Skip to main content

LTM Cost Estimation

LTM charges only for model tokens; storage is currently free. This page explains what is billed and gives a formula and examples for estimating cost from your usage. Actual usage is recorded in the Trace of each call.

What is billed​

PartWhen it occursHow it is billed
Extraction after a writeFor every Event, the background pipeline calls a model once for each enabled memory type (profile, facts, summary); when a new fact is close to an existing one, one more judge call followsAt the token prices of the model family selected for the Space
EmbeddingNew or updated facts and summaries, and the question of every SearchAt the embedding model's price, usually negligible
Reflection and activity insightsOnly when you call the generation endpointsSame as extraction
Redriving a failed derivationWhen you call the redrive endpoint, the model is called again for each failed StrategySame as extraction
Memory injected into the main modelManaged Chat injects recall results into the main model each turn, up to about 4,000 tokens by default, adjustable in the Space's managed_recall; when you call Context yourself, you decide how much to injectCounted as input tokens of the main model call, at the main model's price
ContextMakes no model callFree
Storage—Currently free

Extraction is billed to the Space's owner: a personal Space to the user, a client-scoped Space to that client. When a Managed Chat request configures turing_options.fallbacks (including auto), memory is skipped entirely and incurs no LTM cost.

Formula​

With profile, facts, and summary all enabled, extraction for each Event uses:

  • input tokens ≈ 2,050 + 3 × T + H + 900 × j
  • output tokens ≈ 300 + 20 × j

T is the number of tokens of new conversation in the Event; for Chinese, roughly one token per character. H is the earlier turns of the same session sent as context when facts are extracted, at most about 6,000 characters, and 0 for the first Event of a session or when the Space sets facts_session_history_events to 0, see Does extraction read earlier turns?. j is the number of new facts in the Event that need comparing with existing facts, usually 0–2.

When only some types are enabled, add up the rows:

Memory typeInput tokensOutput tokens
Profile600 + T50
Facts800 + T + H150
Summary650 + T (including the previous summary)100
Fact judge (per new fact that needs comparing)90020

Monthly cost:

monthly cost ≈ Events per month × (input tokens per Event × input price + output tokens × output price)
+ tokens injected into the main model per month × main model input price
+ Reflection cost
  • The coefficients are the measured medians of extraction calls with the platform's current prompts, rounded.
  • Profile, facts, summary and fact-judge calls run with the model's thinking switched off, so the output is only the extraction result; Reflection and activity insights keep it on. Models of the glm and gemini families cannot switch thinking off and run at their lowest level, so their output tokens are higher than the coefficients above.
  • Extraction is counted per Event, not per conversation turn: Managed Chat writes one Event per turn, while with the API you can write one Event per session.
  • Changing the model or strategies affects only Events written afterwards; historical data is not extracted again.

Where to find prices​

Call List LTM model families (GET /families): each family's pricing gives the input and output prices and the currency. Prices are also listed in the chat model list. The table shows example prices from November 2026 (per million tokens); the endpoint is authoritative:

FamilyDefault modelInputOutput
qwenqwen3.8-flashCNY 0.8CNY 2.7
deepseekdeepseek-flashCNY 2 (CNY 1 in discount hours)CNY 8 (CNY 4 in discount hours)
doubaodoubao-seed-2.1-turboCNY 3CNY 15
gptturing/gpt-6-lunaUSD 0.1USD 0.5
geminituring/gemini-3.8-flashUSD 0.75USD 3.75

The gemini price is promotional; see pricing_effective_until in the response for its end date. The deepseek discount hours are 12:00–14:00 and 18:00–09:00 the next day on weekdays, and all day at weekends, Beijing time; this family has a default model only, with no enhanced model. Other families and the enhanced models are listed by the endpoint.

Example​

A consumer assistant: each user has 3 sessions a day with 4 turns each, about 300 tokens per turn; profile, facts, and summary are enabled; extraction uses the qwen family.

Write patternUsage per EventEvents per monthExtraction per user per monthPer user per year
Managed Chat: one Event per turn (T = 300, j = 1)about 3,850 input and 320 output tokens360about CNY 1.4about CNY 17
Your own API calls: one Event per session (T = 1,200, j = 2)about 7,450 input and 340 output tokens90about CNY 0.6about CNY 7

Managed Chat also adds the memory injected into the main model: at 1,000–4,000 tokens per turn with qwen3.8-flash as the main model, that is about CNY 0.3–1.2 per user per month.

With the same Managed Chat pattern, extraction per user per month with other families is about CNY 1.8–3.7 for deepseek (all in discount hours to all in standard hours), CNY 5.9 for doubao, USD 0.20 for gpt, and USD 1.5–1.9 for gemini (it cannot switch thinking off completely, so its real output is above the formula). Total cost scales linearly with the number of active users.

How to reduce cost​

  • Batch writes: with the API, write one Event when a session ends instead of one per turn. That saves about 55% in the example above.
  • Enable only the types you need: an empty enabled list turns a type off, for example the summary if you do not need session summaries.
  • Pick a suitable model family: prices differ several times over between families, and extraction rarely needs a flagship model.
  • Control what you inject: set Context's token_budget to what you need, and skip injection when Search recommends abstaining (abstain_recommended is true); for Managed Chat, adjust the Space's managed_recall, see Managed Chat recall settings.
  • Generate Reflection on demand: do not call the generation endpoints every turn.

Storage size​

Storage is currently free. To estimate its size:

  • each memory is about 8 KB, of which a 1,536-dimension vector takes about 6 KB;
  • each raw Event is about the size of its conversation text;
  • each session has a single summary, which later writes update;
  • every change to a memory writes a revision, kept for 365 days.

In the example above, each user takes about 10–30 MB per year.

Estimate with AI​

Give the following to an AI together with your parameters to get monthly and yearly estimates:

Estimate the monthly and yearly cost of Turing LTM with the formula below, and show the calculation.

Extraction per Event (profile, facts, and summary enabled):
- input tokens ≈ 2050 + 3 × T + 900 × j
- output tokens ≈ 300 + 20 × j
T is the number of tokens of new conversation in the Event; j is the number of new facts that need comparing with existing facts (usually 0–2).
With only some types enabled, add up the parts: profile input 600+T, output 50; facts input 800+T, output 150; summary input 650+T, output 100; fact judge per new fact input 900, output 20.
Managed Chat writes one Event per turn and injects recall results into the main model, up to about 4000 tokens per turn by default, billed at the main model's input price.
Storage is free, Context is free, and embedding cost is negligible.

My parameters:
- Active users:
- Sessions per user per day:
- Turns per session:
- Average tokens per turn:
- Write pattern: Managed Chat (one Event per turn) or own API calls (one Event per session)
- Enabled memory types:
- Extraction model family and its prices (pricing from GET /families):
- Managed Chat main model input price (if applicable):
Self-check