ids to the built-in skip list for this environment only.

• getElizaCloudSkipReason(id): ids not in ELIZACLOUD_SKIP_REASON use default timeout so new skip entries still rotate with a sensible debug line until you assign zero-fix-rate.

• Operational habit: When RESULTS SUMMARY / Model Performance shows 0% fix rate for an ElizaCloud id, add it (with reason + comment) to shared/constants.ts and bump the “last reviewed” line above — same guidance as AGENTS.md.

• Per-run performance: Success/failure is recorded in state; rotation can prefer better-performing models within the same run. PRR_SESSION_MODEL_SKIP_FAILURES skips a tool/model for the rest of the process after repeated verification failures with zero verified fixes.

• PRR_SESSION_MODEL_SKIP_RESET_AFTER_FIX_ITERATIONS: positive integer — every N completed fix iterations (inner loop inside a push iteration), clear session skips so rotation can retry those models without restarting the process. 0 / unset = off. WHY: Long runs otherwise never revisit a model skipped early for transient failures (pill-output #847).

Provider model tables: last curated from linked docs; verify there for current IDs and pricing.