[pitoi]

Price Claude models from Anthropic's list prices, bill 1-hour cache writes separately, flag unknown models as unpriced.

Why

An audit of OpenLaw Benchmarks runs (@openlaw-1) found claude-opus-5 and claude-sonnet-5 billed at the Sonnet 4.x card (3 / 0.30 / 3.75 / 15 per MTok). Cause: aieo has no built-in per-model price for Anthropic, only the bifrost datasheet. When the datasheet is not loaded (STRUT_MODEL_CATALOG=off, a failed fetch, or a host that never calls startModelCatalog) or lacks a model, every Claude model falls back to TOKEN_PRICING.anthropic.

What changed (src/aieo/provider.ts, src/pricing.ts, src/steps/core/agent.ts, src/steps/core/llm.ts)

• Built-in ANTHROPIC_MODELS table with verified list prices (below). Lookup order for anthropic: datasheet (what the LLM gateway bills from), then this table.
• Id normalisation (normalizeAnthropicModelId): anthropic/claude-opus-5, claude-opus-5.5 (dotted), claude-haiku-4-5-20251001 / @20251001 all resolve to the bare entry. getProviderForModel maps a bare claude-* id to anthropic (below the LLM_PROVIDER check, like gpt- / gemini-).
• Unknown models are unpriced: for anthropic, openai, google, xai, openrouter, a named model that neither the table nor the datasheet prices costs 0, logs one [aieo] no price for <provider>/<model>: cost reported as 0 (unpriced) warning, and the agent / llm step output carries costUnpriced: true. New modelPriceSource(provider, model) → table | catalog | provider-default | unpriced. Unchanged: the provider flat rate when no model is named, and the maple / numes / tokenfactory "dearest listed rate" ceiling for unlisted gateway models (a deliberate never-under-report policy, now commented). The explicit tokenfactory $0 entries (glm-5.3, kimi-k3.1-pro, …) are untouched.
• 1-hour cache writes: the agent step already supports cacheTtl: "1h", but usage had one write count priced at the 5-minute rate. TokenUsage.cacheWrite1hTokens (a subset of cacheWriteTokens) is read from Anthropic's raw cache_creation.ephemeral_1h_input_tokens (usage.raw or providerMetadata.anthropic.usage). If there is no 5m/1h split and the step ran with cacheTtl 1h on anthropic, all its writes count as 1h. Priced at cacheWrite1hPrice (table, or the datasheet's cache_creation_input_token_cost_above_1hr).
• Haiku 5.5 long-context tier: a call whose prompt (input + cache read + cache write) is over 100K tokens is priced entirely at the higher tier.

Prices (USD / MTok: input / cache read / 5m write / 1h write / output)

Sources: https://platform.claude.com/docs/en/about-claude/pricing and https://claude.com/pricing, both read 2026-11-11. The bifrost datasheet (https://getbifrost.ai/datasheet, same day) holds the same figures.

Model	Before (fallback when no datasheet)	After
claude-opus-5	3 / 0.30 / 3.75 / 3.75 / 15	5 / 0.50 / 6.25 / 10 / 25
claude-sonnet-5	3 / 0.30 / 3.75 / 3.75 / 15	2 / 0.20 / 2.50 / 4 / 10
claude-opus-5-5	3 / 0.30 / 3.75 / 3.75 / 15	4 / 0.20 / 5 / 8 / 20
claude-sonnet-5-5	3 / 0.30 / 3.75 / 3.75 / 15	2 / 0.10 / 2.50 / 4 / 10
claude-haiku-5-5 (≤100K / >100K prompt)	3 / 0.30 / 3.75 / 3.75 / 15	0.10 / 0.01 / 0.125 / 0.20 / 0.50 · 0.50 / 0.05 / 0.625 / 1 / 2.50
claude-haiku-4-5	3 / 0.30 / 3.75 / 3.75 / 15	1 / 0.10 / 1.25 / 2 / 5
claude-opus-4-8, 4-7, 4-6, 4-5	3 / 0.30 / 3.75 / 3.75 / 15	5 / 0.50 / 6.25 / 10 / 25
claude-opus-4-1, claude-opus-4	3 / 0.30 / 3.75 / 3.75 / 15	15 / 1.50 / 18.75 / 30 / 75
claude-sonnet-4-6, 4-5, 4	3 / 0.30 / 3.75 / 3.75 / 15	3 / 0.30 / 3.75 / 6 / 15 (only the 1h write changes)
claude-fable-5-2, claude-mythos-5-2	3 / 0.30 / 3.75 / 3.75 / 15	8 / 0.20 / 10 / 16 / 40
claude-fable-5-1, claude-mythos-5-1	3 / 0.30 / 3.75 / 3.75 / 15	10 / 0.25 / 12.50 / 20 / 50
claude-fable-5, claude-mythos-5	3 / 0.30 / 3.75 / 3.75 / 15	10 / 1 / 12.50 / 20 / 50
unknown claude-*	3 / 0.30 / 3.75 / 3.75 / 15	unpriced: 0 + warning + costUnpriced

Audit figures checked: Sonnet 5 (2 / 0.20 / 2.50 / 10) and Opus 5 (5 / 0.50 / 6.25 / 25) match the official pages exactly, and the 0.05x cache read on Opus 5.5 / Sonnet 5.5 is confirmed. Opus 5.5 is 4 / 20, not Opus 5's 5 / 25.

Not modelled: Batch (-50%), fast mode, US-only inference_geo (1.1x). Other providers' per-model prices come from the datasheet as before. Their flat rates (openai, google, xai, openrouter) now apply only when no model is named, and their values are unchanged.

Effect on run 1794343520308 (Opus 5 reasoning + Sonnet 5 standard)

Every Opus 5 price is exactly 5/3 of the fallback card and every Sonnet 5 price exactly 2/3, so the billed figures rescale exactly (assuming 5-minute writes, the agent default):

• Opus 5 steps: $18.87 → $31.45
• Sonnet 5 steps: $11.47 → $7.65
• Metered agents: $30.34 → $39.10 (+$8.76, +28.9%)

Tests

New src/aieo/pricing-table.test.ts: table prices with no datasheet, prefix / dotted / dated ids, claude-* provider inference, unpriced behaviour (0, one warning, source), tokenfactory $0 kept, gateway ceiling kept, datasheet precedence, 1h write pricing and datasheet parsing, the Haiku 5.5 tier. Also pricing.test.ts (ephemeral_1h_input_tokens), llm.test.ts and agent.test.ts (costUnpriced). catalog.test.ts, runner.test.ts and job-threads.test.ts were updated for the new behaviour and the new usage field. tsc --noEmit is clean. The aieo, pricing, llm, runner, store, compaction, evaluate, job-threads and graph tests pass. agent.test.ts has 3 bash-process (ESRCH) failures that also fail on the base branch in the pod.