repospend 0.1.6 → 0.1.8
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- package/CHANGELOG.md +24 -0
- package/README.md +11 -1
- package/dist/cli.js +507 -118
- package/docs/llms.txt +6 -1
- package/docs/pricing.md +40 -6
- package/docs/token-accounting.md +19 -2
- package/package.json +25 -7
- package/web-dist/assets/{index-B2mGGjf_.js → index-BWVpHJSG.js} +87 -87
- package/web-dist/assets/index-C3XBInjD.css +1 -0
- package/web-dist/index.html +2 -2
- package/web-dist/assets/index-Bzdfbi5w.css +0 -1
package/docs/llms.txt
CHANGED
|
@@ -55,9 +55,14 @@ Positioning:
|
|
|
55
55
|
- RepoSpend complements CLI tools such as ccusage.
|
|
56
56
|
- ccusage is useful for terminal-first Claude Code totals and daily breakdowns.
|
|
57
57
|
- RepoSpend is useful for visual repo-level analytics, session inspection, source/tool comparison, model mix, token shape, cache reuse, exports, and command/agent friction.
|
|
58
|
-
- RepoSpend totals can intentionally differ from ccusage or Tokscale because it includes Claude Desktop/local-agent session roots when present, treats cache reads/writes as input sub-buckets, prices
|
|
58
|
+
- RepoSpend totals can intentionally differ from ccusage or Tokscale because it includes Claude Desktop/local-agent session roots when present, treats cache reads/writes as input sub-buckets, prices cache writes by their recorded TTL (Claude 5-minute/1-hour, OpenAI 30-minute/12-hour), prices each Codex request at the model active for it, and separates Codex visible output from reasoning output.
|
|
59
59
|
|
|
60
60
|
Cost language:
|
|
61
|
+
- October 27, 2026 pricing update: explicit GPT-6.1 Terra card ($1/$0.05/$5 input/cache-read/output per million tokens), OpenAI 30-minute and 12-hour cache-write rates, and per-request long-context tiers above 272K input tokens for GPT-6/6.1 and GPT-5.x Standard models (5.6 Sol/Terra/Luna, 5.5/5.4 plus Pro cards).
|
|
62
|
+
- October 17, 2026 pricing update: Sonnet 5.5 cache reads are $0.10 per million tokens; Sonnet 5 remains at $0.20.
|
|
63
|
+
- Haiku 5.5 has two per-request prompt-length tiers: input/cache read/output $0.10/$0.01/$0.50 up to 100K, and $0.50/$0.05/$2.50 above 100K. Cache writes have separate 5-minute and 1-hour rates. Unknown request lengths stay unpriced.
|
|
64
|
+
- Gemini 4 Argon introductory input/cache read/output rates are $2/$0.10/$10 through November 30, 2026, then published rates become $4/$0.20/$20. Bundled tables are a dated snapshot, not an automatic future rate switch.
|
|
65
|
+
- Current API-equivalent pricing applied to older usage is not historical billing reconstruction.
|
|
61
66
|
- GPT-6.1 Sol has an explicit Standard short-context card: $2 input, $0.10 cached input, $2.50 cache writes, $10 output per million tokens, checked September 29, 2026.
|
|
62
67
|
- RepoSpend shows estimated API-equivalent cost.
|
|
63
68
|
- RepoSpend does not claim to show actual bills, invoices, subscription usage, savings, credits, or account-specific charges.
|
package/docs/pricing.md
CHANGED
|
@@ -4,21 +4,46 @@ RepoSpend estimates API-equivalent cost from a local pricing table. The bundled
|
|
|
4
4
|
|
|
5
5
|
The bundled table is seeded from public OpenAI, Anthropic, Google, and GitHub Copilot model references and is expressed as USD per 1M tokens. Pricing changes over time, so treat RepoSpend costs as API-equivalent estimates rather than invoice-grade accounting.
|
|
6
6
|
|
|
7
|
-
The GPT-6.1 Sol card was checked on September 29
|
|
7
|
+
The October 27, 2026 update adds the GPT-6.1 Terra card and OpenAI's 30-minute and 12-hour cache-write rates, and checks Claude Haiku 5.5, Sonnet 5.5, Fable 5.1, Mythos 5.1, and Gemini 4 Argon against provider documentation. The GPT-6.1 Sol card was checked on September 29; the other bundled defaults retain the September 28 snapshot. Defaults use published Standard API prices. Sonnet 5.5 cache reads fell from $0.20 to $0.10 per 1M tokens on October 7; its $2 input, $10 output, $2.50 5-minute writes, and $4 1-hour writes are unchanged. Sonnet 5 retains its $0.20 cache-read rate.
|
|
8
8
|
|
|
9
|
-
|
|
9
|
+
### Claude Haiku 5.5 and request-length pricing
|
|
10
|
+
|
|
11
|
+
All rates are USD per 1M tokens, using Anthropic's global Standard API prices:
|
|
12
|
+
|
|
13
|
+
| Prompt length per request | Input | Cache reads | 5m writes | 1h writes | Output |
|
|
14
|
+
| --- | ---: | ---: | ---: | ---: | ---: |
|
|
15
|
+
| Up to and including 100,000 tokens | $0.10 | $0.01 | $0.125 | $0.20 | $0.50 |
|
|
16
|
+
| Over 100,000 tokens | $0.50 | $0.05 | $0.625 | $1 | $2.50 |
|
|
17
|
+
|
|
18
|
+
Prompt length includes uncached input, cache reads, and cache writes. The higher rates apply to the whole qualifying request, not just tokens above 100K. Multiple short requests keep the lower rates even when combined session input exceeds 100K. Claude's deduplicated usage events and Copilot chat/inference records are priced individually, then their unrounded costs are summed before rounding the session estimate once.
|
|
19
|
+
|
|
20
|
+
Settings exposes `claude-haiku-5-5` for the lower tier and `claude-haiku-5-5-long-context` for the higher tier. The latter is a RepoSpend rate card, not a provider model ID; both can be edited separately. The higher-tier card appears with used Haiku models. Editing an inherited Haiku snapshot updates the shared base card and retains automatic tier selection. Use Add model to intentionally create an exact custom snapshot with flat rates; such overrides remain authoritative. Haiku usage without a known request length (including aggregate-only Copilot, Codex, or experimental Cursor totals) stays unpriced unless an explicit custom flat snapshot card or recorded Cursor cost is available. Pricing triage distinguishes missing request lengths from missing rates and exposes any unpriced auxiliary request models. Unknown fast-mode cards remain unpriced.
|
|
21
|
+
|
|
22
|
+
These are current API-equivalent rates applied to the selected usage window, not historical date-effective billing. The Sonnet reduction does not claim that pre-October 7 API requests were billed at the new price. Sources: [Anthropic pricing](https://platform.claude.com/docs/en/about-claude/pricing), [October 7 announcement](https://www.anthropic.com/claude-haiku-5-5).
|
|
23
|
+
|
|
24
|
+
### Gemini 4 Argon introductory rates
|
|
25
|
+
|
|
26
|
+
`gemini-4-argon`, including Copilot-prefixed/suffixed aliases, uses $2 input, $0.10 cached input, and $10 output (including thinking) per 1M tokens through November 30, 2026. Google lists $4 input, $0.20 cached input, and $20 output from December 1. Argon has no explicit cache-write charge. Bundled rates are a verified snapshot: the application does not silently switch prices based on the machine clock. Refresh the table at the promotion boundary or save intentional overrides. Source: [Google's model and pricing guide](https://ai.google.dev/gemini-api/docs/latest-model?hl=en).
|
|
27
|
+
|
|
28
|
+
GPT-6 Astra, Sol, and Luna use OpenAI's Standard short-context rates. Their long-context requests above 272,000 input tokens use different rates, applied per request when local records carry request sizes (Codex per-turn and Copilot per-request usage); aggregate-only usage stays at short-context rates and is flagged with an `assumed_short_context_rates` warning instead of guessing a tier. Fast, batch, and regional processing can also differ from these defaults. RepoSpend keeps the result labeled as an API-equivalent estimate.
|
|
29
|
+
|
|
30
|
+
The GPT-5.x Standard cards carry the same per-request 272K threshold: GPT-5.6 Sol ($8 input / $0.80 cache read / $10 30m and $20 12h writes / $30 output), GPT-5.6 Terra ($4 / $0.40 / $5 and $10 / $18), GPT-5.6 Luna ($0.40 / $0.04 / $0.50 and $1 / $1.80), GPT-5.5 ($10 / $1 / $45), GPT-5.5 Pro ($60 / $270, no published cache rates), GPT-5.4 ($5 / $0.50 / $22.50), and GPT-5.4 Pro ($60 / $270, no published cache rates) each expose an editable `-long-context` card. GPT-5.4 mini and nano have no published long-context tier and stay flat at every request size. The same semantics apply as GPT-6/6.1: a request over 272,000 input tokens reprices at 2x input/cache rates and 1.5x output rates, an intentional custom flat base card stays flat unless the long-context card is also edited, and unknown request lengths use short-context rates with an assumption warning.
|
|
10
31
|
|
|
11
32
|
### GPT-6.1 Sol (September 29, 2026)
|
|
12
33
|
|
|
13
34
|
The explicit `gpt-6.1-sol` card uses $2 input, $0.10 cached input, $2.50 cache writes, and $10 output/reasoning per 1M tokens, verified against [OpenAI's model reference](https://developers.openai.com/api/docs/models/gpt-6.1-sol) on September 29. GPT-6 Sol remains separate at $0.20 for cached input. Case-insensitive and Copilot-prefixed IDs resolve to the same card; terminal snapshot dates preserve the model generation. Snapshot support is defensive parsing, not a claim that OpenAI publishes those snapshot IDs. Custom exact-ID overrides continue to take precedence.
|
|
14
35
|
|
|
15
|
-
Above 272K input tokens per request, GPT-6.1 Sol's Standard rates become $4 input, $0.20 cached input, $5 cache writes, and $15 output per 1M tokens. Fast costs twice Standard; Batch and Flex cost half Standard.
|
|
36
|
+
Above 272K input tokens per request, GPT-6.1 Sol's Standard rates become $4 input, $0.20 cached input, $5 cache writes, and $15 output per 1M tokens, selected per request through the `gpt-6.1-sol-long-context` card (a RepoSpend rate card, not a provider model ID). Fast costs twice Standard; Batch and Flex cost half Standard. Requests without a recorded size use short-context rates with an assumption note; sessions whose saved table lacks the long-context card fall back to base rates and are flagged with `missing_long_context_pricing`. See the [GPT-6.1 Sol pricing and Codex tracking guide](https://repospend.com/guides/gpt-6-1-sol-pricing) for a worked example.
|
|
37
|
+
|
|
38
|
+
### GPT-6.1 Terra (October 26, 2026)
|
|
39
|
+
|
|
40
|
+
The explicit `gpt-6.1-terra` card uses $1 input, $0.05 cached input, $1.25 30-minute cache writes, $2.50 12-hour cache writes, and $5 output/reasoning per 1M tokens, verified against [OpenAI's model reference](https://developers.openai.com/api/docs/models/gpt-6.1-terra) on October 27. Above 272K input tokens per request the whole request reprices at 2x input/cache rates and 1.5x output rates through the editable `gpt-6.1-terra-long-context` card. Terra does not inherit the older GPT-5.6 Terra rates; custom exact-ID overrides continue to take precedence.
|
|
16
41
|
|
|
17
42
|
Claude Fable 5.1 and Mythos 5.1 retain their predecessor's base rates but have lower cache-read pricing. Claude Opus 5.5 has its own lower base and cache rates. These model rows use Anthropic's global Standard rates.
|
|
18
43
|
|
|
19
44
|
OpenAI reduced GPT-5.6 pricing effective July 30, 2026. RepoSpend uses $2 input and $12 output per 1M tokens for Terra, and $0.20 input and $1.20 output for Luna. Cached input remains 90% below the uncached input rate, and cache writes are billed at 1.25x the uncached input rate, so Luna's bundled cache-write rate is $0.25 per 1M tokens.
|
|
20
45
|
|
|
21
|
-
GPT-5.6 Sol and its `gpt-5.6` alias use promotional $4 input, $0.40 cached input, $5 cache writes, and $20 output/reasoning per 1M tokens, available at least through November 21, 2026. Older full bundled pricing files receive this correction; intentionally saved sparse overrides remain unchanged.
|
|
46
|
+
GPT-5.6 Sol and its `gpt-5.6` alias use promotional $4 input, $0.40 cached input, $5 cache writes, and $20 output/reasoning per 1M tokens, available at least through November 21, 2026 — recheck the promotional card on November 22, 2026. Older full bundled pricing files receive this correction; intentionally saved sparse overrides remain unchanged.
|
|
22
47
|
|
|
23
48
|
Gemini 3.6, 3.7, and 3.8 Flash use GitHub Copilot promotional rates through December 31, 2026. MAI-Code-1.1-Flash, Grok 4.5–4.7, and Kimi K3 have explicit Copilot reference cards. Grok rates use the up-to-200K input tier; longer requests cost more.
|
|
24
49
|
|
|
@@ -36,11 +61,20 @@ cache-write rate. If a source only exposes the older aggregate
|
|
|
36
61
|
`cache_creation_input_tokens` field, RepoSpend falls back to the model's generic
|
|
37
62
|
cache-write rate.
|
|
38
63
|
|
|
64
|
+
For OpenAI models, RepoSpend uses the TTL split recorded in
|
|
65
|
+
`usage.cache_write_tokens_details` (or the nested
|
|
66
|
+
`input_tokens_details.cache_write_tokens_details` shape): `ttl_30m_tokens` use
|
|
67
|
+
the 30-minute cache-write rate and `ttl_12h_tokens` use the 12-hour rate. Writes
|
|
68
|
+
without a TTL breakdown use the generic cache-write rate, which equals the
|
|
69
|
+
30-minute rate on bundled OpenAI cards.
|
|
70
|
+
|
|
39
71
|
Each model can define:
|
|
40
72
|
|
|
41
73
|
- input tokens
|
|
42
|
-
- 5-minute cache write input tokens
|
|
43
|
-
- 1-hour cache write input tokens
|
|
74
|
+
- 5-minute cache write input tokens (Claude)
|
|
75
|
+
- 1-hour cache write input tokens (Claude)
|
|
76
|
+
- 30-minute cache write input tokens (OpenAI)
|
|
77
|
+
- 12-hour cache write input tokens (OpenAI)
|
|
44
78
|
- generic cache write input tokens for sources without a TTL split
|
|
45
79
|
- cached input tokens
|
|
46
80
|
- output tokens
|
package/docs/token-accounting.md
CHANGED
|
@@ -6,6 +6,7 @@ RepoSpend uses one normalized token shape across local clients:
|
|
|
6
6
|
- `cachedInputTokens` is the cache-read portion of input.
|
|
7
7
|
- `cacheCreationInputTokens` is the cache-write portion of input, when the source exposes it.
|
|
8
8
|
- `cacheCreationInputTokens5m` and `cacheCreationInputTokens1h` preserve Anthropic's Claude cache-write TTL split when local transcripts expose `usage.cache_creation`.
|
|
9
|
+
- `cacheCreationInputTokens30m` and `cacheCreationInputTokens12h` preserve OpenAI's cache-write TTL split when usage records expose `cache_write_tokens_details` (flat or nested under `input_tokens_details`); `cacheCreationInputTokens` remains the total write bucket for billing.
|
|
9
10
|
- `outputTokens` is visible/non-reasoning output.
|
|
10
11
|
- `reasoningTokens` is reasoning output, when the source exposes it.
|
|
11
12
|
- `totalTokens = inputTokens + outputTokens + reasoningTokens`.
|
|
@@ -55,15 +56,24 @@ session files exist, RepoSpend can show higher Claude totals than `ccusage` or
|
|
|
55
56
|
Tokscale without a parser bug. When comparing tools, first check whether the
|
|
56
57
|
same Claude roots are included.
|
|
57
58
|
|
|
58
|
-
|
|
59
|
+
RepoSpend also preserves the cache-write TTL split when it is present
|
|
59
60
|
in local logs:
|
|
60
61
|
|
|
61
62
|
```text
|
|
62
63
|
5-minute cache write = cacheCreationInputTokens5m
|
|
63
64
|
1-hour cache write = cacheCreationInputTokens1h
|
|
64
|
-
|
|
65
|
+
30-minute cache write = cacheCreationInputTokens30m
|
|
66
|
+
12-hour cache write = cacheCreationInputTokens12h
|
|
67
|
+
unclassified cache write = cacheCreationInputTokens - (5m + 1h + 30m + 12h buckets)
|
|
65
68
|
```
|
|
66
69
|
|
|
70
|
+
OpenAI cache reads and writes may arrive as flat fields (`cached_input_tokens`,
|
|
71
|
+
`cache_write_tokens`) or nested (`input_tokens_details.cached_tokens`,
|
|
72
|
+
`input_tokens_details.cache_write_tokens` with
|
|
73
|
+
`cache_write_tokens_details.ttl_30m_tokens` / `ttl_12h_tokens`). Both shapes
|
|
74
|
+
normalize to the same buckets, and `inputTokens` stays inclusive of reads and
|
|
75
|
+
writes per the OpenAI API contract.
|
|
76
|
+
|
|
67
77
|
This is one reason RepoSpend API-equivalent cost can differ from `ccusage`.
|
|
68
78
|
Anthropic prices 5-minute cache writes at a different rate than 1-hour cache
|
|
69
79
|
writes. RepoSpend applies the recorded TTL-specific rate for each bucket; tools
|
|
@@ -93,6 +103,13 @@ When checking RepoSpend against `ccusage`, Tokscale, or another local usage tool
|
|
|
93
103
|
|
|
94
104
|
For Codex, RepoSpend prefers per-turn `last_token_usage` deltas and skips stale duplicate snapshots when available. This avoids under-counting sessions that compacted and avoids over-counting rebroadcast cumulative checkpoints.
|
|
95
105
|
|
|
106
|
+
Codex sessions can switch models mid-session. Per-turn and direct usage records
|
|
107
|
+
each keep the model that was active for that request in the session's
|
|
108
|
+
`modelUsage` slices, so per-request pricing and model rollups stay correct
|
|
109
|
+
without creating duplicate session rows. Sessions that only expose cumulative
|
|
110
|
+
snapshots fall back to the session model and are priced as an aggregate with an
|
|
111
|
+
explicit assumption warning rather than invented per-request sizes.
|
|
112
|
+
|
|
96
113
|
For GitHub Copilot, RepoSpend aligns with ccusage and Tokscale for OTEL-backed
|
|
97
114
|
records, but keeps `cachedInputTokens` as a subset of `inputTokens` in the
|
|
98
115
|
normalized model. Copilot CLI session-state files can expose output tokens
|
package/package.json
CHANGED
|
@@ -1,6 +1,6 @@
|
|
|
1
1
|
{
|
|
2
2
|
"name": "repospend",
|
|
3
|
-
"version": "0.1.
|
|
3
|
+
"version": "0.1.8",
|
|
4
4
|
"description": "Local-first dashboard for tracking AI coding token usage and API-equivalent spend by repository, session, model, and tool.",
|
|
5
5
|
"license": "Apache-2.0",
|
|
6
6
|
"author": "Mehmet Mustafa Demir",
|
|
@@ -70,24 +70,42 @@
|
|
|
70
70
|
"sqlite"
|
|
71
71
|
],
|
|
72
72
|
"engines": {
|
|
73
|
-
"node": ">=
|
|
73
|
+
"node": ">=22.12"
|
|
74
74
|
},
|
|
75
75
|
"packageManager": "pnpm@9.15.9",
|
|
76
76
|
"dependencies": {
|
|
77
|
-
"@fastify/static": "^
|
|
77
|
+
"@fastify/static": "^10.1.8",
|
|
78
78
|
"better-sqlite3": "^12.10.0",
|
|
79
|
-
"fastify": "^5.
|
|
79
|
+
"fastify": "^5.12.7"
|
|
80
80
|
},
|
|
81
81
|
"devDependencies": {
|
|
82
82
|
"@eslint/js": "^9.39.1",
|
|
83
83
|
"@types/node": "^22.19.1",
|
|
84
|
-
"esbuild": "^0.28.
|
|
84
|
+
"esbuild": "^0.28.2",
|
|
85
85
|
"eslint": "^9.39.1",
|
|
86
86
|
"gifenc": "^1.0.3",
|
|
87
87
|
"playwright": "^1.60.0",
|
|
88
|
-
"sharp": "^0.35.
|
|
88
|
+
"sharp": "^0.35.5",
|
|
89
89
|
"typescript": "^5.9.3",
|
|
90
90
|
"typescript-eslint": "^8.46.4",
|
|
91
|
-
"vitest": "^3.2.
|
|
91
|
+
"vitest": "^3.2.7"
|
|
92
|
+
},
|
|
93
|
+
"pnpm": {
|
|
94
|
+
"overrides": {
|
|
95
|
+
"vite": "^6.4.5",
|
|
96
|
+
"find-my-way": "^9.10.0",
|
|
97
|
+
"fast-uri@3": "^3.1.9",
|
|
98
|
+
"tinypool": "^2.1.2",
|
|
99
|
+
"brace-expansion@1": "^1.1.21",
|
|
100
|
+
"brace-expansion@>=2": "^5.0.12",
|
|
101
|
+
"js-yaml": "^4.3.2",
|
|
102
|
+
"postcss": "^8.5.23",
|
|
103
|
+
"postcss-selector-parser": "^6.1.5",
|
|
104
|
+
"nanoid@3": "^3.3.18",
|
|
105
|
+
"source-map-js": "^1.2.2",
|
|
106
|
+
"browserslist": "^4.28.7",
|
|
107
|
+
"baseline-browser-mapping": "^2.11.0",
|
|
108
|
+
"@babel/core": "^7.29.6"
|
|
109
|
+
}
|
|
92
110
|
}
|
|
93
111
|
}
|