pennyrouter 0.3.38 → 0.3.40

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "pennyrouter",
-  "version": "0.3.38",
+  "version": "0.3.40",
   "description": "Install and manage PennyRouter local coding-agent integrations.",
   "homepage": "https://pennyrouter.com",
   "bugs": {

package/src/cli.js

@@ -2035,8 +2035,6 @@ async function uninstall(flags) {
 
   if (anthropicTokenNotice) {
     console.log("");
-    console.log("Note: Your stored Anthropic token remains on the server for other machines.");
-    console.log("To delete it from the server, run: pennyrouter auth anthropic --forget");
   }
 
   manifest.installations = kept;

package/src/launch.js

@@ -320,6 +320,16 @@ export async function installPennyCompare() {
     await mkdir(dirname(target), { recursive: true });
     await cp(await findInstalledPackage(name), target, { recursive: true });
   }
+  if (process.platform === "darwin") {
+    // node-pty 1.1.0's macOS prebuild ships a non-executable spawn-helper, and the copy
+    // keeps that mode; without +x every PTY launch fails at posix_spawnp.
+    const helper = join(modules, "node-pty", "prebuilds", `darwin-${process.arch}`, "spawn-helper");
+    try {
+      await chmod(helper, (await stat(helper)).mode | 0o111);
+    } catch (error) {
+      if (error.code !== "ENOENT") throw error;
+    }
+  }
   if (process.platform === "win32") {
     const link = join(pennyBinDir(), "penny-compare.cmd");
     const marker = "@REM PennyRouter-managed compare launcher";
@@ -917,6 +927,9 @@ export async function writeClaudeLaunchSettings({
             // An empty value is treated as unset by Claude Code, and prevents a
             // stale API key in the shared settings from creating dual auth.
             ANTHROPIC_API_KEY: "",
+            // Claude Code turns tool search off behind a custom base URL unless told
+            // otherwise; Penny relays deferred tools and their tool_reference blocks.
+            ENABLE_TOOL_SEARCH: "true",
             ANTHROPIC_DEFAULT_HAIKU_MODEL: "",
             ANTHROPIC_DEFAULT_HAIKU_MODEL_NAME: "",
             ANTHROPIC_DEFAULT_HAIKU_MODEL_DESCRIPTION: "",
@@ -1413,6 +1426,7 @@ export function buildPennyInvocation(harness, args, config, models = null) {
     // Ask Claude Code to discover the gateway's model catalog at startup. The gateway
     // advertises Penny-facing ids in /v1/models; this is separate from inference routing.
     env.CLAUDE_CODE_ENABLE_GATEWAY_MODEL_DISCOVERY = "1";
+    env.ENABLE_TOOL_SEARCH = "true";
     delete env.ANTHROPIC_API_KEY;
     for (const slot of ["haiku", "sonnet", "opus", "fable"]) {
       const name = `ANTHROPIC_DEFAULT_${slot.toUpperCase()}_MODEL`;

package/src/launch.test.js

@@ -139,6 +139,7 @@ async function main() {
     assert.equal(claudeLaunchSettings.env.ANTHROPIC_BASE_URL, "https://gateway.example/");
     assert.equal(claudeLaunchSettings.env.ANTHROPIC_AUTH_TOKEN, "pr-super-secret");
     assert.equal(claudeLaunchSettings.env.ANTHROPIC_API_KEY, "");
+    assert.equal(claudeLaunchSettings.env.ENABLE_TOOL_SEARCH, "true");
     assert.equal(claudeLaunchSettings.env.ANTHROPIC_DEFAULT_OPUS_MODEL, "");
     assert.equal(claudeLaunchSettings.env.ANTHROPIC_DEFAULT_OPUS_MODEL_NAME, "");
     await writeClaudeLaunchSettings({
@@ -264,6 +265,7 @@ async function main() {
     assert.equal(claude.env.ANTHROPIC_AUTH_TOKEN, "pr-super-secret");
     assert.equal(claude.env.ANTHROPIC_BASE_URL, "https://gateway.example/");
     assert.equal(claude.env.CLAUDE_CODE_ENABLE_GATEWAY_MODEL_DISCOVERY, "1");
+    assert.equal(claude.env.ENABLE_TOOL_SEARCH, "true");
     // 1M-capable slots carry the `[1m]` suffix; Haiku has no 1M variant and stays bare.
     // Per-slot suffixes, not the global CLAUDE_CODE_MAX_CONTEXT_TOKENS override, which
     // would wrongly hand Haiku a 1M window too.

package/standalone/pricing.py

@@ -7,18 +7,27 @@ that uses 1-hour caching, which the gateway does by default.
 
 from __future__ import annotations
 
-# input, cache_read, write_5m, write_1h, output
-RATES = {
-    "claude-opus-5":     (5.00e-06, 5.00e-07, 6.25e-06, 1.00e-05, 2.50e-05),
-    "claude-sonnet-5":   (3.00e-06, 3.00e-07, 3.75e-06, 6.00e-06, 1.50e-05),
-    "claude-haiku-4-5":  (1.00e-06, 1.00e-07, 1.25e-06, 2.00e-06, 5.00e-06),
-    # OpenAI does not offer a 1-hour cache tier; its write rate applies uniformly.
-    "gpt-5-codex":       (1.25e-06, 1.25e-07, 1.25e-06, 1.25e-06, 1.00e-05),
-    "gpt-5.1-codex":     (1.25e-06, 1.25e-07, 1.25e-06, 1.25e-06, 1.00e-05),
-    # GPT-5.6 standard pricing. Keep Luna explicit so penny-session does not silently
-    # price it as Claude Sonnet when no GovCloud profile was selected.
-    "gpt-5.6-luna":      (0.20e-06, 0.020e-06, 0.250e-06, 0.250e-06, 1.20e-06),
-}
+import json
+from pathlib import Path
+
+
+def _load_standard_rates():
+    """Standard per-token rates from rates.json (scripts/generate_compare_rates.py), the same
+    file the web-demo meters read. Returns (flat rates, tiered rates) keyed by model id; a tier
+    entry is (inclusive threshold, <= threshold rates, > threshold rates)."""
+    document = json.loads((Path(__file__).with_name("rates.json")).read_text())
+    flat, tiered = {}, {}
+    for name, model in document["models"].items():
+        flat[name] = tuple(model["rates"])
+        if "above" in model:
+            above = model["above"]
+            tiered[name] = (above["threshold"], tuple(model["rates"]), tuple(above["rates"]))
+    return flat, tiered
+
+
+# input, cache_read, write_5m, write_1h, output. OpenAI does not offer a 1-hour cache tier;
+# its write rate applies uniformly.
+RATES, STANDARD_TIERED_RATES = _load_standard_rates()
 
 # --native-proxy means the comparison is measuring the GovCloud route. Pricing remains the
 # same on both panes, but both use this schedule instead of the standard one.
@@ -29,9 +38,7 @@ GOVCLOUD_RATES = {
     "claude-haiku-4-5":  (1.10e-06, 1.10e-07, 1.375e-06, 2.20e-06, 5.50e-06),
 }
 
-# GPT GovCloud pricing changes when an individual request's input context exceeds 272K.
-# Each entry is (inclusive threshold, <= threshold rates, > threshold rates), with rate tuples
-# in the same input/read/write-5m/write-1h/output order as RATES.
+# The GovCloud schedule for the same tiered GPT models.
 GOVCLOUD_TIERED_RATES = {
     "gpt-5.6-luna": (
         272_000,
@@ -88,11 +95,11 @@ def _matching(name: str, choices) -> str | None:
 
 def rates_for(model: str, profile: str = "standard", context_tokens: int = 0):
     name = _normalize(model)
-    if profile == "govcloud":
-        tiered = _matching(name, GOVCLOUD_TIERED_RATES)
-        if tiered:
-            threshold, low, high = GOVCLOUD_TIERED_RATES[tiered]
-            return high if context_tokens > threshold else low
+    tiers = GOVCLOUD_TIERED_RATES if profile == "govcloud" else STANDARD_TIERED_RATES
+    tiered = _matching(name, tiers)
+    if tiered:
+        threshold, low, high = tiers[tiered]
+        return high if context_tokens > threshold else low
     rates = GOVCLOUD_RATES if profile == "govcloud" else RATES
     match = _matching(name, rates)
     if match:

package/standalone/rates.json

@@ -0,0 +1,1078 @@
+{
+ "source": "bifrost_datasheet",
+ "generated_at": "2026-09-29T15:31:58+00:00",
+ "order": [
+  "input",
+  "cache_read",
+  "cache_write_5m",
+  "cache_write_1h",
+  "output"
+ ],
+ "unit": "usd_per_token",
+ "models": {
+  "claude-fable-5": {
+   "rates": [
+    1e-05,
+    1e-06,
+    1.25e-05,
+    2e-05,
+    5e-05
+   ]
+  },
+  "claude-fable-5-1": {
+   "rates": [
+    1e-05,
+    2.5e-07,
+    1.25e-05,
+    2e-05,
+    5e-05
+   ]
+  },
+  "claude-haiku-4-5": {
+   "rates": [
+    1e-06,
+    1e-07,
+    1.25e-06,
+    2e-06,
+    5e-06
+   ]
+  },
+  "claude-haiku-4-5-20251001": {
+   "rates": [
+    1e-06,
+    1e-07,
+    1.25e-06,
+    2e-06,
+    5e-06
+   ]
+  },
+  "claude-mythos-5": {
+   "rates": [
+    1e-05,
+    1e-06,
+    1.25e-05,
+    2e-05,
+    5e-05
+   ]
+  },
+  "claude-mythos-5-1": {
+   "rates": [
+    1e-05,
+    2.5e-07,
+    1.25e-05,
+    2e-05,
+    5e-05
+   ]
+  },
+  "claude-mythos-preview": {
+   "rates": [
+    1e-05,
+    1e-06,
+    1.25e-05,
+    2e-05,
+    5e-05
+   ]
+  },
+  "claude-opus-4-5": {
+   "rates": [
+    5e-06,
+    5e-07,
+    6.25e-06,
+    1e-05,
+    2.5e-05
+   ]
+  },
+  "claude-opus-4-5-20251101": {
+   "rates": [
+    5e-06,
+    5e-07,
+    6.25e-06,
+    1e-05,
+    2.5e-05
+   ]
+  },
+  "claude-opus-4-6": {
+   "rates": [
+    5e-06,
+    5e-07,
+    6.25e-06,
+    1e-05,
+    2.5e-05
+   ]
+  },
+  "claude-opus-4-6-20260205": {
+   "rates": [
+    5e-06,
+    5e-07,
+    6.25e-06,
+    1e-05,
+    2.5e-05
+   ]
+  },
+  "claude-opus-4-7": {
+   "rates": [
+    5e-06,
+    5e-07,
+    6.25e-06,
+    1e-05,
+    2.5e-05
+   ]
+  },
+  "claude-opus-4-7-20260416": {
+   "rates": [
+    5e-06,
+    5e-07,
+    6.25e-06,
+    1e-05,
+    2.5e-05
+   ]
+  },
+  "claude-opus-4-8": {
+   "rates": [
+    5e-06,
+    5e-07,
+    6.25e-06,
+    1e-05,
+    2.5e-05
+   ]
+  },
+  "claude-opus-5": {
+   "rates": [
+    5e-06,
+    5e-07,
+    6.25e-06,
+    1e-05,
+    2.5e-05
+   ]
+  },
+  "claude-opus-5-5": {
+   "rates": [
+    4e-06,
+    2e-07,
+    5e-06,
+    8e-06,
+    2e-05
+   ]
+  },
+  "claude-sonnet-4-5": {
+   "rates": [
+    3e-06,
+    3e-07,
+    3.75e-06,
+    6e-06,
+    1.5e-05
+   ],
+   "above": {
+    "threshold": 200000,
+    "rates": [
+     6e-06,
+     6e-07,
+     7.5e-06,
+     1.2e-05,
+     2.25e-05
+    ]
+   }
+  },
+  "claude-sonnet-4-5-20250929": {
+   "rates": [
+    3e-06,
+    3e-07,
+    3.75e-06,
+    6e-06,
+    1.5e-05
+   ],
+   "above": {
+    "threshold": 200000,
+    "rates": [
+     6e-06,
+     6e-07,
+     7.5e-06,
+     1.2e-05,
+     2.25e-05
+    ]
+   }
+  },
+  "claude-sonnet-4-6": {
+   "rates": [
+    3e-06,
+    3e-07,
+    3.75e-06,
+    6e-06,
+    1.5e-05
+   ]
+  },
+  "claude-sonnet-5": {
+   "rates": [
+    2e-06,
+    2e-07,
+    2.5e-06,
+    4e-06,
+    1e-05
+   ]
+  },
+  "claude-sonnet-5-5": {
+   "rates": [
+    2e-06,
+    2e-07,
+    2.5e-06,
+    4e-06,
+    1e-05
+   ]
+  },
+  "gpt-3.5-turbo": {
+   "rates": [
+    5e-07,
+    5e-07,
+    5e-07,
+    5e-07,
+    1.5e-06
+   ]
+  },
+  "gpt-3.5-turbo-0125": {
+   "rates": [
+    5e-07,
+    5e-07,
+    5e-07,
+    5e-07,
+    1.5e-06
+   ]
+  },
+  "gpt-3.5-turbo-1106": {
+   "rates": [
+    1e-06,
+    1e-06,
+    1e-06,
+    1e-06,
+    2e-06
+   ]
+  },
+  "gpt-3.5-turbo-16k": {
+   "rates": [
+    3e-06,
+    3e-06,
+    3e-06,
+    3e-06,
+    4e-06
+   ]
+  },
+  "gpt-3.5-turbo-instruct": {
+   "rates": [
+    1.5e-06,
+    1.5e-06,
+    1.5e-06,
+    1.5e-06,
+    2e-06
+   ]
+  },
+  "gpt-3.5-turbo-instruct-0914": {
+   "rates": [
+    1.5e-06,
+    1.5e-06,
+    1.5e-06,
+    1.5e-06,
+    2e-06
+   ]
+  },
+  "gpt-4": {
+   "rates": [
+    3e-05,
+    3e-05,
+    3e-05,
+    3e-05,
+    6e-05
+   ]
+  },
+  "gpt-4-0613": {
+   "rates": [
+    3e-05,
+    3e-05,
+    3e-05,
+    3e-05,
+    6e-05
+   ]
+  },
+  "gpt-4-1106-preview": {
+   "rates": [
+    1e-05,
+    1e-05,
+    1e-05,
+    1e-05,
+    3e-05
+   ]
+  },
+  "gpt-4-turbo": {
+   "rates": [
+    1e-05,
+    1e-05,
+    1e-05,
+    1e-05,
+    3e-05
+   ]
+  },
+  "gpt-4-turbo-2024-04-09": {
+   "rates": [
+    1e-05,
+    1e-05,
+    1e-05,
+    1e-05,
+    3e-05
+   ]
+  },
+  "gpt-4.1": {
+   "rates": [
+    2e-06,
+    5e-07,
+    2e-06,
+    2e-06,
+    8e-06
+   ]
+  },
+  "gpt-4.1-2025-04-14": {
+   "rates": [
+    2e-06,
+    5e-07,
+    2e-06,
+    2e-06,
+    8e-06
+   ]
+  },
+  "gpt-4.1-mini": {
+   "rates": [
+    4e-07,
+    1e-07,
+    4e-07,
+    4e-07,
+    1.6e-06
+   ]
+  },
+  "gpt-4.1-mini-2025-04-14": {
+   "rates": [
+    4e-07,
+    1e-07,
+    4e-07,
+    4e-07,
+    1.6e-06
+   ]
+  },
+  "gpt-4.1-nano": {
+   "rates": [
+    1e-07,
+    2.5e-08,
+    1e-07,
+    1e-07,
+    4e-07
+   ]
+  },
+  "gpt-4.1-nano-2025-04-14": {
+   "rates": [
+    1e-07,
+    2.5e-08,
+    1e-07,
+    1e-07,
+    4e-07
+   ]
+  },
+  "gpt-4o": {
+   "rates": [
+    2.5e-06,
+    1.25e-06,
+    2.5e-06,
+    2.5e-06,
+    1e-05
+   ]
+  },
+  "gpt-4o-2024-05-13": {
+   "rates": [
+    5e-06,
+    5e-06,
+    5e-06,
+    5e-06,
+    1.5e-05
+   ]
+  },
+  "gpt-4o-2024-08-06": {
+   "rates": [
+    2.5e-06,
+    1.25e-06,
+    2.5e-06,
+    2.5e-06,
+    1e-05
+   ]
+  },
+  "gpt-4o-2024-11-20": {
+   "rates": [
+    2.5e-06,
+    1.25e-06,
+    2.5e-06,
+    2.5e-06,
+    1e-05
+   ]
+  },
+  "gpt-4o-mini": {
+   "rates": [
+    1.5e-07,
+    7.5e-08,
+    1.5e-07,
+    1.5e-07,
+    6e-07
+   ]
+  },
+  "gpt-4o-mini-2024-07-18": {
+   "rates": [
+    1.5e-07,
+    7.5e-08,
+    1.5e-07,
+    1.5e-07,
+    6e-07
+   ]
+  },
+  "gpt-5": {
+   "rates": [
+    1.25e-06,
+    1.25e-07,
+    1.25e-06,
+    1.25e-06,
+    1e-05
+   ]
+  },
+  "gpt-5-2025-08-07": {
+   "rates": [
+    1.25e-06,
+    1.25e-07,
+    1.25e-06,
+    1.25e-06,
+    1e-05
+   ]
+  },
+  "gpt-5-chat": {
+   "rates": [
+    1.25e-06,
+    1.25e-07,
+    1.25e-06,
+    1.25e-06,
+    1e-05
+   ]
+  },
+  "gpt-5-chat-latest": {
+   "rates": [
+    1.25e-06,
+    1.25e-07,
+    1.25e-06,
+    1.25e-06,
+    1e-05
+   ]
+  },
+  "gpt-5-codex": {
+   "rates": [
+    1.25e-06,
+    1.25e-07,
+    1.25e-06,
+    1.25e-06,
+    1e-05
+   ]
+  },
+  "gpt-5-mini": {
+   "rates": [
+    2.5e-07,
+    2.5e-08,
+    2.5e-07,
+    2.5e-07,
+    2e-06
+   ]
+  },
+  "gpt-5-mini-2025-08-07": {
+   "rates": [
+    2.5e-07,
+    2.5e-08,
+    2.5e-07,
+    2.5e-07,
+    2e-06
+   ]
+  },
+  "gpt-5-nano": {
+   "rates": [
+    5e-08,
+    5e-09,
+    5e-08,
+    5e-08,
+    4e-07
+   ]
+  },
+  "gpt-5-nano-2025-08-07": {
+   "rates": [
+    5e-08,
+    5e-09,
+    5e-08,
+    5e-08,
+    4e-07
+   ]
+  },
+  "gpt-5-pro": {
+   "rates": [
+    1.5e-05,
+    1.5e-05,
+    1.5e-05,
+    1.5e-05,
+    0.00012
+   ]
+  },
+  "gpt-5-pro-2025-10-06": {
+   "rates": [
+    1.5e-05,
+    1.5e-05,
+    1.5e-05,
+    1.5e-05,
+    0.00012
+   ]
+  },
+  "gpt-5-search-api": {
+   "rates": [
+    1.25e-06,
+    1.25e-07,
+    1.25e-06,
+    1.25e-06,
+    1e-05
+   ]
+  },
+  "gpt-5-search-api-2025-10-14": {
+   "rates": [
+    1.25e-06,
+    1.25e-07,
+    1.25e-06,
+    1.25e-06,
+    1e-05
+   ]
+  },
+  "gpt-5.1": {
+   "rates": [
+    1.25e-06,
+    1.25e-07,
+    1.25e-06,
+    1.25e-06,
+    1e-05
+   ]
+  },
+  "gpt-5.1-2025-11-13": {
+   "rates": [
+    1.25e-06,
+    1.25e-07,
+    1.25e-06,
+    1.25e-06,
+    1e-05
+   ]
+  },
+  "gpt-5.1-chat-latest": {
+   "rates": [
+    1.25e-06,
+    1.25e-07,
+    1.25e-06,
+    1.25e-06,
+    1e-05
+   ]
+  },
+  "gpt-5.1-codex": {
+   "rates": [
+    1.25e-06,
+    1.25e-07,
+    1.25e-06,
+    1.25e-06,
+    1e-05
+   ]
+  },
+  "gpt-5.1-codex-max": {
+   "rates": [
+    1.25e-06,
+    1.25e-07,
+    1.25e-06,
+    1.25e-06,
+    1e-05
+   ]
+  },
+  "gpt-5.1-codex-mini": {
+   "rates": [
+    2.5e-07,
+    2.5e-08,
+    2.5e-07,
+    2.5e-07,
+    2e-06
+   ]
+  },
+  "gpt-5.2": {
+   "rates": [
+    1.75e-06,
+    1.75e-07,
+    1.75e-06,
+    1.75e-06,
+    1.4e-05
+   ]
+  },
+  "gpt-5.2-2025-12-11": {
+   "rates": [
+    1.75e-06,
+    1.75e-07,
+    1.75e-06,
+    1.75e-06,
+    1.4e-05
+   ]
+  },
+  "gpt-5.2-chat-latest": {
+   "rates": [
+    1.75e-06,
+    1.75e-07,
+    1.75e-06,
+    1.75e-06,
+    1.4e-05
+   ]
+  },
+  "gpt-5.2-codex": {
+   "rates": [
+    1.75e-06,
+    1.75e-07,
+    1.75e-06,
+    1.75e-06,
+    1.4e-05
+   ]
+  },
+  "gpt-5.2-pro": {
+   "rates": [
+    2.1e-05,
+    2.1e-05,
+    2.1e-05,
+    2.1e-05,
+    0.000168
+   ]
+  },
+  "gpt-5.2-pro-2025-12-11": {
+   "rates": [
+    2.1e-05,
+    2.1e-05,
+    2.1e-05,
+    2.1e-05,
+    0.000168
+   ]
+  },
+  "gpt-5.3-chat-latest": {
+   "rates": [
+    1.75e-06,
+    1.75e-07,
+    1.75e-06,
+    1.75e-06,
+    1.4e-05
+   ]
+  },
+  "gpt-5.3-codex": {
+   "rates": [
+    1.75e-06,
+    1.75e-07,
+    1.75e-06,
+    1.75e-06,
+    1.4e-05
+   ]
+  },
+  "gpt-5.3-codex-spark": {
+   "rates": [
+    1.75e-06,
+    1.75e-07,
+    1.75e-06,
+    1.75e-06,
+    1.4e-05
+   ]
+  },
+  "gpt-5.4": {
+   "rates": [
+    2.5e-06,
+    2.5e-07,
+    2.5e-06,
+    2.5e-06,
+    1.5e-05
+   ],
+   "above": {
+    "threshold": 272000,
+    "rates": [
+     5e-06,
+     5e-07,
+     5e-06,
+     5e-06,
+     2.25e-05
+    ]
+   }
+  },
+  "gpt-5.4-2026-03-05": {
+   "rates": [
+    2.5e-06,
+    2.5e-07,
+    2.5e-06,
+    2.5e-06,
+    1.5e-05
+   ],
+   "above": {
+    "threshold": 272000,
+    "rates": [
+     5e-06,
+     5e-07,
+     5e-06,
+     5e-06,
+     2.25e-05
+    ]
+   }
+  },
+  "gpt-5.4-mini": {
+   "rates": [
+    7.5e-07,
+    7.5e-08,
+    7.5e-07,
+    7.5e-07,
+    4.5e-06
+   ]
+  },
+  "gpt-5.4-mini-2026-03-17": {
+   "rates": [
+    7.5e-07,
+    7.5e-08,
+    7.5e-07,
+    7.5e-07,
+    4.5e-06
+   ]
+  },
+  "gpt-5.4-nano": {
+   "rates": [
+    2e-07,
+    2e-08,
+    2e-07,
+    2e-07,
+    1.25e-06
+   ]
+  },
+  "gpt-5.4-nano-2026-03-17": {
+   "rates": [
+    2e-07,
+    2e-08,
+    2e-07,
+    2e-07,
+    1.25e-06
+   ]
+  },
+  "gpt-5.4-pro": {
+   "rates": [
+    3e-05,
+    3e-06,
+    3e-05,
+    3e-05,
+    0.00018
+   ],
+   "above": {
+    "threshold": 272000,
+    "rates": [
+     6e-05,
+     6e-06,
+     6e-05,
+     6e-05,
+     0.00027
+    ]
+   }
+  },
+  "gpt-5.4-pro-2026-03-05": {
+   "rates": [
+    3e-05,
+    3e-06,
+    3e-05,
+    3e-05,
+    0.00018
+   ],
+   "above": {
+    "threshold": 272000,
+    "rates": [
+     6e-05,
+     6e-06,
+     6e-05,
+     6e-05,
+     0.00027
+    ]
+   }
+  },
+  "gpt-5.5": {
+   "rates": [
+    5e-06,
+    5e-07,
+    5e-06,
+    5e-06,
+    3e-05
+   ],
+   "above": {
+    "threshold": 272000,
+    "rates": [
+     1e-05,
+     1e-06,
+     1e-05,
+     1e-05,
+     4.5e-05
+    ]
+   }
+  },
+  "gpt-5.5-2026-04-23": {
+   "rates": [
+    5e-06,
+    5e-07,
+    5e-06,
+    5e-06,
+    3e-05
+   ],
+   "above": {
+    "threshold": 272000,
+    "rates": [
+     1e-05,
+     1e-06,
+     1e-05,
+     1e-05,
+     4.5e-05
+    ]
+   }
+  },
+  "gpt-5.5-cyber": {
+   "rates": [
+    1.25e-05,
+    1.25e-06,
+    1.25e-05,
+    1.25e-05,
+    7.5e-05
+   ]
+  },
+  "gpt-5.5-pro": {
+   "rates": [
+    3e-05,
+    3e-06,
+    3e-05,
+    3e-05,
+    0.00018
+   ],
+   "above": {
+    "threshold": 272000,
+    "rates": [
+     6e-05,
+     6e-06,
+     6e-05,
+     6e-05,
+     0.00027
+    ]
+   }
+  },
+  "gpt-5.5-pro-2026-04-23": {
+   "rates": [
+    3e-05,
+    3e-06,
+    3e-05,
+    3e-05,
+    0.00018
+   ],
+   "above": {
+    "threshold": 272000,
+    "rates": [
+     6e-05,
+     6e-06,
+     6e-05,
+     6e-05,
+     0.00027
+    ]
+   }
+  },
+  "gpt-5.6": {
+   "rates": [
+    4e-06,
+    4e-07,
+    5e-06,
+    5e-06,
+    2e-05
+   ],
+   "above": {
+    "threshold": 272000,
+    "rates": [
+     8e-06,
+     8e-07,
+     1e-05,
+     1e-05,
+     3e-05
+    ]
+   }
+  },
+  "gpt-5.6-cyber": {
+   "rates": [
+    1.25e-05,
+    1.25e-06,
+    1.5625e-05,
+    1.5625e-05,
+    7.5e-05
+   ],
+   "above": {
+    "threshold": 272000,
+    "rates": [
+     2.5e-05,
+     2.5e-06,
+     3.125e-05,
+     3.125e-05,
+     0.0001125
+    ]
+   }
+  },
+  "gpt-5.6-luna": {
+   "rates": [
+    2e-07,
+    2e-08,
+    2.5e-07,
+    2.5e-07,
+    1.2e-06
+   ],
+   "above": {
+    "threshold": 272000,
+    "rates": [
+     4e-07,
+     4e-08,
+     5e-07,
+     5e-07,
+     1.8e-06
+    ]
+   }
+  },
+  "gpt-5.6-sol": {
+   "rates": [
+    4e-06,
+    4e-07,
+    5e-06,
+    5e-06,
+    2e-05
+   ],
+   "above": {
+    "threshold": 272000,
+    "rates": [
+     8e-06,
+     8e-07,
+     1e-05,
+     1e-05,
+     3e-05
+    ]
+   }
+  },
+  "gpt-5.6-terra": {
+   "rates": [
+    2e-06,
+    2e-07,
+    2.5e-06,
+    2.5e-06,
+    1.2e-05
+   ],
+   "above": {
+    "threshold": 272000,
+    "rates": [
+     4e-06,
+     4e-07,
+     5e-06,
+     5e-06,
+     1.8e-05
+    ]
+   }
+  },
+  "gpt-6-astra": {
+   "rates": [
+    1e-05,
+    1e-06,
+    1.25e-05,
+    1.25e-05,
+    5e-05
+   ],
+   "above": {
+    "threshold": 272000,
+    "rates": [
+     2e-05,
+     2e-06,
+     2.5e-05,
+     2.5e-05,
+     7.5e-05
+    ]
+   }
+  },
+  "gpt-6-luna": {
+   "rates": [
+    1e-07,
+    1e-08,
+    1.25e-07,
+    1.25e-07,
+    5e-07
+   ],
+   "above": {
+    "threshold": 272000,
+    "rates": [
+     2e-07,
+     2e-08,
+     2.5e-07,
+     2.5e-07,
+     7.5e-07
+    ]
+   }
+  },
+  "gpt-6-sol": {
+   "rates": [
+    2e-06,
+    2e-07,
+    2.5e-06,
+    2.5e-06,
+    1e-05
+   ],
+   "above": {
+    "threshold": 272000,
+    "rates": [
+     4e-06,
+     4e-07,
+     5e-06,
+     5e-06,
+     1.5e-05
+    ]
+   }
+  },
+  "gpt-daybreak-blue-latest": {
+   "rates": [
+    4e-06,
+    4e-07,
+    5e-06,
+    5e-06,
+    2e-05
+   ],
+   "above": {
+    "threshold": 272000,
+    "rates": [
+     8e-06,
+     8e-07,
+     1e-05,
+     1e-05,
+     3e-05
+    ]
+   }
+  },
+  "gpt-daybreak-red-latest": {
+   "rates": [
+    1.25e-05,
+    1.25e-06,
+    1.5625e-05,
+    1.5625e-05,
+    7.5e-05
+   ],
+   "above": {
+    "threshold": 272000,
+    "rates": [
+     2.5e-05,
+     2.5e-06,
+     3.125e-05,
+     3.125e-05,
+     0.0001125
+    ]
+   }
+  },
+  "gpt-rosalind-research": {
+   "rates": [
+    5e-06,
+    5e-07,
+    5e-06,
+    5e-06,
+    2.5e-05
+   ]
+  }
+ }
+}

package/standalone/web-demo/claude-meter.mjs

@@ -1,18 +1,44 @@
+import { readFileSync } from 'node:fs';
 import { mkdir, open, readdir, stat, writeFile } from 'node:fs/promises';
 import { homedir } from 'node:os';
 import { join } from 'node:path';
 
-// Mirrors the standard Claude rates in ../pricing.py (USD per token).
-const rates = {
-  'claude-opus-5': [5e-6, .5e-6, 6.25e-6, 10e-6, 25e-6],
-  'claude-sonnet-5': [3e-6, .3e-6, 3.75e-6, 6e-6, 15e-6],
-  'claude-haiku-4-5': [1e-6, .1e-6, 1.25e-6, 2e-6, 5e-6],
-};
+// Standard per-token rates from ../rates.json (scripts/generate_compare_rates.py), the same
+// file pricing.py reads. Each rate list is [input, cache read, cache write 5m, cache write 1h,
+// output]; a model may carry an `above` list that applies once one request's input context
+// exceeds its threshold.
+const { models } = JSON.parse(readFileSync(new URL('../rates.json', import.meta.url), 'utf8'));
 
-function rateFor(model) {
-  const normalized = String(model || '').replace(/^.*\//, '').replace('[1m]', '');
-  const key = Object.keys(rates).find((name) => normalized.startsWith(name));
-  return rates[key] || rates['claude-sonnet-5'];
+function longestPrefix(name, table) {
+  const keys = Object.keys(table).filter((key) => name.startsWith(key));
+  return keys.sort((a, b) => b.length - a.length)[0];
+}
+
+function rateFor(model, contextTokens = 0) {
+  const normalized = String(model || '').replace(/^.*\//, '').replace('[1m]', '').trim();
+  const entry = models[longestPrefix(normalized, models)] || models['claude-sonnet-5'];
+  return entry.above && contextTokens > entry.above.threshold ? entry.above.rates : entry.rates;
+}
+
+// Codex's session_meta line embeds the full base instructions (~20 KB), so a fixed small
+// read truncates it mid-JSON. Read until the first newline, bounded.
+async function readFirstLine(path, limit = 1024 * 1024) {
+  const handle = await open(path, 'r');
+  try {
+    const chunks = [];
+    const buffer = Buffer.alloc(16 * 1024);
+    let position = 0;
+    while (position < limit) {
+      const { bytesRead } = await handle.read(buffer, 0, buffer.length, position);
+      if (!bytesRead) break;
+      const chunk = buffer.subarray(0, bytesRead);
+      const newline = chunk.indexOf(10);
+      if (newline >= 0) { chunks.push(Buffer.from(chunk.subarray(0, newline))); break; }
+      chunks.push(Buffer.from(chunk));
+      position += bytesRead;
+    }
+    return Buffer.concat(chunks).toString('utf8');
+  } finally { await handle.close(); }
 }
 
 function emptyTotals() {
@@ -56,17 +82,12 @@ export class ClaudeMeter {
       for (const entry of entries) {
         if (!entry.endsWith('.jsonl')) continue;
         const path = join(this.root, project.name, entry);
-        let handle;
         try {
-          handle = await open(path, 'r');
-          const buffer = Buffer.alloc(4096);
-          const { bytesRead } = await handle.read(buffer, 0, buffer.length, 0);
-          const first = buffer.subarray(0, bytesRead).toString('utf8').split('\n', 1)[0];
-          if (JSON.parse(first).sessionId === this.sessionId) {
+          if (JSON.parse(await readFirstLine(path)).sessionId === this.sessionId) {
             this.transcriptPath = path;
             return path;
           }
-        } catch { /* next */ } finally { await handle?.close(); }
+        } catch { /* next */ }
       }
     }
     return null;
@@ -120,7 +141,7 @@ export class ClaudeMeter {
         const input = Number(usage.input_tokens) || 0;
         const output = Number(usage.output_tokens) || 0;
         const model = message.model || this.model;
-        const prices = rateFor(model);
+        const prices = rateFor(model, read + write + input);
         const cost = input * prices[0] + read * prices[1] + write5 * prices[2] +
           write1 * prices[3] + output * prices[4];
         const call = { timestamp: obj.timestamp || null, recorded_at: Date.now() / 1000,
@@ -177,11 +198,7 @@ export class CodexMeter extends ClaudeMeter {
         if (entry.isDirectory()) found.push(...await visit(path));
         else if (entry.name.endsWith('.jsonl')) {
           try {
-            const handle = await open(path, 'r');
-            const buffer = Buffer.alloc(4096);
-            const { bytesRead } = await handle.read(buffer, 0, buffer.length, 0);
-            await handle.close();
-            const meta = JSON.parse(buffer.subarray(0, bytesRead).toString('utf8').split('\n', 1)[0]).payload;
+            const meta = JSON.parse(await readFirstLine(path)).payload;
             const info = await stat(path);
             const expectedProvider = this.role === 'NATIVE' ? 'openai' : 'pennyrouter';
             if (meta?.cwd && meta.cwd.toLowerCase() === this.cwd.toLowerCase() &&
@@ -228,15 +245,16 @@ export class CodexMeter extends ClaudeMeter {
         this.previous = fingerprint;
         const [input, cached, write, output, reasoning] = delta;
         const fresh = Math.max(0, input - cached - write);
-        const prices = [1.25e-6, .125e-6, 1.25e-6, 1.25e-6, 10e-6];
-        const cost = fresh * prices[0] + cached * prices[1] + write * prices[2] +
-          (output - reasoning) * prices[4] + reasoning * prices[4];
+        const context = Number(payload.info.last_token_usage?.input_tokens) || input;
+        const prices = rateFor(this.model, context);
+        const cost = fresh * prices[0] + cached * prices[1] + write * prices[2] + output * prices[4];
         const call = { timestamp: obj.timestamp || null, recorded_at: Date.now() / 1000,
           model: this.model, cache_read: cached, cache_write: write, write_5m: write,
           write_1h: 0, fresh_input: fresh, output, reasoning_output: reasoning,
-          total: input + output, cost, context_tokens: input, pricing_profile: 'standard',
-          rates_per_million: { input: 1.25, cache_read: .125, cache_write_5m: 1.25,
-            cache_write_1h: 1.25, output: 10 } };
+          total: input + output, cost, context_tokens: context, pricing_profile: 'standard',
+          rates_per_million: Object.fromEntries(
+            ['input', 'cache_read', 'cache_write_5m', 'cache_write_1h', 'output']
+              .map((name, index) => [name, prices[index] * 1_000_000])) };
         this.calls.push(call);
         this.totals.cache_read += cached; this.totals.cache_write += write;
         this.totals.write_5m += write; this.totals.fresh_input += fresh;