OpenAI October 2026 Model Deprecation Planning Guide

October 2026 brings two separate shutdown waves for OpenAI API users. On October 1, the specialized gpt-5.4-cyber model goes dark. On October 23, a larger batch of legacy models — gpt-3.5-turbo-0125, gpt-4-0613, gpt-4-turbo-2024-04-09, and o1-2024-12-17 — stops accepting requests. If your codebase pins any of these model IDs, you have days (not months) to act.

This guide provides the exact replacement mapping for each affected model, before-and-after code diffs, a pricing delta analysis, Azure Foundry retirement context, and a step-by-step production checklist. We also cover how TheRouter's fallback routing can automate the transition so your application keeps serving requests through the shutdown window.

We route OpenAI-compatible requests through configured providers and support provider/model routing and fallback where the live product path supports it. We do not claim universal model support, zero downtime, or guaranteed cheapest pricing. Every date and pricing figure below is sourced from OpenAI's official documentation at retrieval time.

OpenAI-compatible means a provider exposes a chat-completions endpoint whose request and response shape matches the OpenAI API contract closely enough that an unmodified OpenAI SDK call works against it after swapping three values: API key, base URL, and model name. The minimum surface in practice is POST /v1/chat/completions with messages, model, and an OpenAI-shaped streaming response.

30-Second Answer

• October 1, 2026: gpt-5.4-cyber shuts down. Replace with gpt-5.6-cyber (Daybreak Red access required) (OpenAI Deprecations, retrieved 2026-09-18).
• October 23, 2026: gpt-3.5-turbo-0125, gpt-4-0613, gpt-4-turbo-2024-04-09, and o1-2024-12-17 shut down (OpenAI Deprecations, retrieved 2026-09-18).
• Recommended replacements: GPT-5.6 family — Sol for flagship, Terra for mid-tier, Luna for lightweight workloads.
• Action required now: Audit model IDs, shadow-test replacements, configure fallback routing. The October 1 deadline is 13 days out.

Complete October 2026 Deprecation Timeline

Two distinct waves hit in October. The first is a specialized cybersecurity model retirement. The second is the final cleanup of legacy GPT-3.5, GPT-4, and early reasoning model snapshots.

Wave 1 — October 1: GPT-5.4-Cyber Shutdown

GPT-5.4-Cyber was OpenAI's first purpose-trained cybersecurity model, launched in April 2026 as part of the Daybreak program. Its successor, GPT-5.6-Cyber, was announced on August 10, 2026 and is built on the Sol architecture with significantly improved completion rates on advanced cybersecurity prompts — 95% versus 57.3% for GPT-5.5-Cyber on OpenAI's internal advanced cybersecurity eval (OpenAI Daybreak announcement, retrieved 2026-09-18).

GPT-5.6-Cyber is gated. There is no public API model ID you can drop into a request. Access requires enrollment in the Daybreak Red program with identity verification, approved-use restrictions, and a hardware security key on your account from September 1, 2026. Teams not enrolled in Daybreak Red should use standard gpt-5.6-sol, which scores 96.7% on published capture-the-flag benchmarks and 73.5% on ExploitBench without any Daybreak access (eesel.ai GPT-5.6-Cyber overview, retrieved 2026-09-18).

Wave 2 — October 23: Legacy Model Mass Retirement

This wave removes the last remaining date-stamped snapshots of GPT-3.5, GPT-4, and the initial O1 reasoning model. OpenAI announced these retirements on April 22, 2026, providing the standard 6-month notice window for generally available models (OpenAI Community Forum, retrieved 2026-09-18).

After October 23, the undated gpt-3.5-turbo alias will also stop working if it points to the retired 0125 snapshot. Verify which snapshot your alias currently resolves to before the deadline.

Migration Path: Model-by-Model Replacement Guide

GPT-3.5-Turbo to GPT-5.6-Terra

GPT-5.6-Terra is the mid-tier model in the GPT-5.6 family. For workloads that used GPT-3.5-Turbo for cost-efficiency, Terra provides a significant capability upgrade at a higher price point.

The cost increase is substantial — 4x on input, 8x on output at standard rates. Two strategies reduce the impact. First, GPT-5.6-Terra supports prompt caching at $0.20/1M cached input tokens, a 90% discount over uncached. For repeated-prefix workloads, this can offset much of the price increase. Second, batch processing cuts rates by 50% for non-latency-sensitive jobs (OpenAI Pricing, retrieved 2026-09-18).

For high-volume, cost-sensitive workloads where GPT-3.5-Turbo's capability was sufficient, consider gpt-5.6-luna ($0.20/$1.20 per 1M tokens) as an alternative. Luna's pricing is closer to legacy GPT-3.5 rates while still offering GPT-5.6-class capabilities.

GPT-4 and GPT-4-Turbo to GPT-5.6-Sol

Both gpt-4-0613 and gpt-4-turbo-2024-04-09 map to gpt-5.6-sol, the flagship tier.

For GPT-4 (0613) users, the migration to Sol is a 87% cost reduction on input and 67% on output. For GPT-4-Turbo users, the savings are 60% on input and 33% on output. This is one of the rare deprecation migrations where the replacement model is both more capable and less expensive (OpenAI Pricing, retrieved 2026-09-18).

O1 to GPT-5.6-Sol

The o1-2024-12-17 reasoning model maps to gpt-5.6-sol. OpenAI unified the reasoning and general-purpose model lines with the GPT-5.6 family. The O-series naming convention is discontinued for new releases.

Sol provides reasoning capabilities through reasoning.mode parameters rather than a separate model family. For workloads that relied on O1's extended reasoning, use gpt-5.6-sol with reasoning.mode: "pro" for comparable depth.

Before/After Code: Updating Model IDs

Python SDK

Before — pinned to deprecated models:

from openai import OpenAI

client = OpenAI()

# Legacy GPT-4 call
response = client.chat.completions.create(
    model="gpt-4-turbo-2024-04-09",  # shuts down Oct 23
    messages=[{"role": "user", "content": "Analyze this vulnerability report"}]
)

After — migrated to replacement:

from openai import OpenAI

client = OpenAI()

# Migrated to GPT-5.6-Sol
response = client.chat.completions.create(
    model="gpt-5.6-sol",  # recommended replacement
    messages=[{"role": "user", "content": "Analyze this vulnerability report"}]
)

Node.js / TypeScript

Before:

const response = await openai.chat.completions.create({
  model: "gpt-3.5-turbo-0125",  // shuts down Oct 23
  messages: [{ role: "user", content: "Summarize this document" }],
});

After:

const response = await openai.chat.completions.create({
  model: "gpt-5.6-terra",  // recommended for cost-efficient workloads
  messages: [{ role: "user", content: "Summarize this document" }],
});

cURL

Before:

curl https://api.openai.com/v1/chat/completions \
  -H "Authorization: Bearer $OPENAI_API_KEY" \
  -d '{"model": "o1-2024-12-17", "messages": [{"role": "user", "content": "Solve this step by step"}]}'

After:

curl https://api.openai.com/v1/chat/completions \
  -H "Authorization: Bearer $OPENAI_API_KEY" \
  -d '{"model": "gpt-5.6-sol", "messages": [{"role": "user", "content": "Solve this step by step"}]}'

Azure Foundry Implications

Azure Foundry (formerly Azure OpenAI Service) runs an independent retirement schedule that generally trails the OpenAI direct API by several months. Microsoft's model retirement schedule follows an 18-month lifecycle from GA deployment.

Key differences for Azure users:

• GPT-4 and GPT-4-Turbo retirements on Azure may differ from the October 23 OpenAI API date. Check the Azure Foundry model retirement schedule for your specific deployment region.
• Azure already retired some GPT-4o versions (2024-05-13 and 2024-08-06 snapshots) on March 31, 2026 (Microsoft Learn, retrieved 2026-09-18).
• Fine-tuned model deployments on deprecated base models will continue to work until the base model is fully retired, but no new fine-tuning jobs can be created.

If you deploy through both Azure and the OpenAI direct API, audit each environment separately. A model that is still live on Azure may already be gone from the direct API, or vice versa.

TheRouter Fallback Configuration for October Shutdowns

TheRouter's model fallback routing lets you configure automatic rerouting so that when a deprecated model ID starts returning errors, traffic shifts to the replacement without code changes.

Example fallback configuration for the October 23 wave:

# therouter.yaml — fallback rules for October 2026 deprecations
routes:
  - match:
      model: "gpt-3.5-turbo-0125"
    fallback:
      - model: "gpt-5.6-luna"    # cost-optimized replacement
      - model: "gpt-5.6-terra"   # capability upgrade if Luna is insufficient

  - match:
      model: "gpt-4-0613"
    fallback:
      - model: "gpt-5.6-sol"

  - match:
      model: "gpt-4-turbo-2024-04-09"
    fallback:
      - model: "gpt-5.6-sol"

  - match:
      model: "o1-2024-12-17"
    fallback:
      - model: "gpt-5.6-sol"

This configuration catches requests still using deprecated model IDs and routes them to the recommended replacements. Deploy the fallback rules before October 1 to cover the GPT-5.4-Cyber shutdown, and before October 23 for the legacy model wave.

For more on fallback configuration, see our LLM API fallback routing guide.

How This Fits the Broader 2026 Deprecation Landscape

October 2026 is one of several deprecation waves across providers this year. Understanding the full timeline helps teams plan capacity and avoid surprise outages.

Production Migration Checklist

1. Swap three values, not three SDKs. Change api_key, base_url, and model in the existing OpenAI client. Keep your request/response code unchanged.
2. Map model IDs explicitly. The target provider's model id is almost never identical to the OpenAI id. Keep a single dict of { openai_id: target_id } outside business logic.
3. Verify streaming format. SSE chunks must follow the OpenAI data: {...} + data: [DONE] contract. Test one streaming call before moving production traffic.
4. Check rate-limit headers. Some providers omit x-ratelimit-* headers. Add a wrapper that defaults safely when headers are absent.
5. Keep a rollback path. Ship the swap behind a feature flag, run both endpoints in shadow for 24 hours, then cut over.

October-specific steps

1. Audit model IDs — Search your codebase, environment variables, and configuration files for gpt-5.4-cyber, gpt-3.5-turbo-0125, gpt-3.5-turbo, gpt-4-0613, gpt-4, gpt-4-turbo-2024-04-09, gpt-4-turbo, and o1-2024-12-17. Include third-party libraries and workflow tools that may hardcode model IDs.

2. Check alias resolution — The undated aliases gpt-3.5-turbo, gpt-4, and gpt-4-turbo may resolve to the retiring snapshots. Call the /v1/models endpoint and verify what your alias currently points to.

3. Shadow-test replacements — Run your evaluation suite against gpt-5.6-sol and gpt-5.6-terra in parallel with your current models. Compare output quality, latency, and cost.

4. Configure fallback routing — Deploy TheRouter fallback rules (see above) as a safety net. Even if you update model IDs in code, fallback routing catches any missed references.

5. Monitor gateway logs — After deploying, watch for requests still hitting deprecated model IDs. These indicate missed references in your codebase or third-party integrations.

6. Verify Azure separately — If you use Azure Foundry, check Microsoft's retirement schedule independently. Do not assume OpenAI API dates apply to Azure deployments.

7. Test error handling — After shutdown, deprecated model requests will return HTTP 400 or 404 errors. Verify your application handles these gracefully and surfaces actionable error messages.

FAQ

Q: Will the undated gpt-4 alias keep working after October 23?

The undated gpt-4 alias currently resolves to the gpt-4-0613 snapshot, which shuts down October 23. After that date, OpenAI may remap the alias to a newer model or remove it entirely. Do not rely on alias stability through a deprecation window — pin to the replacement model ID explicitly.

Q: I use GPT-3.5-Turbo for high-volume, low-cost work. What is the cheapest replacement?

gpt-5.6-luna at $0.20/$1.20 per 1M input/output tokens is the closest price match to GPT-3.5-Turbo's $0.50/$1.50 rates. Luna offers substantially better capabilities while keeping costs in the same order of magnitude. For batch workloads, Luna's batch pricing drops to $0.10/$0.60.

Q: Is gpt-5.6-cyber available through the standard OpenAI API?

No. GPT-5.6-Cyber is gated behind the Daybreak Red program. You must apply, pass identity verification, agree to approved-use restrictions, and have a hardware security key on your account. For general cybersecurity work, OpenAI recommends Daybreak Blue (which uses standard Sol with guardrails removed) or standard Sol without any Daybreak access.

Q: What happens if I send a request to a shutdown model?

The API returns an HTTP 400 or 404 error indicating the model does not exist or has been retired. Your application receives no completion — the request fails immediately.

---

Sources cited in this guide were retrieved on September 18, 2026. OpenAI may update deprecation dates, pricing, or replacement recommendations after publication. Always verify against the official OpenAI deprecations page before making production changes.