.agents/skills/add-model-price/references/provider-sources-and-price-keys.md
Always fetch pricing from the provider's official docs before editing.
| Provider | Source |
|---|---|
| Anthropic Claude | https://platform.claude.com/docs/en/about-claude/pricing |
| OpenAI | https://developers.openai.com/api/docs/pricing |
| Google Gemini (AI Studio) | https://ai.google.dev/pricing |
| Google Gemini (Vertex AI) | https://cloud.google.com/vertex-ai/generative-ai/pricing#gemini-models |
| AWS Bedrock | https://aws.amazon.com/bedrock/pricing/ |
| Azure OpenAI | https://azure.microsoft.com/pricing/details/cognitive-services/openai-service/ |
openai.com/api/pricing/ often returns HTTP 403 to automated fetchers.
Use https://developers.openai.com/api/docs/pricing instead as that is often permitted.
Use https://developers.openai.com/api/docs/models/all to discover model-by-model info and pricing.
If this page fails, leave OpenAI prices unchanged and report the 403 as an unresolved finding.(openai\/)?
as an optional prefix in their matchPattern (e.g., (?i)^(openai\/)?(gpt-4o)$).
Entries missing this prefix will not match model IDs sent with the openai/ prefix.
The o4-mini and o4-mini-2025-04-16 entries were found missing this prefix in
June 2026 and corrected. Verify any new OpenAI entries include it.ai.google.dev/pricing) and the Vertex AI
page (cloud.google.com/vertex-ai/generative-ai/pricing) can show different prices
for the same model (e.g. Gemini 2.0 Flash: AI Studio $0.10/MTok vs Vertex $0.15/MTok
as of June 2026). When they differ, prefer the AI Studio page for AI Studio–specific
models and Vertex for Vertex-specific ones; leave the file unchanged and report the
discrepancy when uncertain which applies.gemini-1.5-pro, gemini-1.5-flash, and gemini-1.5-flash-8b
are no longer listed on either official Gemini pricing page as of June 2026. They
appear to be retired/deprecated. Do not add or modify their pricing without a concrete
official source.gemini-2.0-flash-exp, gemini-2.0-pro-exp-02-05, gemini-2.0-flash-thinking-exp-01-21,
gemini-2.5-flash-preview-09-2025, and gemini-2.5-flash-lite-preview-09-2025 are
in the selectable model lists but have no standalone pricing entry on official pages.
Do not add pricing for them without explicit official evidence.gemini-2.0-flash and gemini-2.0-flash-001 are in the
selectable model lists and have pricing entries in the file, but as of June 2026
these models are no longer listed on the official AI Studio pricing page. Treat the
existing prices as the last known values; do not update without a concrete official
source.gemini-3-pro-preview is in the selectable model lists
and the pricing file but is NOT listed on the official AI Studio pricing page as of
June 2026. Its prices ($2.00/≤200k, $4.00/>200k input; $12.00/$18.00 output) were
set when the model was first added; do not update without explicit official evidence.ai.google.dev/pricing has separate Free-tier and Paid-tier columns — do not confuse
them (resolved July 31 2026) — The official Gemini pricing table has both a "Free of
charge" column and a "Paid tier" column per row. A model row can legitimately read
"Context caching price: Not available | $0.025/MTok (paid)" — the "Not available" only
describes the free tier. Prior audits (July 23, 25, 27 2026) saw contradictory
"available" vs "not available" summaries for gemini-3.1-flash-lite context caching
because WebFetch's summarizer sometimes collapsed the two columns into one answer. A
July 31 2026 fetch that explicitly asked to quote the row verbatim confirmed: Free tier
= "Not available", Paid tier = "$0.025/MTok (text/image/video), $0.05/MTok (audio)",
plus a storage price of $1.00 per 1M tokens per hour for the paid tier (a
time-based holding cost with no equivalent usage key in Langfuse's pricing schema —
do not attempt to represent it). Langfuse prices the paid/API tier, so
gemini-3.1-flash-lite's existing cache pricing ($0.025/MTok text/image/video,
$0.05/MTok audio = 2.5e-8 / 5e-8 per token) is CONFIRMED CORRECT; no change was made.
This resolves unresolved finding #5 from the July 27 2026 audit memory. Lesson: when a
provider pricing page has multiple tiers/columns per model, ask WebFetch to quote the
exact row verbatim (not "does caching exist") to avoid column-collapse artifacts.claude-sonnet-5 (no
date suffix; pinned snapshot, not an alias). Introductory pricing of $2/$10 per
input/output MTok is in effect through August 31, 2026; standard pricing of $3/$15 will
apply from September 1, 2026. Cache write 5m = $2.50/MTok, 1h = $4/MTok, read =
$0.20/MTok during introductory period. Since the pricing schema cannot express
time-based tiers, the file holds the current introductory prices; update to $3/$15 and
cache equivalents ($3.75/$6/$0.30) after August 31, 2026. AWS Bedrock ID:
anthropic.claude-sonnet-5. The model is in the flat long-context list (no Large
Context tier). Added to pricing file and anthropicModels in July 2026 audit.developers.openai.com/api/docs/pricing fetch fails for either reason, leave OpenAI
prices unchanged and report it as an unresolved finding.gpt-5.6-sol (flagship, $5/$0.50/$30 per MTok input/cached/output),
gpt-5.6-terra (balanced, $2.50/$0.25/$15), and gpt-5.6-luna (cost-efficient,
$1.00/$0.10/$6.00). All three are reasoning models; no date-stamped snapshot versions were
present at launch. If dated versions appear (e.g. gpt-5.6-sol-2026-07-xx), add them
as separate pricing entries following the gpt-5.4 / gpt-5.5 precedent.
Long context pricing applies when input tokens exceed 272,000: prices are 2× input
and 1.5× output for the full request (cached input also doubles). Individual model page
URLs: https://developers.openai.com/api/docs/models/gpt-5.6-sol (and -terra, -luna).
Long context prices: sol $10/$1.00/$45, terra $5/$0.50/$22.50, luna $2/$0.20/$9
per MTok input/cached/output. Added Large Context (>272K) tiers to the pricing file in
July 2026. The threshold of 272K is unique to this family; most other models use 200K.gemini-3.6-flash appeared on the official AI Studio
pricing page in July 2026 at $1.50/MTok input, $7.50/MTok output, cache read $0.15/MTok
(10% cache-read ratio). No large-context tier. Added to pricing file and selectable model lists
(vertexAIModels, googleAIStudioModels) in the July 22 2026 audit. Note: despite the higher
version number, output price ($7.50) is lower than gemini-3.5-flash ($9.00); this is correct
per the official page (improved efficiency at same input price).gemini-3.5-flash-lite appeared on the official
AI Studio pricing page in July 2026 at $0.30/MTok input, $2.50/MTok output. No large-context
tier. Context caching is NOT available for this model (explicitly listed as "Not available"
on the official AI Studio pricing page for all tiers — Standard, Batch, Flex, Priority).
The entry was initially added on July 22 2026 with cache pricing; the cache keys were
removed on July 23 2026 after official confirmation. Do NOT add cache pricing for this model
unless the official page explicitly adds it. Follows the same selectable-model pattern as
gemini-3.1-flash-lite.claude-opus-5 appeared on the official Anthropic pricing and models pages in July 2026. API ID: claude-opus-5 (no date suffix, pinned snapshot). Bedrock ID: anthropic.claude-opus-5. Google Cloud ID: claude-opus-5. Pricing: $5/$25 MTok input/output, 5m cache $6.25/MTok, 1h cache $10/MTok, cache read $0.50/MTok — same as Opus 4.8/4.7/4.6. The model is in the flat long-context list (1M token context at standard pricing; no Large Context tier). Fast mode is available at $10/$50 MTok (shared price point with Opus 4.8). Added to pricing file and anthropicModels in the July 25 2026 audit. matchPattern: (?i)^((anthropic\/)?claude-opus-5|(eu\.|us\.|apac\.|global\.)?anthropic\.claude-opus-5(-v1(:0)?)?)$.https://developers.openai.com/api/docs/models/gpt-5-chat-latest (July 2026 audit).
A prior audit WebFetch of the overview pricing page returned an artifact suggesting
"$5/$30", which was confusion with gpt-5.6-sol pricing. When a pricing summary for a
model alias appears inconsistent with what the file holds, always fetch the specific
model page (https://developers.openai.com/api/docs/models/<model-id>) to confirm.gpt-5.3-codex appeared on the OpenAI pricing
page and model page in July 2026, described as "the most capable agentic coding model".
Pricing: $1.75/MTok input, $0.175/MTok cached input, $14.00/MTok output. Context window:
400k tokens; max output 128k tokens. No large-context tier. No date-stamped snapshot at
launch. Standard OpenAI matchPattern: (?i)^(openai\/)?(gpt-5.3-codex)$. Added to pricing
file and openAIModels in July 27 2026 audit. Official sources:
https://developers.openai.com/api/docs/pricing and
https://developers.openai.com/api/docs/models/gpt-5.3-codex.gpt-5.6-terra and gpt-5.6-luna sometime between the July 27 and July 31 2026 audits;
gpt-5.6-sol was unchanged. Confirmed via 4 independent WebFetch calls (the overview
pricing page fetched twice plus each model's dedicated page): gpt-5.6-terra is now
$2.00/MTok input, $0.20/MTok cached input, $12.00/MTok output (previously
$2.50/$0.25/$15.00); gpt-5.6-luna is now $0.20/MTok input, $0.02/MTok cached input,
$1.20/MTok output (previously $1.00/$0.10/$6.00). The >272K Large Context tier still
applies at 2x input / 1.5x output, with cached input also doubling (preserving the 10%
cache-to-input ratio): terra large-context $4.00/$0.40/$18.00, luna large-context
$0.40/$0.04/$1.80. gpt-5.6-sol remains $5.00/$0.50/$30.00 standard,
$10.00/$1.00/$45.00 large-context — unchanged. Updated in the pricing file during the
July 31 2026 audit. Official sources: https://developers.openai.com/api/docs/pricing,
https://developers.openai.com/api/docs/models/gpt-5.6-terra,
https://developers.openai.com/api/docs/models/gpt-5.6-luna. Lesson: do not assume a
model family's siblings keep moving in lockstep — verify each model ID's own page even
when the whole family was fully priced in a recent prior audit.Capture:
Values in default-model-prices.json are per token, not per million tokens.
| Provider Price | JSON Value |
|---|---|
$5 / MTok | 5e-6 |
$25 / MTok | 25e-6 |
$0.50 / MTok | 0.5e-6 |
$6.25 / MTok | 6.25e-6 |
Formula:
price_per_token = price_per_mtok / 1_000_000
Use provider-usage-key-matrix.md as the single source of truth for OpenAI, Gemini, Anthropic, and Bedrock usage aliases. Do not copy a partial key set from this pricing-source reference or from an older model entry.