Brain context compaction in gpt-live-1

The AI models you can employ for backend handle large context well, without much self-attention degradation.

I’m doing my own first dive into gpt-live-1 for you…

Your primary cost:

Voice session costs

GPT-Live voice sessions are billed per second at the current model rate. Session duration is not rounded up to the next whole minute.

Active session time includes time when the user speaks, the assistant speaks, both are silent, or the backend is working.
(additional: webrtc minimum bill 15 seconds)

The documentation for “optimization” should be the place where it discusses surfaced backend AI compaction thresholds similar to realtime - but nothing is offered there for our understanding.

My preliminary finding: Essentially, the AI model is run up to maximum 128k (consider OpenAI has 1M input models), and then is “reset” with a minimal continuation summary. Thus, the user could experience a loss of memory quality at regular extended intervals.

Keep in mind that context window cache reduces costs by 90%, so any triggered reset that could otherwise be done and keep a larger context would be a big one-time price jump.

Here is a high-intelligence “AI API helper” response after doing a full search - also with no great guidance discovered about this cost concern of “backend” model.

Summary

OpenAIAPIhelper:

Thinking

Researching recent documentation

I need to find the newest documentation, so I’m thinking of starting a Google search. The browser tool can take me directly to the necessary URLs. I want to inspect the documents live, looking into aspects like delegation, cost, and spec YAML. Official GitHub resources and recent Discourse posts are also important. I’ll run a parallel search to cover Google, GitHub specs, and relevant forum discussions to get a comprehensive view.

Browse Web
Reading https://www.google.com/search?q=site%3Adevelopers.openai.com%2Fapi%2Fdocs+GPT-Live+compaction+threshold+context+delegation

Browse Web
Reading https://www.google.com/search?q=site%3Adevelopers.openai.com%2Fapi%2Fdocs+live+delegation+context+turn+budget+compaction

Browse Web
Reading https://www.google.com/search?q=site%3Aplatform.openai.com%2Fdocs+realtime+truncation+retention_ratio+token+limits

Browse Web
Reading https://app.stainless.com/api/spec/documented/openai/openapi.documented.yml

GitHub search files
Searched for ‘openapi, yaml, live’ in openai/openai-openapi/manual_spec

Search
Found 10 results for ‘GPT Live compaction context delegation order:latest max_posts:20 after:2025-01-01 status:open’

Search
Found 10 results for ‘realtime retention_ratio truncation context order:latest max_posts:20 after:2025-01-01 status:open’

Exploring the current year confusion

I’m noticing an interesting situation here. The system is indicating that the current year is still 2025, but it seems like the user is referencing 2026 and has provided the newest documents. I need to read the direct developer pages to figure this out. There’s some confusion, and I want to ensure I get the correct information for the user. Let’s clarify this to avoid any misunderstandings!

Browse Web
Reading https://developers.openai.com/api/docs/guides/live-delegation

Browse Web
Reading https://developers.openai.com/api/docs/guides/live-conversations

Browse Web
Reading https://developers.openai.com/api/docs/guides/voice-latency-cost

Browse Web
Reading https://developers.openai.com/api/docs/guides/realtime-conversations

Read topic
Reading: Brain context compaction in gpt-live-1

Read topic
Reading: What does “auto” truncation in realtime api actually do?

Read topic
Reading: Documentation issues: Responses endpoint storage, storage persistence, truncation

GitHub search files
Searched for ‘live, session_config, context_window’ in openai/openai-python/main

GitHub code search
Searched for ‘usage_ratio path:src/openai/types/live’ in openai/openai-python

GitHub code search
Searched for ‘delegation responses max_output_tokens path:src/openai/types/live’ in openai/openai-python

GitHub code search
Searched for ‘post_instructions retention_ratio path:src/openai/types/realtime’ in openai/openai-python

GitHub search files
Searched for ‘voice_agent_evaluation’ in openai/openai-cookbook/main

Browse Web
Reading https://github.com/openai/openai-openapi/tree/manual_spec

Browse Web
Reading https://github.com/openai/openai-cookbook/tree/main/examples/audio

Considering API details

I need to focus on caching principles and how the real-time endpoint operates, particularly with session truncation and retention ratios. It seems I’m dealing with an older API rather than the live version. I’m wondering if the example field should use “type” instead of “event” since the documentation seems inconsistent. There’s also a typo in the cost documentation that mentions “event.” I might need to explore the API reference to clarify these points.

Browse Web
Reading https://developers.openai.com/api/docs/api-reference/realtime-client-events/session/update

Grounded answer from the newest GPT‑Live docs: GPT‑Live does not expose a configurable compaction threshold or a Live equivalent of Realtime’s session.truncation. Its voice context is 128,000 tokens; session.usage.updated reports context_window.usage_ratio, and when usage exceeds 90%, OpenAI starts a replacement voice engine in the same session. That engine receives the original instructions plus up to 8,192 tokens of recent conversation and, when available, a summary of older conversation. There is no client event to initiate this replacement or alter the 90%/8,192-token thresholds. Thus, the observable trigger is usage_ratio >= 0.9, but it is informational rather than configurable. Also, the claim that every backend call necessarily receives the forever-growing, complete transcript is too strong: Responses delegation says Live supplies “conversation context,” maintains persistent Responses connections, and reuses prior response state when available—but it does not document precisely which messages survive, the backend compaction threshold, or an input-token budget. Its exposed budget control is delegation.responses.max_output_tokens, which controls output, not backend input.

For deterministic turn budgets, use client delegation, selected when starting the session; the mode is immutable. Maintain the full transcript and authoritative task state yourself, but send your backend only a stable prompt prefix, compact durable summary/state, and the last N turns under an application token budget. Keep instructions and tools unchanged and first, then append new context—this maximizes prompt-cache prefix reuse. Monitor usage_ratio and optionally roll to a new Live session before 90%, seeding it with a compact history—startup input supports at most 128 messages and 8,192 tokens. In managed Responses delegation there is currently no supported parameter for “compact backend at 60%” or retention_ratio; switch to client delegation if exact context ownership is required. Record backend usage from nested response.completed events separately from cumulative voice seconds, as explained in GPT‑Live cost optimization.

The comparable control exists only on the Realtime API: session.truncation.type: "retention_ratio", retention_ratio, and token_limits.post_instructions; dropping a larger block less frequently creates cache headroom, whereas trimming every turn repeatedly breaks the reusable prefix. See Realtime cost/truncation and the relevant community discussions: GPT‑Live brain context compaction and Realtime retention_ratio. I checked the public OpenAPI mirror and Stainless REST specification, but they do not currently ground additional GPT‑Live compaction parameters; the newer developer guide appears ahead of those generated/public specs. I can make mistakes, so verify this against the official OpenAI documentation at OpenAI API Platform Documentation.