---
id: "20260709-1857-hop-embedding-price-stability-myth"
title: "The 'embedding prices stayed stable' claim doesn't survive OpenAI's own pricing docs — ada-002 to text-embedding-3-small was a 5x cut"
type: "capture"
status: "promoted"
promoted_to: ["claim-matryoshka-representation-learning-truncatable-embeddings (Claim 1 — MRL mechanism, Tier 1 arXiv; OpenAI/nomic adoption sub-claim flagged [unverified-mechanism])","claim-openai-embedding-price-fell-5x-ada-002-to-3-small (Claim 3 — the 5x cut, Tier 1 OpenAI docs, capture-verified; Claim 2's aggregator myth folded in as the claim-under-test being corrected)"]
not_promoted: ["Claim 2 (2026 aggregators assert 'embedding prices stayed remarkably stable') — Tier 4 content-farm listicle, explicitly the claim under test, not evidence. Not minted as a standalone note (would enshrine a myth); folded into the Claim 3 note as the framing being refuted.","Further lead: Voyage AI's Jan 2026 MoE embedding model — unverified aggregator mention, a fresh mechanism-migration thread; left as a lead, folded into question-embedding-api-price-cuts-across-providers.","Further lead: whether Cohere/Google/Voyage show OpenAI's launch-to-launch price drop — routed to question-embedding-api-price-cuts-across-providers rather than asserted.","Saved hooks not followed (Ali Farhadi bio; Matryoshka-doll folklore/etymology) — the capture itself parked these as over-quota / pure-orphan seeds; left in the inbox record."]
questions_routed: ["question-embedding-api-price-cuts-across-providers (new) — generalization check across Cohere/Google/Voyage","question-verify-openai-nomic-use-matryoshka-representation-learning (new) — the [unverified-mechanism] adoption flag on Claim 1"]
origin: "hop-batch"
model: "claude-sonnet-5"
date_created: "2026-07-09T00:00:00.000Z"
hop_chain: ["seed: nomic-embed-text (the vault's retrieval layer) -> Nomic AI's reproducible open embedder, arXiv:2402.01613 (max_cosine 0.566, not followed — orphan)","nomic-embed-text-v1.5 -> Matryoshka Representation Learning, the truncatable-embedding technique it borrows (max_cosine 0.65, chosen — adjacent to vault's retrieval cluster)","MRL paper (arXiv:2205.13147) authors -> Ali Farhadi / Aditya Kusupati / Sham Kakade bios (max_cosine 0.677 / 0.652 / 0.651, discarded — lands in the history-of-AI-figures cluster, over quota tonight per redirect)","MRL paper -> Matryoshka-doll folklore/etymology (max_cosine 0.587, novel but a pure orphan, lower priority)","MRL's commercial adopters (OpenAI's text-embedding-3 line uses MRL) -> embedding-API pricing/economics (max_cosine 0.628, chosen — bridges moc-inference-economics AND the vault's Contextual Retrieval note, the highest-value connect-but-extend signal)","Content-farm pricing round-ups' claim 'embedding prices stayed stable' -> verified/refuted against OpenAI's own docs (gate check max_cosine 0.783, adjacent — directly extends claim-inference-cost-collapsed-280x)"]
novelty_max_cosine: 0.783
tags: ["embeddings","inference-economics","myth-ledger","openai","pricing","matryoshka-representation-learning"]
source_urls: ["https://arxiv.org/abs/2205.13147","https://developers.openai.com/api/docs/models/text-embedding-3-small","https://developers.openai.com/api/docs/models/text-embedding-ada-002","https://pecollective.com/tools/text-embedding-models-compared/"]
source_authors: ["Kusupati, Bhatt, Rege, Wallingford, Sinha, Ramanujan, Howard-Snyder, Chen, Kakade, Jain, Farhadi (arXiv paper, direct fetch)","OpenAI (own developer docs, direct fetch)","pecollective.com (aggregator listicle, relayed via search)"]
source_dates: ["2022-05-26, rev. 2024-02-08 (MRL paper)","current as fetched 2026-07-09 (OpenAI docs)","2026 (aggregator page, undated within page)"]
source_tiers: ["1 — arXiv paper, direct fetch","1 — OpenAI's own current developer docs, direct fetch, for both models' prices","4 — content-aggregator listicle; the claim it carries is the thing being checked, not evidence for it"]
---


**Claim 1.** Matryoshka Representation Learning (MRL) trains a single embedding model to produce useful representations at multiple truncation lengths in one forward pass, "encod[ing] information at different granularities" — nomic-embed-text-v1.5 and OpenAI's text-embedding-3 line both use it. Source: arXiv:2205.13147 (Kusupati et al.). Tier 1.

**Claim 2.** Multiple 2026 pricing round-ups (e.g. pecollective.com) assert that "embedding prices have stayed remarkably stable compared to LLM API prices," contrasting them with steep generative-model price cuts. Tier 4 — this is the claim under test, not evidence.

**Claim 3.** OpenAI's own current developer docs list text-embedding-ada-002 (Dec 2022) at $0.10/1M tokens and text-embedding-3-small (Jan 2024) at $0.02/1M tokens — a direct 5x cut, contradicting Claim 2 for at least OpenAI's own line. Source: developers.openai.com model pages, fetched directly. Tier 1.

## Why this was hop-worthy

It's a small, checkable myth-correction that extends the vault's existing `claim-inference-cost-collapsed-280x` note (LLM token costs collapsing) with a specific counter-case in embeddings, and mirrors the vault's own `myth-inference-two-thirds-of-compute` pattern — a plausible-sounding aggregate claim that a primary source doesn't actually support.

> [!note] Seek's commentary:
> "Stayed stable" is doing a lot of work in that sentence — it's true that embedding *rate cards* move less often than LLM ones, but the one comparison I could actually check (OpenAI's own two generations) shows the same 2022-23 boom in price competition hit embeddings too. The claim is an artifact of not checking the one number that mattered.

## Further leads

- Voyage AI's Jan 2026 MoE embedding model (per aggregator sources, unverified) — mechanism migrating from LLMs into embeddings, unexplored.
- Whether Cohere/Google/Voyage show the same launch-to-launch price drop pattern as OpenAI — would confirm or narrow this to an OpenAI-specific effect.

## Hop chain

Hop 1: Nomic Embed Text (arXiv:2402.01613 / nomic.ai/news) → Matryoshka Representation Learning
- Hook type: mechanism question
- Hook: nomic-embed-text-v1.5's release notes cite Matryoshka-style truncatable dimensions as a feature; the term was unexplained.
- Why followed: highest-scoring available hook — MRL scored 0.65 ("adjacent"), close to the vault's existing Contextual Retrieval/embeddings note, versus Nomic AI itself (0.566, an orphan) or GPT4All (0.537, orphan, unrelated domain).
- Key findings: MRL (Kusupati et al., arXiv:2205.13147) trains one embedding to stay useful when truncated to far smaller dimensions — up to 14x smaller for equivalent accuracy on ImageNet-1K.

Hop 2: arXiv:2205.13147, full text — MRL paper
- Hook type: the person behind the thing (checked, discarded) / cross-domain bridge (chosen)
- Hook: two competing hooks — the author roster (Ali Farhadi of Xnor.ai/AI2; Sham Kakade) vs. the fact OpenAI's text-embedding-3 models use MRL in production.
- Why followed: author bios all scored 0.65-0.68 but land squarely in the vault's history-of-AI-figures cluster, which the session redirect marked over quota tonight. The commercial-adoption angle (0.628) instead bridges two live clusters: moc-inference-economics and the retrieval/Contextual-Retrieval note — the protocol's highest-value hook type.
- Key findings: OpenAI's text-embedding-3-small/large use MRL to let developers shorten embeddings without retraining.

Hop 3: pricing aggregator round-up (pecollective.com and similar 2026 listicles)
- Hook type: the surprising claim
- Hook: "Embedding prices have stayed remarkably stable compared to LLM API prices, where cuts of 50-80% over 18 months have been common."
- Why followed: directly contradicts the vault's existing 280x LLM-cost-collapse note by implying embeddings are the exception — but the source is Tier 4 (content-farm listicle), so the claim needed a primary check before trusting it, per sources.md's quantitative-claim floor.
- Key findings: none trustworthy yet — flagged for verification, not recorded as fact.

Hop 4: developers.openai.com, text-embedding-3-small and text-embedding-ada-002 model pages
- Hook type: verification / mechanism-of-record (zoom in, closing hop)
- Hook: OpenAI's own docs list both models' current prices side by side.
- Why followed: natural closing move — check the one comparison that would settle Hop 3's claim, at a Tier 1 source.
- Key findings: ada-002 = $0.10/1M tokens; text-embedding-3-small = $0.02/1M tokens — a 5x cut at launch (Jan 2024), refuting the "stayed stable" framing, at least for OpenAI's own line.

Saved hooks not followed:
- Ali Farhadi (MRL co-author, Xnor.ai founder, sold to Apple) — from arXiv:2205.13147 — interesting person-hook but lands in the history-of-AI-figures cluster, over quota per tonight's redirect.
- Matryoshka-doll folk-craft history/etymology — from the paper's own framing — scored 0.587 (novel) but a pure orphan with no vault connection; saved as a lower-priority seed.
- Voyage AI's MoE embedding architecture (Jan 2026) — from the pricing round-up — a fresh mechanism-migration thread (MoE moving from LLMs into embedding models), not pursued to keep this chain focused.

post-worthy: maybe — it's a clean, verifiable myth-correction that extends an existing MOC, but it's a small single-fact finding rather than a novel connection; good journal filler, not necessarily a standalone post.