Module pricing 

Centralised pricing logic for Anthropic models.

All cost calculations live here so that the engine, TUI, and any future consumer share a single source of truth. Update this file — and only this file — whenever Anthropic changes its pricing.

Prices are in USD per million tokens. Source: https://platform.claude.com/docs/en/about-claude/pricing (“Model pricing”, “Batch processing”, “Long context pricing” and “Feature-specific pricing → Prompt caching”).

Model	Input	Output	Cache read
Fable 5.1 / Mythos 5.1	$10.00	$50.00	0.025×
Fable 5 / Mythos 5	$10.00	$50.00	0.10×
Opus 5.5	$4.00	$20.00	0.05×
Opus 4.5 – 5	$5.00	$25.00	0.10×
Opus 4 / 4.1 (and Opus 3)	$15.00	$75.00	0.10×
Sonnet 5.5	$2.00	$10.00	0.05×
Sonnet 5	$2.00	$10.00	0.10×
Sonnet (other)	$3.00	$15.00	0.10×
Haiku 5.5, prompt ≤ 100k	$0.10	$0.50	0.10×
Haiku 5.5, prompt > 100k	$0.50	$2.50	0.10×
Haiku 3.5	$0.80	$4.00	0.10×
Haiku 3 (historical)	$0.25	$1.25	fallback*
Haiku (other)	$1.00	$5.00	0.10×

Historical Opus 3 / Haiku 3 input and output rates: https://www.anthropic.com/news/claude-3-family (verified 2026-10-10). *Legacy cache rates use the standard multiplier fallback; historical cache support and prices were not verified from that launch source.

Sonnet 5. The official model overview explicitly lists $2 input / $10 output per million tokens (verified 2026-10-10): https://platform.claude.com/docs/en/models/sonnet-5/overview. These are standard global direct-API estimates, not complete invoices: fast-mode/residency multipliers and server-tool charges are not modeled. Unknown-model fallbacks and routed/cloud/subscription model IDs are estimates, not verified provider invoice rates or net charges after credits/discounts.

Haiku 5.5 is priced by prompt length (“Long context pricing”): a prompt of over 100,000 tokens pays the higher row. The prompt is everything the request sent — uncached input + cache reads + cache writes — so calculate_cost_split picks the tier per request from those counts.

Cache pricing (relative to input price):

• Cache reads: 0.10× input price, except where the docs list a per-model multiplier (see cache_read_multiplier)
• Cache write (5m): 1.25× input price (5-minute TTL, the default)
• Cache write (1h): 2.00× input price (1-hour TTL, opt-in via cache_ttl)

Functions

cache_read_multiplier

Cache-read price as a multiple of the input price (https://platform.claude.com/docs/en/about-claude/pricing, “Prompt caching”). The docs list 0.025× for Fable 5.1 / Mythos 5.1 and 0.05× for Opus 5.5 / Sonnet 5.5; every other model uses the standard 0.1× (also the conservative default for models this table does not know).

calculate_cost

Calculate the USD cost of a single model turn (no cache-write TTL split).

calculate_cost_optional_split

Split-aware cost for callers holding an aggregate plus an optional TTL split (the shape of SessionEvent::Usage). When either split bucket is present, preserve its known count and assign any aggregate remainder to the conservative 1h rate. This is an estimate, not an inferred TTL. A split sum exceeding the aggregate is malformed metadata: retain the larger split sum rather than dropping known writes or lowering the request’s price tier. Arithmetic saturates. When both are None, TTL is explicitly unknown and the existing aggregate-at-5m estimate is retained (potentially underbilling).

calculate_cost_split

Calculate the USD cost of a single model turn with the cache-write TTL split made first-class.