What this run shows. The Assistants API stored conversation history in a server-side Thread, so each new turn only sent the new message. The Responses API has no equivalent server-managed thread unless you attach a Conversation object; the common port pattern is to carry prior turns yourself. This two-run check measures what that manual state-carrying costs on the input side.

Measured result: carrying a single prior turn by hand added 66 input tokens on turn 2 (196 vs 130), a 51% increase in input tokens for that turn. Both runs used model gpt-5.5-2026-04-23 via the /v1/responses endpoint. The gap compounds with every turn and grows much faster once tool outputs (file-search chunks, code-interpreter results) join the history.

Honesty boundary for this run. This demonstrates the manual resend pattern only. It does not exercise previous_response_id chaining or the Conversations API. Per OpenAI's own docs, previous_response_id still bills all prior input tokens in the chain, so the token-growth reality holds there too; the Conversations API changes storage/retention, not per-turn token billing. Real billed cost, latency, and any production migration outcome are [DATA NOT AVAILABLE] — this is a bounded lab check, not a benchmark.

Effloow Lab OpenAI API Run: openai-assistants-api-sunset-responses-conversations-port-poc-2026

• Date: 2026-07-06T00:39:36.755580+00:00
• Model: gpt-5.5-2026-04-23
• Purpose: Migration PoC turn 1/2: baseline Responses call, no carried state
• Request ID: 52556b68-fd30-4090-8d62-1ef79a41c558
• Usage: {"input_tokens": 130, "input_tokens_details": {"cached_tokens": 0}, "output_tokens": 50, "output_tokens_details": {"reasoning_tokens": 22}, "total_tokens": 180}
• Prompt SHA-256: b595699958c07ba2e50f79ab7686a849519c2981dee853a3f40efe677b1c7482
• Safety boundary: no confidential, customer, credential, or private data should be included in this run.

Prompt Excerpt

Simulated migration test: a stateful assistant ported to the Responses API, turn 1 of 2.

[Conversation so far]
User: We shipped two features in "Beacon" v2.3 — a dark mode and CSV export. Briefly acknowledge this, then ask one short question about who the release notes are for.

Respond only as the assistant, in 2 sentences maximum.

Output

Got it — Beacon v2.3 includes dark mode and CSV export. Who are the release notes for?

Limitations

This API run is a bounded lab check. It is not a production benchmark, user study, or proof that an external product works in a real customer environment.

Effloow Lab OpenAI API Run: openai-assistants-api-sunset-responses-conversations-port-poc-2026

• Date: 2026-07-06T00:39:52.521501+00:00
• Model: gpt-5.5-2026-04-23
• Purpose: Migration PoC turn 2/2: manual state carry, prior turn re-sent as input
• Request ID: cc338ee0-77ba-4d90-a45a-726fb391e832
• Usage: {"input_tokens": 196, "input_tokens_details": {"cached_tokens": 0}, "output_tokens": 116, "output_tokens_details": {"reasoning_tokens": 44}, "total_tokens": 312}
• Prompt SHA-256: c2301f53cfdc59306347c25a6394f8de21c1545c462a9da6173bb825d681576b
• Safety boundary: no confidential, customer, credential, or private data should be included in this run.

Prompt Excerpt

Simulated migration test: a stateful assistant ported to the Responses API, turn 2 of 2. Because there is no server-managed Thread, the full prior turn is re-sent as input.

[Conversation so far]
User: We shipped two features in "Beacon" v2.3 — a dark mode and CSV export. Briefly acknowledge this, then ask one short question about who the release notes are for.
Assistant: Got it — Beacon v2.3 includes dark mode and CSV export. Who are the release notes for?
User: The audience is non-technical small-business owners. Draft a 3-sentence release note for them.

Respond only as the assistant, in 3 sentences maximum.

Output

Beacon v2.3 is here with a new dark mode that makes the app easier on your eyes, especially in low light.  
You can also export your data to CSV, making it simple to open and share information in spreadsheets.  
These updates are designed to help you work more comfortably and keep your business data more flexible.

Limitations

This API run is a bounded lab check. It is not a production benchmark, user study, or proof that an external product works in a real customer environment.