Portfolio 11 · request envelope

LLM token and context budget planner

Work only with pasted or estimated counts—not prompt content—to reconcile system text, tool definitions, retained history, retrieval, provider overhead, and reserved output.

Decision this answers: How many conversation turns fit before the first invalid request, and which component should be trimmed without pretending an estimate is tokenizer-exact?

Scenario

Decision summary

Complete envelope

79200model estimate

The evidence class cannot be stronger than its weakest component.

[1][2][3]

Remaining capacity

48800model estimate

Default request fits: true.

[1]

Complete turns retained

15exact-derived

Simulation stops before the first invalid request.

[3]

Context utilization

61.9%model estimate

Envelope divided by the selected limit.

[1]

Direct-labeled result profile

Complete envelope Remaining capacity Complete turns retained Context utilization

The tables below are the accessible source of truth; bar lengths never carry identity alone.

Count provenance ledger

Count provenance ledger
Candidate	Count	Evidence	Included in links	Boundary	Action	Evidence
Provider CountTokens result	Paste numeric result	Measured	Number only	Model-specific	Use directly	measured[1]
Local tokenizer estimate	Numeric estimate	Model estimate	Number only	Tokenizer/version-specific	Do not call exact	model estimate[2]
Unknown provider overhead	Unknown	Unknown	No content	Cannot prove fit	Paste official count	unknown[3]

Official surface versus this site

Gemini CountTokens API

Use the official surface to confirm native prices, counting rules, limits, and product eligibility.

This offline workbench

Provider counting APIs return authoritative model-specific counts. This offline page lets those counts be pasted locally, combines them into a complete envelope, and simulates retention without transmitting content.

Sources and methodology

Every result-affecting reference is visible here without JavaScript and is retained in the JSON export.

[1] Gemini CountTokens API

Authority: Google
Native identifier: models.countTokens
Unit: tokens
Evidence: measured
As of: 2026-08-01
Retrieved: 2026-08-01
Freshness: reference
Normalization: Official counts can be pasted as numbers; no content is transmitted by this page.

Open the authoritative source

[2] OpenAI token counting guide

Authority: OpenAI
Native identifier: token counting guide
Unit: tokens
Evidence: measured
As of: 2026-08-01
Retrieved: 2026-08-01
Freshness: reference
Normalization: Model-specific counts remain distinct from local estimates and raw content is never exported.

Open the authoritative source

[3] Anthropic token counting

Authority: Anthropic
Native identifier: token counting API
Unit: tokens
Evidence: measured
As of: 2026-08-01
Retrieved: 2026-08-01
Freshness: reference
Normalization: Provider results can be entered numerically while prompt and tool text stay local and out of URLs.

Open the authoritative source

Limits and non-claims

Share links contain numeric counts only—never prompts, messages, documents, or tool schemas.
A total containing an estimate remains an estimate even when the addition is exact.
Context fit does not imply quality, latency, output completion, or provider acceptance.