coding-agent-cost 0.2.2__tar.gz → 0.4.0__tar.gz

{coding_agent_cost-0.2.2/coding_agent_cost.egg-info → coding_agent_cost-0.4.0}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: coding-agent-cost
-Version: 0.2.2
+Version: 0.4.0
 Summary: Estimate AI coding agent (Claude Code / Codex CLI) token usage and cost from local logs
 Author: shiki-yusuke
 License: MIT
@@ -232,10 +232,28 @@ network, never calls `gh`, and never resolves branches or PRs.
   flagged, since the reader's own cross-check of real transcripts found
   exactly that pattern in every observed duplicated group. See
   `CHANGELOG.md`'s 0.2.0 entry.
+  If the row a group adopts has no `stop_reason` (common in subagent
+  transcripts, where a message ending in `tool_use` may never get its final
+  line), its `output_tokens` is only the streaming head's value: the
+  group's `output` fact is flagged `source_quality: "output_lower_bound"`
+  -- the observed lower bound of that message's output tokens, not
+  guaranteed to be >= the true count. The group's input-side facts stay
+  `"ok"`, and the token amount itself is unchanged (still priced and
+  included in rows/totals); the flag is a warning only.
   When Anthropic's prompt-cache TTL breakdown (5-minute vs 1-hour writes) is
   present in the log, it's used; otherwise the cache-write tokens are priced
   at the 5-minute rate as an explicit **lower bound** and flagged
   `lower_bound` rather than guessed at the (more expensive) 1-hour rate.
+  Each Claude fact also carries its request's prompt length
+  (`prompt_tokens` = `input_tokens` + `cache_read_input_tokens` +
+  `cache_creation_input_tokens`), used only to pick the tier of a model
+  whose price depends on prompt length (`claude-haiku-5-5`).
+- **Prompt-length tiers need a known prompt length.** A fact whose prompt
+  length is unknown -- every Codex fact (a delta of cumulative totals, not
+  one request), a Claude fact from a conflicting dedup group, or a Claude
+  row whose cache-write TTL breakdown exceeds its `cache_creation_input_tokens`
+  -- is priced at the model's base (cheapest) tier and flagged
+  `lower_bound`. Models without tiers are unaffected.
 - **Codex CLI**: rollout files record a *cumulative* token count after each
   turn; agent-cost turns that into per-turn deltas. Under the Codex token-rate
   tariff used here, cache writes are not added as a separate charge. The current
@@ -295,6 +313,37 @@ carries a `catalog_version`, a list of `sources` (the pricing page a rate
 came from), and is validated on load (no duplicate model keys or aliases,
 no negative rates, no overlapping periods for the same model).
 
+A rate period can price by prompt length (catalog `schema_version` `"2"`).
+The period's own five values are the base tier, for prompts up to the first
+threshold; each entry in `prompt_tiers` applies to prompts *strictly over*
+its `prompt_tokens_over`, and a request uses the largest threshold it
+exceeds. Prompt length is the request's input + cache-read + cache-write
+tokens, and the chosen tier prices every token kind of that request,
+output included:
+
+```json
+{
+  "rate_id": "claude-haiku-5-5-launch-2026-10-07",
+  "effective_from": "2026-10-07T00:00:00+00:00",
+  "effective_until": null,
+  "input_nocache": "0.10", "cache_read": "0.01", "cache_write_5m": "0.125", "cache_write_1h": "0.20", "output": "0.50",
+  "prompt_tiers": [
+    { "prompt_tokens_over": 100000,
+      "input_nocache": "0.50", "cache_read": "0.05", "cache_write_5m": "0.625", "cache_write_1h": "1.0", "output": "2.50" }
+  ]
+}
+```
+
+In a period with `prompt_tiers`, all five values of the base and of every
+tier must be present and non-null, thresholds must be positive integers in
+ascending order without duplicates, and every value must be greater than or
+equal to the same value in the previous tier (the base first). That
+monotonicity is what makes the base tier a true lower bound for a fact whose
+prompt length is unknown. A `schema_version` `"1"` catalog may not contain
+`prompt_tiers`; agent-cost 0.3.0 and earlier reject a `"2"` catalog as
+unsupported rather than mispricing it, so a custom `--rates` file that uses
+tiers must declare `"2"`.
+
 To use your own catalog instead of the one bundled with the package, pass
 `--rates path/to/rates.json` to `report` or `export` -- this fully replaces
 the bundled catalog, it does not merge with it. Inspect any catalog with:
@@ -342,7 +391,7 @@ agent-cost measure --session-id <id> [--session-id <id> ...] \
 ```json
 {
   "protocol_version": "measure/v1",
-  "producer_version": "0.2.2",
+  "producer_version": "0.3.0",
   "accounting_basis": "agent-cost-raw-total/v2",
   "generated_at": "...",
   "window": { "since": "...", "until": null },
@@ -363,7 +412,7 @@ agent-cost measure --session-id <id> [--session-id <id> ...] \
     "duplicate_rows_skipped": 0,
     "conflicting_duplicate_groups": 0,
     "missing_dedup_identity_rows": 0,
-    "source_quality": { "ok": 41, "first_event_delta": 2, "identity_missing": 0 }
+    "source_quality": { "ok": 41, "first_event_delta": 2, "identity_missing": 0, "output_lower_bound": 1 }
   }
 }
 ```
@@ -394,9 +443,11 @@ agent-cost makes zero network calls. `agent-cost export`'s JSONL never
 includes absolute file paths, rollout paths, prompt/message content, or git
 branch names -- only the fields needed to reproduce a cost estimate:
 `occurred_at_utc`, `agent`, `session_id`, `model_raw`, `model_key`,
-`token_kind`, `tokens`, `mode`, and `source_quality` (a fixed-vocabulary
-caveat about how that one fact was derived, e.g. `"ok"` or Codex's
-`"first_event_delta"` -- never null).
+`token_kind`, `tokens`, `mode`, `source_quality` (a fixed-vocabulary
+caveat about how that one fact was derived: `"ok"`, Codex's
+`"first_event_delta"`, or Claude's `"identity_missing"` /
+`"output_lower_bound"` -- never null), and `prompt_tokens` (the request's
+prompt length as an integer, or null when unknown).
 
 ## License
 
@@ -420,6 +471,10 @@ MIT. See [LICENSE](LICENSE).
   未知のモデル・単価表にない token 種別は `unpriced` として扱い、憶測の価格を出しません。
 - 単価表 (`agent_cost/rates.json`) は履歴型カタログで、値上げは新しい期間として追加します。
   `--rates PATH` で別カタログに完全差し替えできます。
+- prompt 長で単価が変わるモデル（`claude-haiku-5-5`、100,000 tokens 超で高い単価）は期間の
+  `prompt_tiers`（schema_version `"2"`）で表し、リクエストごとの prompt 長（input + cache read +
+  cache write）で tier を選びます。prompt 長が不明な fact（Codex 由来、conflicting group、usage
+  不整合）は基底 tier（最安）で **下限推計**（`lower_bound`）になります。
 - `agent-cost measure --session-id ID [--session-id ID ...] --format json` は他プログラムから
   subprocess で叩くための機械可読な契約です（`protocol_version: "measure/v1"`）。指定した
   session_id が1件も見つからなくても終了コードは0（空集計として表現）、`--session-id` 未指定など

{coding_agent_cost-0.2.2 → coding_agent_cost-0.4.0}/README.md

@@ -206,10 +206,28 @@ network, never calls `gh`, and never resolves branches or PRs.
   flagged, since the reader's own cross-check of real transcripts found
   exactly that pattern in every observed duplicated group. See
   `CHANGELOG.md`'s 0.2.0 entry.
+  If the row a group adopts has no `stop_reason` (common in subagent
+  transcripts, where a message ending in `tool_use` may never get its final
+  line), its `output_tokens` is only the streaming head's value: the
+  group's `output` fact is flagged `source_quality: "output_lower_bound"`
+  -- the observed lower bound of that message's output tokens, not
+  guaranteed to be >= the true count. The group's input-side facts stay
+  `"ok"`, and the token amount itself is unchanged (still priced and
+  included in rows/totals); the flag is a warning only.
   When Anthropic's prompt-cache TTL breakdown (5-minute vs 1-hour writes) is
   present in the log, it's used; otherwise the cache-write tokens are priced
   at the 5-minute rate as an explicit **lower bound** and flagged
   `lower_bound` rather than guessed at the (more expensive) 1-hour rate.
+  Each Claude fact also carries its request's prompt length
+  (`prompt_tokens` = `input_tokens` + `cache_read_input_tokens` +
+  `cache_creation_input_tokens`), used only to pick the tier of a model
+  whose price depends on prompt length (`claude-haiku-5-5`).
+- **Prompt-length tiers need a known prompt length.** A fact whose prompt
+  length is unknown -- every Codex fact (a delta of cumulative totals, not
+  one request), a Claude fact from a conflicting dedup group, or a Claude
+  row whose cache-write TTL breakdown exceeds its `cache_creation_input_tokens`
+  -- is priced at the model's base (cheapest) tier and flagged
+  `lower_bound`. Models without tiers are unaffected.
 - **Codex CLI**: rollout files record a *cumulative* token count after each
   turn; agent-cost turns that into per-turn deltas. Under the Codex token-rate
   tariff used here, cache writes are not added as a separate charge. The current
@@ -269,6 +287,37 @@ carries a `catalog_version`, a list of `sources` (the pricing page a rate
 came from), and is validated on load (no duplicate model keys or aliases,
 no negative rates, no overlapping periods for the same model).
 
+A rate period can price by prompt length (catalog `schema_version` `"2"`).
+The period's own five values are the base tier, for prompts up to the first
+threshold; each entry in `prompt_tiers` applies to prompts *strictly over*
+its `prompt_tokens_over`, and a request uses the largest threshold it
+exceeds. Prompt length is the request's input + cache-read + cache-write
+tokens, and the chosen tier prices every token kind of that request,
+output included:
+
+```json
+{
+  "rate_id": "claude-haiku-5-5-launch-2026-10-07",
+  "effective_from": "2026-10-07T00:00:00+00:00",
+  "effective_until": null,
+  "input_nocache": "0.10", "cache_read": "0.01", "cache_write_5m": "0.125", "cache_write_1h": "0.20", "output": "0.50",
+  "prompt_tiers": [
+    { "prompt_tokens_over": 100000,
+      "input_nocache": "0.50", "cache_read": "0.05", "cache_write_5m": "0.625", "cache_write_1h": "1.0", "output": "2.50" }
+  ]
+}
+```
+
+In a period with `prompt_tiers`, all five values of the base and of every
+tier must be present and non-null, thresholds must be positive integers in
+ascending order without duplicates, and every value must be greater than or
+equal to the same value in the previous tier (the base first). That
+monotonicity is what makes the base tier a true lower bound for a fact whose
+prompt length is unknown. A `schema_version` `"1"` catalog may not contain
+`prompt_tiers`; agent-cost 0.3.0 and earlier reject a `"2"` catalog as
+unsupported rather than mispricing it, so a custom `--rates` file that uses
+tiers must declare `"2"`.
+
 To use your own catalog instead of the one bundled with the package, pass
 `--rates path/to/rates.json` to `report` or `export` -- this fully replaces
 the bundled catalog, it does not merge with it. Inspect any catalog with:
@@ -316,7 +365,7 @@ agent-cost measure --session-id <id> [--session-id <id> ...] \
 ```json
 {
   "protocol_version": "measure/v1",
-  "producer_version": "0.2.2",
+  "producer_version": "0.3.0",
   "accounting_basis": "agent-cost-raw-total/v2",
   "generated_at": "...",
   "window": { "since": "...", "until": null },
@@ -337,7 +386,7 @@ agent-cost measure --session-id <id> [--session-id <id> ...] \
     "duplicate_rows_skipped": 0,
     "conflicting_duplicate_groups": 0,
     "missing_dedup_identity_rows": 0,
-    "source_quality": { "ok": 41, "first_event_delta": 2, "identity_missing": 0 }
+    "source_quality": { "ok": 41, "first_event_delta": 2, "identity_missing": 0, "output_lower_bound": 1 }
   }
 }
 ```
@@ -368,9 +417,11 @@ agent-cost makes zero network calls. `agent-cost export`'s JSONL never
 includes absolute file paths, rollout paths, prompt/message content, or git
 branch names -- only the fields needed to reproduce a cost estimate:
 `occurred_at_utc`, `agent`, `session_id`, `model_raw`, `model_key`,
-`token_kind`, `tokens`, `mode`, and `source_quality` (a fixed-vocabulary
-caveat about how that one fact was derived, e.g. `"ok"` or Codex's
-`"first_event_delta"` -- never null).
+`token_kind`, `tokens`, `mode`, `source_quality` (a fixed-vocabulary
+caveat about how that one fact was derived: `"ok"`, Codex's
+`"first_event_delta"`, or Claude's `"identity_missing"` /
+`"output_lower_bound"` -- never null), and `prompt_tokens` (the request's
+prompt length as an integer, or null when unknown).
 
 ## License
 
@@ -394,6 +445,10 @@ MIT. See [LICENSE](LICENSE).
   未知のモデル・単価表にない token 種別は `unpriced` として扱い、憶測の価格を出しません。
 - 単価表 (`agent_cost/rates.json`) は履歴型カタログで、値上げは新しい期間として追加します。
   `--rates PATH` で別カタログに完全差し替えできます。
+- prompt 長で単価が変わるモデル（`claude-haiku-5-5`、100,000 tokens 超で高い単価）は期間の
+  `prompt_tiers`（schema_version `"2"`）で表し、リクエストごとの prompt 長（input + cache read +
+  cache write）で tier を選びます。prompt 長が不明な fact（Codex 由来、conflicting group、usage
+  不整合）は基底 tier（最安）で **下限推計**（`lower_bound`）になります。
 - `agent-cost measure --session-id ID [--session-id ID ...] --format json` は他プログラムから
   subprocess で叩くための機械可読な契約です（`protocol_version: "measure/v1"`）。指定した
   session_id が1件も見つからなくても終了コードは0（空集計として表現）、`--session-id` 未指定など

{coding_agent_cost-0.2.2 → coding_agent_cost-0.4.0}/agent_cost/__init__.py

@@ -6,4 +6,4 @@ against a versioned rate catalog. Everything happens locally; there are no
 network calls.
 """
 
-__version__ = "0.2.2"
+__version__ = "0.4.0"

{coding_agent_cost-0.2.2 → coding_agent_cost-0.4.0}/agent_cost/aggregate.py

@@ -28,6 +28,11 @@ def price_fact(catalog: RateCatalog, fact: Fact) -> Tuple[Optional[Decimal], str
     never at the more expensive 1-hour rate -- and flagged accordingly.
     An unrecognized model or an unpriced token kind for an otherwise-known
     model both return ``(None, "unpriced", None)`` rather than a guess.
+
+    For a period with ``prompt_tiers``, the fact is priced at the tier with
+    the largest ``prompt_tokens_over`` its ``prompt_tokens`` exceeds (the
+    base values when none). A fact whose ``prompt_tokens`` is unknown is
+    priced at the base values -- the cheapest tier -- as a lower bound.
     """
     resolved_key, period = catalog.rate_for(fact.model_key, fact.occurred_at_utc)
     if resolved_key is None or period is None:
@@ -47,7 +52,16 @@ def price_fact(catalog: RateCatalog, fact: Fact) -> Tuple[Optional[Decimal], str
         rate_field = "cache_write_5m"
         lower_bound = True
 
-    rate = period.values.get(rate_field)
+    values = period.values
+    if period.prompt_tiers:
+        if fact.prompt_tokens is None:
+            lower_bound = True
+        else:
+            for tier in period.prompt_tiers:
+                if fact.prompt_tokens > tier.prompt_tokens_over:
+                    values = tier.values
+
+    rate = values.get(rate_field)
     if rate is None:
         return None, "unpriced", None
 

{coding_agent_cost-0.2.2 → coding_agent_cost-0.4.0}/agent_cost/cli.py

@@ -168,6 +168,7 @@ def cmd_export(args) -> int:
                 "tokens": f.tokens,
                 "mode": f.mode,
                 "source_quality": f.source_quality,
+                "prompt_tokens": f.prompt_tokens,
             }
             out.write(json.dumps(record, ensure_ascii=False) + "\n")
     finally:
@@ -305,6 +306,11 @@ def cmd_rates_show(args) -> int:
             print(f"  [{period.rate_id}] {period.effective_from.isoformat()} .. {until}")
             for field_name, value in period.values.items():
                 print(f"      {field_name}: {value if value is not None else '(unpriced)'}")
+            if period.prompt_tiers:
+                print("      prompt_tiers:")
+                for tier in period.prompt_tiers:
+                    values = " ".join(f"{name}={value}" for name, value in tier.values.items())
+                    print(f"        > {tier.prompt_tokens_over}: {values}")
         return 0
 
     print(f"catalog_version: {catalog.catalog_version}")

{coding_agent_cost-0.2.2 → coding_agent_cost-0.4.0}/agent_cost/facts.py

@@ -32,10 +32,14 @@ MODES = ("fast", "normal", "unknown")
 # (e.g. Codex's first delta in a rollout is measured against an assumed
 # zero baseline; Claude's reader can't dedup a row that lacks a full
 # (message.id, requestId) pair, so it emits that row individually and
-# flags it "identity_missing" rather than silently treating it as "ok").
+# flags it "identity_missing" rather than silently treating it as "ok";
+# "output_lower_bound" marks a Claude output fact whose tokens are only the
+# observed lower bound of that message's output_tokens -- the reader's
+# adopted row had no stop_reason, so the value is not guaranteed to be
+# >= the true count).
 # This is never left unset -- every Fact defaults to "ok" so export never
 # emits a null source_quality.
-SOURCE_QUALITY_VALUES = ("ok", "first_event_delta", "identity_missing")
+SOURCE_QUALITY_VALUES = ("ok", "first_event_delta", "identity_missing", "output_lower_bound")
 
 _BRACKET_SUFFIX = re.compile(r"\[[^\]]*\]$")
 _DATE_SUFFIX = re.compile(r"@\d{6,8}$")
@@ -80,6 +84,12 @@ class Fact:
     tokens: int
     mode: str = "unknown"
     source_quality: str = "ok"
+    # The prompt length of the one request that produced this fact (usage's
+    # input_tokens + cache_read_input_tokens + cache_creation_input_tokens).
+    # Used only to pick the tier of a prompt-length-tiered rate. None means
+    # unknown (a Codex fact, or a Claude fact whose usage was inconsistent or
+    # came from a conflicting dedup group).
+    prompt_tokens: Optional[int] = None
 
     def __post_init__(self) -> None:
         if self.agent not in AGENTS:
@@ -92,5 +102,11 @@ class Fact:
             raise ValueError(f"invalid source_quality: {self.source_quality!r}")
         if self.tokens < 0:
             raise ValueError(f"tokens must be >= 0, got {self.tokens}")
+        if self.prompt_tokens is not None and (
+            isinstance(self.prompt_tokens, bool)
+            or not isinstance(self.prompt_tokens, int)
+            or self.prompt_tokens < 0
+        ):
+            raise ValueError(f"prompt_tokens must be None or an int >= 0, got {self.prompt_tokens!r}")
         if self.occurred_at_utc.tzinfo is None:
             raise ValueError("occurred_at_utc must be timezone-aware")

{coding_agent_cost-0.2.2 → coding_agent_cost-0.4.0}/agent_cost/rates.json

@@ -1,6 +1,6 @@
 {
-  "schema_version": "1",
-  "catalog_version": "2026-09-29",
+  "schema_version": "2",
+  "catalog_version": "2026-10-09",
   "currency": "USD",
   "unit": "per_mtok",
   "usd_per_credit": "0.04",
@@ -12,7 +12,8 @@
     "claude-fable-5-1 added 2026-09-09: rate_id claude-fable-5-1-local-first-observed-2026-09-02. effective_from (2026-09-02T01:17:55Z) is the timestamp the raw model ID was first observed across all 54 local transcript files (none found in August), not a confirmed official launch date -- if an authoritative launch date is confirmed later, add it as a separate rate_id with effective_until on this one rather than editing this period in place. Values (cache_read $0.25/MTok = 0.025x base input) differ from claude-fable-5's cache_read ($1.00/MTok = 0.1x); this is a genuine per-model price difference confirmed against the primary source (see sources below), not an inconsistency to reconcile.",
     "claude-sonnet-5 corrected 2026-09-09: a prior period (rate_id claude-sonnet-5-standard, effective_from 2026-09-01, values 3.0/0.30/3.75/6.0/15.0) priced a previously-announced rate increase that the pricing page's claude-sonnet-5-introductory-pricing note states did not occur (\"The previously scheduled increase to $3/$15 ... on September 1, 2026 will not occur\"). That period has been replaced by claude-sonnet-5-standard-2026-09-01, which carries the same $2/$0.20/$2.50/$4/$10 values as the preceding claude-sonnet-5-launch-promo period, since the launch price is confirmed to be the ongoing price with no change at the 2026-09-01 boundary. Any measure/v1 or report output computed against the pre-correction catalog for claude-sonnet-5 sessions occurring on or after 2026-09-01 overstated cost by 1.5x for that model. To detect this class of error going forward, cross-check each priced model's rate periods against its primary pricing page's Note text on a recurring (e.g. monthly) basis for mentions of a rate-change announcement being withdrawn or not taking effect -- a period can be internally valid (schema-wise) and still price a rate that was announced but never actually charged.",
     "claude-opus-5-5 added 2026-09-23: rate_id claude-opus-5-5-launch-2026-09-22. effective_from 2026-09-22T00:00:00Z is 00:00:00Z of the official public launch date (anthropic.com/claude-opus-5-5, 'September 22, 2026'); the launch time of day is not published, but a transcript row whose raw model ID is claude-opus-5-5 cannot predate the launch, so pricing from the start of that UTC day cannot price a pre-launch event. Values 4.0/0.20/5.0/8.0/20.0 per the pricing page: cache_read is 0.05x base input (pricing-page footnote 2), a genuine per-model difference from the standard 0.1x and from claude-fable-5-1's 0.025x. fast_multiplier 2.0 (fast mode $8/$40 per MTok on the pricing page). Claude Code's CHANGELOG entry for 2.1.280 states Opus 5.5 became the default model (see sources); rows whose raw model ID is claude-opus-5-5 would be unpriced without this entry. Claude Sonnet 5.5 and Claude Haiku 5.5 are announced to follow 'in the coming weeks' -- add them as separate entries when their model IDs and prices are published, never as aliases of the 5.x entries.",
-    "claude-sonnet-5-5 added 2026-09-29: rate_id claude-sonnet-5-5-launch-2026-09-28. effective_from 2026-09-28T00:00:00Z is 00:00:00Z of the official public launch date (anthropic.com/claude-sonnet-5-5, 'September 28, 2026'); as with claude-opus-5-5, the launch time of day is not published, but a transcript row whose raw model ID is claude-sonnet-5-5 cannot predate the launch, so pricing from the start of that UTC day cannot price a pre-launch event. Values 2.0/0.20/2.50/4.0/10.0 per the pricing page (Sonnet 5.5 row, no footnote: cache_read is the standard 0.1x, unlike claude-opus-5-5's 0.05x and claude-fable-5-1's 0.025x). Identical to claude-sonnet-5's ongoing values, but kept as a separate entry, never an alias: the pricing page lists the two rows independently and a future change to one need not apply to the other. fast_multiplier 1.0: the pricing page's fast-mode table lists only the Opus models. Claude Code's CHANGELOG entry for 2.1.284 states Sonnet 5.5 is now the default Sonnet model, so the `sonnet` alias in agent frontmatter resolves to it and rows whose raw model ID is claude-sonnet-5-5 are expected from that version onward; no such row existed in the local transcripts on 2026-09-29. Claude Haiku 5.5 is announced to follow 'in the coming weeks' -- add it as a separate entry when its model ID and prices are published."
+    "claude-sonnet-5-5 added 2026-09-29: rate_id claude-sonnet-5-5-launch-2026-09-28. effective_from 2026-09-28T00:00:00Z is 00:00:00Z of the official public launch date (anthropic.com/claude-sonnet-5-5, 'September 28, 2026'); as with claude-opus-5-5, the launch time of day is not published, but a transcript row whose raw model ID is claude-sonnet-5-5 cannot predate the launch, so pricing from the start of that UTC day cannot price a pre-launch event. Values 2.0/0.20/2.50/4.0/10.0 per the pricing page (Sonnet 5.5 row, no footnote: cache_read is the standard 0.1x, unlike claude-opus-5-5's 0.05x and claude-fable-5-1's 0.025x). Identical to claude-sonnet-5's ongoing values, but kept as a separate entry, never an alias: the pricing page lists the two rows independently and a future change to one need not apply to the other. fast_multiplier 1.0: the pricing page's fast-mode table lists only the Opus models. Claude Code's CHANGELOG entry for 2.1.284 states Sonnet 5.5 is now the default Sonnet model, so the `sonnet` alias in agent frontmatter resolves to it and rows whose raw model ID is claude-sonnet-5-5 are expected from that version onward; no such row existed in the local transcripts on 2026-09-29. Claude Haiku 5.5 is announced to follow 'in the coming weeks' -- add it as a separate entry when its model ID and prices are published.",
+    "claude-haiku-5-5 added 2026-10-09: rate_id claude-haiku-5-5-launch-2026-10-07. effective_from 2026-10-07T00:00:00Z is 00:00:00Z of the official release date (platform.claude.com/docs/en/models/haiku-5-5/overview, 'Released October 7, 2026'); as with claude-sonnet-5-5, the release time of day is not published, but a transcript row whose raw model ID is claude-haiku-5-5 cannot predate the release. Its prices are tiered by prompt length: the period's top-level values (0.10/0.01/0.125/0.20/0.50) apply to prompts up to 100,000 tokens, and prompt_tiers[0] (prompt_tokens_over 100000: 0.50/0.05/0.625/1.0/2.50) to prompts over 100,000 tokens, per the pricing page's two Haiku 5.5 rows (no footnote: cache_read is the standard 0.1x). Prompt length follows the pricing page's Long context pricing section: it counts all of a request's input tokens including cache reads and cache writes, and each request is priced on its own, so a request over the threshold pays the higher prices for every token kind, output included. A fact whose prompt length is unknown (Codex, a conflicting Claude dedup group, inconsistent usage) is priced at the top-level values as a lower_bound. A separate entry, never an alias of claude-haiku-4-5. fast_multiplier 1.0: the pricing page's fast-mode table lists only the Opus models. Claude Code's CHANGELOG entry for 2.1.293 states Haiku 5.5 is now the default Haiku model, so the `haiku` alias resolves to it from that version onward. schema_version is 2 because prompt_tiers is new: agent-cost 0.3.0 and earlier reject schema_version 2 as unsupported instead of silently pricing every claude-haiku-5-5 request at the cheaper tier, and a schema_version 1 catalog may not carry prompt_tiers."
   ],
   "sources": [
     {
@@ -97,6 +98,21 @@
       "url": "https://github.com/anthropics/claude-code/blob/main/CHANGELOG.md",
       "retrieved_at": "2026-09-29",
       "note": "Claude Code 2.1.284 entry: added Claude Sonnet 5.5 (claude-sonnet-5-5), now the default Sonnet model; the reason claude-sonnet-5-5 rows are expected in transcripts from that version"
+    },
+    {
+      "url": "https://platform.claude.com/docs/en/about-claude/pricing",
+      "retrieved_at": "2026-10-09",
+      "note": "used for claude-haiku-5-5 (two Haiku 5.5 rows: prompts up to 100K tokens $0.10 / $0.125 / $0.20 / $0.01 / $0.50, prompts over 100K tokens $0.50 / $0.625 / $1 / $0.05 / $2.50; no footnote so cache_read is the standard 0.1x; fast-mode table lists Opus models only); Long context pricing section: a request's prompt length counts all of its input tokens including cache reads and cache writes, and each request is priced on its own"
+    },
+    {
+      "url": "https://platform.claude.com/docs/en/models/haiku-5-5/overview",
+      "retrieved_at": "2026-10-09",
+      "note": "Haiku 5.5 overview: Released October 7, 2026, model ID claude-haiku-5-5, Retirement not sooner than October 7, 2027"
+    },
+    {
+      "url": "https://github.com/anthropics/claude-code/blob/main/CHANGELOG.md",
+      "retrieved_at": "2026-10-09",
+      "note": "Claude Code 2.1.293 entry: Added Claude Haiku 5.5 (claude-haiku-5-5), now the default Haiku model; 1M context, $0.10/$0.50 per MTok ($0.50/$2.50 for prompts over 100K); the reason claude-haiku-5-5 rows are expected in transcripts from that version"
     }
   ],
   "models": [
@@ -314,6 +330,33 @@
         }
       ]
     },
+    {
+      "model_key": "claude-haiku-5-5",
+      "aliases": [],
+      "fast_multiplier": "1.0",
+      "rates": [
+        {
+          "rate_id": "claude-haiku-5-5-launch-2026-10-07",
+          "effective_from": "2026-10-07T00:00:00+00:00",
+          "effective_until": null,
+          "input_nocache": "0.10",
+          "cache_read": "0.01",
+          "cache_write_5m": "0.125",
+          "cache_write_1h": "0.20",
+          "output": "0.50",
+          "prompt_tiers": [
+            {
+              "prompt_tokens_over": 100000,
+              "input_nocache": "0.50",
+              "cache_read": "0.05",
+              "cache_write_5m": "0.625",
+              "cache_write_1h": "1.0",
+              "output": "2.50"
+            }
+          ]
+        }
+      ]
+    },
     {
       "model_key": "gpt-5.6-sol",
       "aliases": [],

{coding_agent_cost-0.2.2 → coding_agent_cost-0.4.0}/agent_cost/rates.py

@@ -16,7 +16,7 @@ from datetime import datetime, timezone
 from decimal import Decimal, InvalidOperation
 from importlib import resources
 from pathlib import Path
-from typing import Optional
+from typing import Optional, Tuple
 
 RATE_FIELDS = (
     "input_nocache",
@@ -30,8 +30,11 @@ RATE_FIELDS = (
 # knows how to interpret. A catalog claiming anything else is rejected
 # outright rather than silently treated as USD-per-mtok -- a catalog in a
 # different currency or unit priced as if it were USD-per-mtok would be
-# wrong by a fixed, silent factor.
-SUPPORTED_SCHEMA_VERSIONS = ("1",)
+# wrong by a fixed, silent factor. Schema "2" adds optional per-period
+# ``prompt_tiers``; it is a separate version (and "1" may not carry tiers)
+# so that a reader that predates tiers rejects a tiered catalog as
+# unsupported instead of silently pricing every request at the base tier.
+SUPPORTED_SCHEMA_VERSIONS = ("1", "2")
 SUPPORTED_CURRENCIES = ("USD",)
 SUPPORTED_UNITS = ("per_mtok",)
 
@@ -40,12 +43,28 @@ class RatesValidationError(ValueError):
     """The rates catalog failed structural or business-rule validation."""
 
 
+@dataclass(frozen=True)
+class PromptTier:
+    """Rate values for requests whose prompt length is strictly greater
+    than ``prompt_tokens_over``."""
+
+    prompt_tokens_over: int
+    values: dict
+
+
 @dataclass(frozen=True)
 class RatePeriod:
     rate_id: str
     effective_from: datetime
     effective_until: Optional[datetime]
+    # Base tier: prompt length <= the first tier's threshold (or every
+    # request, when there are no tiers).
     values: dict
+    # Ascending by threshold; a request uses the tier with the largest
+    # threshold its prompt length exceeds. Every value in a tiered period
+    # is non-null and non-decreasing tier to tier, so ``values`` is the
+    # cheapest (a lower bound) when the prompt length is unknown.
+    prompt_tiers: Tuple[PromptTier, ...] = ()
 
     def covers(self, moment: datetime) -> bool:
         if moment < self.effective_from:
@@ -125,7 +144,46 @@ def _parse_datetime(value, *, field_name: str) -> datetime:
     return dt.astimezone(timezone.utc)
 
 
-def _build_model_entry(raw: dict) -> ModelEntry:
+def _build_prompt_tiers(tiers_raw, base_values: dict, *, where: str) -> Tuple[PromptTier, ...]:
+    if not isinstance(tiers_raw, list):
+        raise RatesValidationError(f"{where}: prompt_tiers must be a list")
+    if not tiers_raw:
+        return ()
+
+    for f in RATE_FIELDS:
+        if base_values[f] is None:
+            raise RatesValidationError(f"{where}: {f} is required when prompt_tiers is present")
+
+    tiers = []
+    previous_values = base_values
+    previous_over = None
+    for idx, tier_raw in enumerate(tiers_raw):
+        tier_where = f"{where}.prompt_tiers[{idx}]"
+        if not isinstance(tier_raw, dict):
+            raise RatesValidationError(f"{tier_where}: must be an object")
+        over = tier_raw.get("prompt_tokens_over")
+        if isinstance(over, bool) or not isinstance(over, int) or over <= 0:
+            raise RatesValidationError(f"{tier_where}: prompt_tokens_over must be a positive int, got {over!r}")
+        if previous_over is not None and over <= previous_over:
+            raise RatesValidationError(
+                f"{tier_where}: prompt_tokens_over must be ascending without duplicates "
+                f"({over} after {previous_over})"
+            )
+        values = {f: _parse_decimal(tier_raw.get(f), field_name=f"{tier_where}.{f}") for f in RATE_FIELDS}
+        for f in RATE_FIELDS:
+            if values[f] is None:
+                raise RatesValidationError(f"{tier_where}: {f} is required")
+            if values[f] < previous_values[f]:
+                raise RatesValidationError(
+                    f"{tier_where}: {f} {values[f]} is lower than the previous tier's {previous_values[f]}"
+                )
+        tiers.append(PromptTier(prompt_tokens_over=over, values=values))
+        previous_values = values
+        previous_over = over
+    return tuple(tiers)
+
+
+def _build_model_entry(raw: dict, *, schema_version: str) -> ModelEntry:
     model_key = raw.get("model_key")
     if not model_key or not isinstance(model_key, str):
         raise RatesValidationError(f"model entry missing model_key: {raw!r}")
@@ -169,12 +227,22 @@ def _build_model_entry(raw: dict) -> ModelEntry:
             f: _parse_decimal(rate_raw.get(f), field_name=f"{model_key}.{rate_id}.{f}")
             for f in RATE_FIELDS
         }
+        prompt_tiers: Tuple[PromptTier, ...] = ()
+        if "prompt_tiers" in rate_raw:
+            if schema_version == "1":
+                raise RatesValidationError(
+                    f"{model_key}.{rate_id}: prompt_tiers requires schema_version \"2\""
+                )
+            prompt_tiers = _build_prompt_tiers(
+                rate_raw["prompt_tiers"], values, where=f"{model_key}.{rate_id}"
+            )
         periods.append(
             RatePeriod(
                 rate_id=rate_id,
                 effective_from=effective_from,
                 effective_until=effective_until,
                 values=values,
+                prompt_tiers=prompt_tiers,
             )
         )
 
@@ -227,7 +295,7 @@ def _validate_and_build(data: dict) -> RateCatalog:
 
     models: dict = {}
     for raw in data.get("models") or []:
-        entry = _build_model_entry(raw)
+        entry = _build_model_entry(raw, schema_version=schema_version)
         if entry.model_key in models:
             raise RatesValidationError(f"duplicate model_key: {entry.model_key}")
         models[entry.model_key] = entry