Obtaining model API prices and estimating token costs

The cost engine and optional models.dev refresh are implemented and default off. See online refresh, cost-engine results, refresh results and current API references. Occurrence-time estimates on real data still require verified/configured provider channels; reference-price comparison does not verify actual bills. Spending reminders are implemented/default off, using user daily/monthly occurrence-time estimates per currency without zero-filling unknowns. Specified Coding Plans do not verify actual pay-as-you-go payments; other scope remains in Plan.md. Research completed 2026-09-25. Data pricing rules are authoritative; this document adds channel references/local snapshots. Prices below describe dated research snapshots and test values, rather than a promise of current rates.

Research scope and method

Prioritize locally observed providers: OpenAI Codex gpt-6-astra/gpt-6-sol; Zhipu GLM omp/pi/zcode glm-5.3-flash/GLM-5.3 including synthetic glm-5.2; Moonshot kimi-code/kimi-for-coding/k3. Anthropic/Google are specified integration targets without local data at the research date. Compare official machine-readable sources/official pages/community catalogs and manual snapshots. External lookup dates: 2026-09-25 unless a later dated entry says otherwise; source list below. Absent verifiable information stays unverified. Page bodies/public curl JSON establish facts; search summaries locate sources only. Most official pages lack versions/dates; snapshot identity uses URL/retrieval date/visible updates, e.g. Gemini Last updated. Community catalogs use data hash characteristics/repository push times.

Price-source comparison

Official machine-readable interfaces

The checked OpenAI/Anthropic/Gemini model-list APIs require authentication and contain no price fields. Moonshot’s Mintlify platform supplies official .md/llms.txt from the same source as HTML. No anonymous pricing JSON/YAML API was found for the researched providers; absence of a verified endpoint does not prove one cannot exist.

Provider	Model-list API	Price fields	Public machine-readable pricing	References
OpenAI	GET /v1/models, Authorization: Bearer	No; id/created/object/owned_by/shutdown_date only	Not found; HTML docs, platform.openai.com redirects 301 to developers.openai.com	S01/S02
Anthropic	GET /v1/models, X-Api-Key	No; id/display_name/created_at/max_tokens, etc.	Not found; HTML docs	S03/S04
Google Gemini	GET /v1beta/models?key=, API key	No; Model has token limits, etc.	publishPriceDocs not found (S23); HTML with Last updated date	S07/S08
Moonshot Kimi	Public model-list endpoint unverified	—	Anonymous docs/pricing/chat.md/docs/llms.txt via Mintlify; CN moonshot.cn/kimi.com and global moonshot.ai/kimi.ai separate	S09/S10
Zhipu GLM	bigmodel public model-list endpoint unverified	—	Not found; bigmodel pricing is rendered JS SPA; international docs.z.ai static docs	S15/S17

OpenAI/Anthropic/z.ai/bigmodel pages lack date/version markers, so page age cannot be established from them alone. Gemini dates updates but retrieval language can yield machine translations; one fetched page was Polish with original USD figures (S07). No update frequency promised. OpenAI mentions tier rename 2026-07-30/promotion end 2026-11-21, showing irregular revisions (S01).

Community catalogs

Catalog	Data / license	Structure	Official comparison on 2026-09-25	References
models.dev	MIT, anomalyco/models.dev moved from sst; pushed 2026-09-25; api.json about 4.9 MiB/200+ providers	Provider groups; cost.input/output/cache_read/cache_write USD/million tokens; context tiers; reasoning/tools; many coding-plan entries	gpt-6-astra 10/50/1/12.5 and 272K tier 20/75 match; zai glm-5.3 1.4/4.4/0.26 matches docs.z.ai. zhipuai bigmodel endpoint links z.ai/USD without CN CNY. Coding-plan cost 0 denotes subscription placeholders	S18/S19
OpenRouter	Anonymous GET /api/v1/models 200,460 models; USD/token decimal strings; license unverified	pricing.prompt/completion/input_cache_read/input_cache_write; overrides min_prompt_tokens long tiers; separate :batch SKU	gpt-6-astra/moonshotai/kimi-k3/z-ai/glm-5.3 match; Flash 0.045/0.14 below official 0.15/0.50; k2.7-code 0.6562/3.3 below 0.95/4.0. Aggregator rates are not official rates	S20
pydantic/genai-prices	MIT; pushed 2026-09-25; provider YAML plus prices/new_data/v2/data.json/JSON Schema	prices_checked/pricing_urls; input/cache_read/output USD/million; historical/tiered/daily rates; provider conversion/selection notes	kimi-k3 3/0.3/15 and zai GLM-5.3 1.4/0.26/4.4 match. zhipuai explicitly converts CNY at 1 USD = 7.25 CNY/selects lowest flagship tier; an explicit CN approximation, not official data	S21
libgen pricing	No LLM catalog with that name found after multiple searches; Library Genesis/unrelated results only	—	—	S22

Catalogs do not promise freshness. Cache-write TTL5m/1h, batch and subscription/API distinctions vary. Aggregator rates are a different price category. Use catalogs for cross-checks/candidates, without replacing official references.

Manually maintained local snapshots

See versioned snapshots. Repository seeds are versioned JSON with source URLs/ lookup dates/effective intervals/extraction methods per row. UI imports local JSON to override/ supplement, validated by schema/interval conflicts. Work concentrates around flagship price changes/new releases; the small locally used model set makes manual checking manageable.

Source selection

Prefer manually verified versioned snapshots: official machine-readable availability is limited, while community catalogs have licensing/structural benefits and observed discrepancies. models.dev/ genai-prices cross-check official pages and supply candidates when pages are unavailable; adopted rows retain community provenance. OpenRouter is comparison-only because aggregator rates differ. Optional public-price metadata downloads default off. The implemented built-in refresh source is models.dev api.json only, MIT with data-driven official-provider selection, raw caching/long TTL/failure fallback; import official pay-as-you-go entries only, excluding plan placeholders. Manual imports/repository seeds remain available; see online-refresh design.

Verified billing dimensions

The five researched providers use pages for observed/target models. Prices are per million tokens; unverified means the page did not establish that dimension, without guessed values.

Dimension matrix

Dimension	OpenAI	Anthropic	Gemini	Kimi	GLM
Ordinary input	Yes	Yes	Yes	Cache-miss price	Yes
Cache read	One tenth input	0.1x, some0.025x/0.05x	Context-cache price	One tenth miss price	Cache-hit price
Cache write5m	1.25x input	1.25x input	No separate write price	Miss-input price	No separate write price
Cache write1h	Unverified/not found	2x input	—	2x miss-input	—
Cache storage	—	—	Million tokens/hour	Unverified	Temporarily free by policy
Output	Yes	Yes	Yes	Yes	Yes
Reasoning output	No separate rate; input/output rates	Thinking as output	Included in output	reasoning_content in input/output quotas	Official rule unspecified below
Batch	Batch/Flex 50%	50%, combines cache discount	Batch/Flex 50%	BatchJob 60%	bigmodel50%; z.ai unverified
Long context	Two tiers; GPT-6 threshold absent on page, catalogs272K	4.6+ standard at 1M, no tiers	Pro >200K	No tiers found	GLM-5.1 32K; none found5.2/5.3
Currency	USD	USD	USD	CN CNY/global USD excluding tax	bigmodel CNY/z.ai USD
Region/channel	FedRAMP/data residency+10%	US-only inference1.1x for4.6+	Not found/unverified	Two platforms/two rates	Two channels/two rates

Provider details and observed-model prices

OpenAI S01, undated page, locally observed Codex gpt-6-astra/gpt-6-sol: short astra input $10, cache read $1, write $12.50, output $50; sol $2/$0.20/$2.50/$10. Long astra$20/$2/$25/$75. The page does not show GPT-6’s tier threshold although older families show272K; models.dev/ OpenRouter use272,000 (S18/S20), requiring official recheck before formal adoption. Fast, renamed from Priority2026-07-30, costs2x; FedRAMP/residency+10% for models released after 2026-03-05. No separate reasoning rate; tokens use the chosen model’s input/output rates.

Anthropic S03, research target with no local Claude data then: Opus 5.5 input $4/output $20, write $5/$8, read $0.20 (0.05x); Sonnet 5$2/$10, Haiku 4.5$1/$5. Sonnet 5 promotional$2/$10 became permanent instead of the planned2026-09-01 increase. Batch 50% combines cache discounts (S06). Thinking is output; usage.output_tokens_details.thinking_tokens breaks out billed internal reasoning (S05). Retained historical thinking blocks count as input for Opus 4.5/4.6+.

Gemini S07, Last updated 2026-09-24 UTC:3.8 Flash standard input $0.75 through 2026-12-31, $1.50 from 2027-01-01; output $3.75 including thinking; cache read $0.075/storage$0.50 per million tokens/hour. Batch/Flex 50%;2.5 Pro >200K input $1.25/$2.50/output $10/$15. English/machine-translated3.1Pro preview fetches disagree on >200K presentation; recheck before adoption, as noted in S07.

Kimi S09–S13, observed Kimi Code/Work kimi-code/kimi-for-coding, wire inputOther/inputCacheRead/ inputCacheCreation align with dimensions. kimi-k3 CN hit¥2/miss¥20/output¥100; global $0.30/$3/$15, excluding tax. kimi-k2.7-code CN¥1.30/¥6.50/¥27, global $0.19/$0.95/$4. Separate cache-write policy S12: global k3 five-minute $3=miss input, one-hour $6=2x; hits renew expiry. CN separate rates not found/unverified. BatchJob 60%, not50%(S13); reasoning_content counts toward input/output quotas(S11).

GLM S15/S17, observed omp/pi/zcode glm-5.3-flash/GLM-5.3: bigmodelCN GLM-5.3 input¥8/cache¥2/ output¥28 at 1M context. Flash temporary50% rates¥0.4/¥0.115/¥1.4 versus standard¥0.8/¥0.23/¥2.8. GLM-5.2 matches5.3;5.1 tiers[0,32K)/[32K+); Batch 50%; cache storage temporarily free. International z.ai GLM-5.3$1.4/$0.26/$4.4, Flash $0.15/$0.03/$0.50, FlashX $0.37/$0.075/$1.25. Batch/tier statements not found/unverified. Official thinking docs only mention additional tokens; examples combine reasoning_content/content in completion_tokens without reasoning_tokens(S16). Billing remains unverified; snapshots omit a separate reasoning price, and known output subsets must not be billed again.

Observed models and price mapping

Observed model	Source	Pay-as-you-go availability	Treatment
gpt-6-astra/gpt-6-sol	Codex ChatGPT subscription	S01	API reference estimate applied to subscription usage
GLM-5.3/glm-5.3-flash	omp/pi/zcode	S15/S17, two channels/currencies	Exact channel first; unknown channels permit unambiguous official same-model reference only, without actual-payment inference
kimi-code/k3	Kimi Code/Work/omp	k3 S09/S10	Coding Plan usage gets API reference estimate
kimi-for-coding	Kimi Code/Work	Official mapping since 2026-09-11 to K2.8 Preview; no verified original-vendor API price, S14	Authorized K2.8 exception: current K2.7 Code official reference with substitute label; occurrence-time estimate still lacks price
k28-agent-preview	Kimi K2.8 Preview	User confirmed identity/substitution2026-10-05	Identity kimi-k2.8-preview; only missing exact prices use kimi-k2.7-code, not arbitrary unknown K2/K3
Synthetic glm-5.2	Kilo test samples	S15/S17	Same two-channel handling as GLM-5.3

Coding Plans/API usage are different billing systems. models.dev kimi-code-plan-*/zai-coding-plan cost 0 is a subscription placeholder, not free API pricing; never import placeholder zeros.

Versioned price-snapshot requirements

Snapshot format

A metadata header and row array. Each billing context is provider×model×region/channel×service tier×context tier×effective interval. Implemented2026-09-30/schema 8: price_snapshots metadata, price_versions mandatory region/channel, service_tier/context_threshold_tokens/five *_per_mtok_hundredths price columns/hourly cache_storage, matching these decisions:

Dimension	Research schema	Implemented decision
Cache-write TTL	One cache_write	cache_write_5m/cache_write_1h; absent tier NULL
Service tiers	Absent	service_tier standard default; separate tier rows; auto-estimation standard only
Context tiers	Absent	context_threshold_tokens minimum input, NULL=0; per-request input determines tier; unknown input with multiple tiers stays ambiguous
Cache storage	Absent	cache_storage_hour; temporary free0 with policy metadata; engine omits without storage duration
Precision	Integer minor currency units/million	Hundredths of minor units/million, i64; $0.075/M=750, supports ¥0.115/M
Provenance	price_version string only	Snapshot ID/source type/URLs/fetched_at/content_hash/license/checker/method
Reasoning price	Absent	No separately verified rate; GLM remains unverified; no second charge for output subsets

Versioned seeds desktop/src-tauri/crates/core/prices/seed-2026-09-25.json and seed-2026-10-02.json import once on startup. Later seed adds Opus 4.8/GPT-4o mini2024-07-18 standard rates effective from verification date, without backdating. Settings costs imports use the same JSON/manual source_type. Monetary values/prices never use binary floats; data rules. Format llm-usage-price-snapshot/1; ISO UTC YYYY-MM-DD dates, half-open effective interval, null end=open; prices hundredths of minor currency units/million; null component=no applicable rate:

{
  "format": "llm-usage-price-snapshot/1",
  "snapshot": {
    "id": "manual-2026-10-01",
    "source_type": "manual",
    "source_urls": ["https://example.com/pricing"],
    "fetched_at": "2026-10-01"
  },
  "rows": [
    {
      "price_id": "my-openai-gpt-6-astra",
      "provider_id": "openai",
      "model": "gpt-6-astra",
      "region": "global",
      "channel": "api",
      "service_tier": "standard",
      "context_threshold_tokens": null,
      "effective_from": "2026-10-01",
      "effective_to": null,
      "currency": "USD",
      "input": 100000, "cache_read": 10000,
      "cache_write_5m": 125000, "cache_write_1h": null,
      "output": 500000, "cache_storage_hour": null,
      "note": null
    }
  ]
}

Validate USD/CNY/service-tier enums/nonnegative prices/legal nonoverlapping same-key provider/ model/region/channel/tier/threshold intervals/unique price_id without imported-row conflicts. Same snapshot ID skips only when price-field hashes/source metadata/notes all match; corrections require a new snapshot ID. Optional official_vendor bool since schema 11 marks original-vendor API prices as fallback candidates: seed/community default true, manual false; manual official rows can explicitly set true. Exact provider_id/region/channel matches selected defaults case-insensitively. Missing exact rows may use official same-model references below without guessing the actual channel. Across snapshots: applicable date/threshold, manual>seed>community, then newer fetched_at/import time within one class. Below-threshold inputs do not use a row; unknown inputs with high-tier rows do not guess tiers.

Effective intervals and updates

Half-open [effective_from_ms,effective_to_ms); same-key provider/model/channel/tier/context rows within a snapshot cannot overlap. Match occurred_at_ms.
Append rows only. New snapshots with later starts override old open intervals without changing imported old rows. Announced future changes, e.g. Gemini2027-01-01/GLM Flash discount expiry, can be registered for automatic date-based switching.
Prefer manual import/official-page or .md seeds with lookup dates. Optional downloads default off per cost/reminder settings. HTTPS GET public price endpoints sends no local usage/host-source identity/session/account keys; no remote usage/billing APIs.
Validate JSON Schema/nonnegative values/currencies/intervals; preview additions/interval closures/price changes per row. Community/official discrepancies warn and permit rejection.
Persist locally; statistics/prior estimates work offline. Failed downloads/validation retain verified snapshots/show fetched_at/verified_at age; no silent amount changes.
Reproduce history through data_revision/price_version. Keep occurrence-time amounts distinct from current-price simulation. Dashboard displays current API pay-as-you-go references rather than duplicate historic estimates. Background updates do not recalculate saved amounts; explicit user requests/data revisions may. Matching sealed days mark limited detail coverage.

Online refresh: models.dev community catalog

Task5 implementation. Settings costs online refresh defaults off. Enabled requests only HTTPS GET models.dev api.json (MIT,S18/S19/S24); no local usage/identities/session bodies/keys; User-Agent contains app/version only. Offline statistics/prior estimates and the no-remote- usage/billing boundary remain unchanged.

Cache and failure fallback

Raw response <DB-directory>/price-cache/models-dev-api.json and .meta.json URL/fetch time/ content hash/bytes; files survive DB rebuild.
Default TTL three days, configurable 1–365. Fresh cache imports without network; explicit refresh-now bypasses TTL.
Download/validation failures reuse the last successful cached response. No cache returns error/preserves all snapshots(A9); invalid responses never replace cache.
Successful download IDs models-dev-YYYYMMDD-hash8 from date/hash; unchanged content skips; snapshot imports never recalculate saved occurrence estimates.
Check after collection; enabled/expired cache downloads in background. Failures produce operation logs/diagnostics without blocking collection.

Catalog filtering: official pay-as-you-go prices

2026-10-01 S24 api.json:225 providers/8339 models; context-only tiers; cost fields input/output/ cache_read/cache_write/input_audio/output_audio/reasoning/tiers/context_over_200k. No currency field: USD/million list prices throughout.

Data-driven official selection: provider IDs prefixed in at least one canonical_model_id, plus corresponding <id>-cn variants. Measured 23+3: openai/anthropic/google/zhipuai/moonshotai/ alibaba/deepseek/xai/mistral/etc., plus moonshotai-cn/minimax-cn/alibaba-cn. Exclude aggregators/ gateways/subscription placeholders.
Skip provider IDs containing -plan, including checked coding-plan/token-plan/step-plan all-zero placeholders. Skip models whose input/output are both zero/missing; zero is not free pricing.
Zero cache/read/write/input/output components import NULL for unavailable rate.
cost.input/output/cache_read map directly; cache_write=>cache_write_5m, one-hour NULL because catalog lacks TTL tiers. context tiers create context_threshold_tokens rows. Ignore duplicate context_over_200k and unsupported audio/reasoning dimensions; reasoning already in output.
region cn for -cn/global otherwise, channel api,currencyUSD; CNY remains seed/manual. Only standard service; experimental.modes fast/ultrafast excluded. effective_from=fetch date without native price-effective intervals; freshness describes lookup time.
USD/million×10^4 converts to hundredths of cents/million, rounded. Small prices such as DeepSeek cache $0.003625/M incur <=0.5 unit/<$0.00005/M rounding, noted in row metadata.
Twenty-six eligible providers yield408 models/473 rows per snapshot. Empty-priced poolside/ sarvam are skipped; persisted rows cover24 providers, as S24/test output distinguishes.

Official-provider fallback

official_vendor schema 11: seed/community true, manual false unless explicitly set. Only when exact provider=>selected region/channel=>model=>standard=>interval=>tier finds no_price_row, search official rows for the same model/standard/date; selected region/channel first, otherwise manual>seed>community/newer within class. Unknown input/multiple tiers remains tier_ambiguous. Mark official_fallback and daily/summary fallback_event_count; UI shows official-reference count while retaining subscription-reference semantics. Missing provider/unconfigured channel enters reference search without premature no_provider/channel_unknown. Vendor-family restriction still requires an exact model: GPT/o OpenAI, Claude Anthropic, Gemini Google, Kimi/Moonshot Moonshot, GLM Zhipu/Z.ai, DeepSeek DeepSeek; references/IDs in dashboard rules. Prefer selected channel, then global official api (including vendor-named compatible channels). Ambiguous remaining regions/channels/currencies stay unpriced; no invented rates/conversion.

Official Kimi model table maps k3/k3-256k toK3. Local kimi-code/ profiles can reference kimi-k3 for price matching only, without changing raw statistics; exact profile prices win. Missing official same-model row yields no_price_row; unknown family plus missing provider yields no_provider. Real provider stays unchanged; reference is not actual payment. Known usage_observation tokens can be partially priced; priced-event counts are not call counts. One verified correction of existing unsealed/detail- retained estimates is permitted; later background prices leave occurrence amounts unchanged. Overview/trend show references/coverage; results.

Queries retain historic responses and current model/currency subtotals/daily provider-model amounts/ matched price_id units. Dashboard model cost column/footer, currency/model curves and expandable units share current references. One row per source-provider/model; show reference provider/snapshot/ context tier. Query the same retained details/sealed daily-weekly-monthly partitions as usage, exclusively by complete source dimension; cleanup neither loses costs nor double-counts coexisting details. Weekly/monthly archives cannot invent daily curve points. Price saved archive components only; different sample-population sums cannot derive missing input components. Keep known totals in coverage denominators. Without per-call tiers use applicable tiers from the same preferred snapshot to bound known components; these intervals do not price missing components. Indivisible archives crossing dynamic-alias changes retain model ambiguity. Usage/prices share one SQLite read transaction. Today/range/selected-hour share references; local-hour filtering includes DST repeated hours and does not substitute whole-day costs. Queries leave saved estimates/snapshots unchanged. Hour-granularity usage may still show daily cost curves without prorating day costs. Candidate reduction retains providers/priorities/dates/tiers; see interactions/results. Never fallback after an exact-chain match, including tier_ambiguous.

Detailed cost rules

These refine authoritative data rules:

Unpriced required components/model without API rate or authorized current substitute/ ambiguous channel-tier/unknown writeTTL remain empty, labeled unpriced; no zero/average rate.
Known priceable components contribute subtotals even with missing others; coverage is priced tokens/known tokens, total marked partial estimate.
Store hundredths of minor currency units/million; sum i128 token×price products before dividing 100,000,000, round each billing component to minor units. Current references group day/provider/model/currency; indivisible archives by period, then combine model/grand totals consistently with tables. Occurrence-time daily cost partitions retain sealed estimates. Never discard tiny event fractions before aggregation. Negative values/cache exceeding known input diagnose; overflow explicitly fails instead of becoming zero/unknown.
Separate currencies. Preserve CNY originals/units, show approximate USD beside them with verified ECB2026-10-02 snapshot/date/source; no live exchange refresh/stored-value changes; dashboard repair. User conversion estimates retain source/time; absent applicable rate displays separate currencies only.
API rates applied to ChatGPT/Coding Plan/GLM subscriptions are reference estimates; actual subscription payments differ. Estimated cache savings are not actual refunds.
Missing writeTTL, e.g. Kimi inputCacheCreation, defaults unpriced. Explicit per-provider default TTL such as5m permits estimated pricing under that tier.

V29 acceptance samples

Validation checklist V29 uses fixed snapshots/manual expected amounts/errors. Values from 2026-09-25 official S01/S09–S13/S15 are fixed tests, not ongoing price guarantees; version them with seeds.

Fixed price samples

Row	Provider/model	Channel/currency	Tier	Threshold	Input	Cache read	Write5m	Write1h	Output
P1	glm-5.3	bigmodel-cn/CNY	standard	—	800	200	NULL	NULL	2800
P2	gpt-6-astra	openai/USD	standard	0	1000	100	1250	NULL	5000
P3	gpt-6-astra	openai/USD	standard	272000	2000	200	2500	NULL	7500
P4	kimi-k3	moonshot-global/USD	standard	—	300	30	300	600	1500
P5	glm-5.1	bigmodel-cn/CNY	standard	0	600	NULL	NULL	NULL	2400
P6	glm-5.1	bigmodel-cn/CNY	standard	32768	800	NULL	NULL	NULL	2800

Table units: minor currency units/million (fen/cents): P1input800=¥8/M, P2input1000=$10/M. Engine storage multiplies table values by100, e.g.P1input80000. NULL cache write means no verified separate rate; temporary-free policies stay metadata, without substituting a zero rate.

Manually expected amounts

Single-event expected components round to fen/cents. Multi-event groups sum integer products before rounding, rather than summing rounded event values. All events occur inside half-open snapshot intervals. Archive/precision regression: 2026-10-05.

Case	Tokens: uncached/read/write/output	Row	Expected
E1 mixed input	1,234,567/500,000/200,000/345,678	P1	988+100+968=2056 fen, ¥20.56; writes unpriced/coverage notice
E2 all components	1,000,000/200,000/100,000/250,000	P3	Input 1.3M≥272K selects P3: manually calculated 2000+40+250+1875=4165 cents/$41.65. The original P2 calculation of 2395 cents does not apply at this input size. Current tests check scaled 130K-input E2′ under P2 (240 cents) and P2/P3 threshold boundaries; results
E3 one-hour TTL	400,000/300,000/600,000(1h)/150,000	P4	120+9+360+225=714 cents/$7.14
E4 upper-tier boundary	Input272,000/0/0/output 8,000	P3	544+60=604 cents
E5 below boundary	Input271,999/0/0/output 8,000	P2	272+40=312 cents
E6 GLM tier	Input32,768/0/0/output 1,000	P6	26+3=29 fen
E7 below GLM tier	Input32,767/0/0/output 1,000	P5	20+2=22 fen
E8 separate currencies	E1+original E2 example	P1+P2	Original example2056CNYfen/2395USDcents separate; no conversion/combination; E2 tier correction remains applicable

Error and boundary scenarios

Case	Input	Expected
A1 no occurrence-time price	kimi-for-coding usage	Empty occurrence estimate; current substitute tested separately
A2 partial pricing	Known input/unknown output	Price input; partial total/correct coverage
A3 no duplicate reasoning	reasoning50,000 subset of output 200,000	Charge output_total once
A4 no wrong service tier	Ordinary realtime call	No batch price
A5 absent cache-read price	Known reads/NULL rate	Component unpriced; others count/coverage notice
A6 unknown TTL	Known writes/no TTL/default	Writes unpriced until explicit default, then estimated
A7 historical reproduction	Within/outside old interval	Occurrence rates; missing history gets separate current-price simulation, never mixed
A8 invalid tokens	Negative/cache exceeds known input	Reject cost calculation/diagnose; no max(0,…)
A9 refresh failure	Network/validation error	Retain snapshots/age/amounts; last successful cache fallback; no-cache error preserves snapshots
A10 immutable snapshot	Same snapshot repeated	Skip unchanged; corrected price uses new price_version

Implementation checklist and dated progress

2026-09-30 results, with later completion updates explicitly noted below:

Implemented schema 7=>8 price_snapshots/redefined price_versions/daily_cost_usage; prerelease backup/rebuild policy.
Implemented seed 2026-09-25/currency-tier/nonnegative/overlap/repeat-import validation. Official-page extraction script optional/unimplemented; seed manually authored.
Implemented matching/components/rounding/unpriced-partial coverage/i128/overflow. UI labels references.
Implemented saved occurrence-time/current simulation/history data_revision/price_basis sets.
Implemented optional models.dev-only/raw-cache defaultthree-day/failure fallback/official filter/fallback schema 11/A9; design/results.
Implemented default-off cost UI/currency groups/snapshot view/manual import/explicit recompute. Reminders were deferred in the original2026-09-30 checklist; the opening update records later default-off implementation, with advisory behavior rather than blocking usage.
Executed V29 fixed P1–P6/E/A cases and manual expectations. Real occurrence estimates still require verified/configured channels; current-reference checks recorded separately.

Risks and limits

Undated official pages measure freshness by retrieval, not effective revision time. Community cross-checks/visible age reduce but cannot eliminate silent price changes.
Region/channel differences matter: K3 CN¥20/global $3 input cannot substitute each other. Verify actual subscription endpoints per Agent; unknown channels may use unambiguous official same-model/currency references without endpoint inference.
Resolve kimi-for-coding by usage date, never apply today’s dynamic route retroactively. K2.8 Preview lacks verified Moonshot API prices; user authorizes current K2.7 Code official substitution with label. Future exact prices win without deleting the substitute rule. estimate_at_time never substitutes models or rewrites historical amounts. Other channels need independent labels/references; spelling/aliases/channel rules.
Observed catalog discrepancies include OpenRouter aggregator rates/models.dev zhipuai with international rates; retain recognizable community provenance.
Limited-time GLM Flash discounts/free cache storage/Kimi separate-write policy may expire; rows express effective changes but detection depends on checking/refresh.
Unverified/inconsistent: GPT-6 official threshold, GLM thinking billing, KimiCN write components, z.ai batch discount, Gemini3.1Pro preview tier presentation. Recheck officially before adoption.
The research batch had no direct OpenAI/Anthropic/Gemini provider usage samples for API-bill comparison; page-level price checks do not establish local end-to-end billing matches.
models.dev remains community data. zhipuai links z.aiUSD; -cn USD entries’ derivation from CNY is unverified. Changed structures fail validation/fallback instead of silent import. Community snapshots containUSD only; CN CNY prices require seed/manual data.
Seeds/engine/online refresh are implemented; V29/validation records establish checked scope.

Research record

Windows 11 x64/Git Bash/curl-jq public JSON/Markdown/WebFetch-web-reader page bodies; dates 2026-09-25 unless stated. Status: research references, not runtime acceptance.

ID	Source	Adopted facts
S01	OpenAI pricing, old platform.openai.com/docs/pricing redirects301; undated	gpt-6-astra/sol/luna/gpt-5.6 rates/short-long/Batch-Flex50%/Fast2x renamed 2026-07-30/FedRAMP-residency+10%/no separate reasoning rate/promotions
S02	OpenAI models API	Bearer; no price fields
S03	Anthropic pricing, no version	Input/output/write5m-1h/read/Sonnet 5 promotion permanent/4.6+ no long tier/US-only1.1x
S04	Anthropic models API	X-Api-Key; no price fields
S05	Extended thinking	Thinking as output/thinking_tokens billed subset/retained thinking as input
S06	Anthropic batch	Standard API50%; cache discounts combine
S07	Gemini pricing, Last updated 2026-09-24 UTC; one translated fetch, USD originals	Flash 3.8 input/output/cache/storage/2027changes; Batch-Flex50%;2.5 Pro >200K; thinking output;3.1 Pro preview inconsistent fetches
S08	Gemini models	Key; no price fields
S09	Kimi CN, canonical platform.kimi.com	k3/k2.7-code/highspeed/k2.6 CNY/1M/billing concepts
S10	Kimi global, canonical platform.kimi.ai; anonymous .md	Same modelsUSD excluding tax; Mintlify .md/llms.txt
S11	Kimi thinking	reasoning_content input/output quota
S12	Kimi caching	k3 write5m$3/1h$6/read $0.30/hit renewal/separation leaves total costs unchanged
S13	Kimi batch	BatchJob 60%
S14	Kimi Coding Plan models	k3/k3-256k/kimi-for-coding K2.8 Preview/highspeed; subscription quotas without API rates
S15	Zhipu pricing, rendered JS/undated	GLM-5.3/Flash/5.2/5.1 CNY/32K/Batch 50%/temporary free storage/hit rates
S16	Zhipu thinking	Additional tokens only; combined completion_tokens/no reasoning breakdown; billing unspecified
S17	Z.AI pricing, undated	GLM-5.3/Flash/FlashXUSD; batch/tier statements absent
S18	models.dev api.json,4.9 MiB/2026-09-25	cost/tiers/reasoning_options/200+providers/plan0/zhipuai z.ai rates/astra official match
S19	anomalyco/models.dev, moved from sst	MIT/pushed 2026-09-25/community README/api.json/OpenCode use
S20	OpenRouter models,anonymous200/460 models/2026-09-25	USD/token/overrides min_prompt_tokens272000/:batch; matches and lower Flash/k2.7-code examples
S21	pydantic/genai-prices,MIT/pushed 2026-09-25	YAML/v2JSON-Schema/prices_checked/history-tiers-daily/zhipuai conversion/zai-moonshot official matches
S22	libgen search2026-09-25	No named LLM-price catalog found; unrelated Library Genesis
S23	publishPriceDocs search2026-09-25	No public Gemini pricing JSON found under that name
S24	models.dev api.json2026-10-01,5.28 MB/HTTP200	Provider{id,env,npm,name,doc,models}; cost{input,output,cache_read,cache_write,tiers,context_over_200k,input_audio,output_audio,reasoning}; canonical_model_id 5274/8339;658 context tiers; no currency,USD/million. Official canonical-prefix/-cn selection26 eligible providers/408 models/473 rows; empty-priced poolside/sarvam excluded, rows24 providers, verified by probe. Plans all0. Official matching astra$10/$50,K3 $3/$15,GLM-5.3$1.4/$4.4,Opus 5.5$4/$20,Flash 3.8$0.75/$3.75; zhipuai/zai same docs.z.ai price
S25	ECB daily rates XML,fetched 2026-10-03/effective 2026-10-02	EUR/USD1.1225,EUR/CNY7.5259; CNY=>USD1.1225/7.5259 for reference display, not bills/trades
S26	Kimi model table,global API,models.dev README,reread2026-10-03	k3/k3-256k mapK3; USD/million input 3/read 0.30/output 15/write5m-1h 3/6. models.dev alreadyUSD/million; CN provider names do not justify reconversion; exact bare-profile aliases leave statistics names unchanged

Still unverified: official GPT-6 threshold/OpenAI1h/Anthropic page version/Gemini3.1Pro tiers/ KimiCN write rates/public model-list endpoint/z.ai batch/GLM thinking billing/bigmodel model endpoint/OpenRouter license. The original research wrote no business code/seeds and made no credentialed requests; research alone did not establish automated pricing. Later implementation is explicitly recorded above.