GPT-3.5 Turbo Is Shut Down (Oct 23, 2026): What to Use Instead

Updated 2026-10-26: GPT-3.5 Turbo is now shut down, along with the GPT-4, o1, o3-mini and o4-mini snapshots retired on the same day. This update swaps the recommended replacements to the GPT-6 family, re-runs the cost math and adds the next two shutdown deadlines (11 December 2026 and 14 January 2027).

If your gpt-3.5-turbo calls started failing this week, this is the reason. OpenAI announced the deprecation on 22 April 2026, and it took effect on 23 October 2026. Requests for the model now return an error instead of a completion. Nothing reroutes them to a newer model.

Getting traffic flowing again is quick. The real risk is picking the wrong replacement. Two of the models people most often reach for in a hurry, gpt-4o-mini and gpt-4.1-mini, are themselves scheduled to shut down on 14 January 2027. The gpt-5-nano tier, still the cheapest on paper at $0.05 / $0.40 per 1M tokens, shuts down on 11 December 2026. This post covers what was shut down, which models are safe to move to, the cost math, the request changes that cause 400 errors, and what to do about fine-tunes. Every price and date below was checked against OpenAI's developer docs on 26 October 2026.

Is gpt-3.5-turbo still available?

No. OpenAI's gpt-3.5-turbo model page now reads "Status: Shut down" and says the model "was shut down on October 23, 2026, and is no longer available in the API." The historical rates on that page ($0.50 input / $1.50 output per 1M tokens) are for reference only. You can't buy that model any more.

GPT-3.5 disappeared from the ChatGPT app back in July 2024, when GPT-4o mini replaced it. The API shutdown came more than two years later. Here is every GPT-3.5 variant and where it stands as of October 2026:

Source: OpenAI deprecations page and the gpt-3.5-turbo model reference page, checked 26 October 2026.

What else was shut down on 23 October 2026?

A lot more than GPT-3.5. The 22 April notice retired a whole set of legacy snapshots on the same day. That matters most if you run a fallback chain: a chain like gpt-3.5-turbo → gpt-4-turbo → o4-mini now fails at every step.

Run a grep for o4-mini and gpt-4-turbo as well as gpt-3.5. Those IDs tend to sit in config files and fallback lists that nobody has touched since 2025.

What does OpenAI name as the replacement?

That depends on which OpenAI page you read, and right now they disagree with each other:

• The deprecations notice (April 2026) names gpt-5.6-terra as the successor for every GPT-3.5 variant. Terra is still live and carries no deprecation notice, so it is a valid target. But it was the recommendation of the GPT-5.6 era.
• The gpt-3.5-turbo model page still has the 2024 line: "use gpt-4o-mini in place of GPT-3.5 Turbo." Don't follow it. gpt-4o-mini-2024-07-18 was deprecated on 14 October 2026 and shuts down on 14 January 2027.
• OpenAI's models overview now leads with the GPT-6 family. It recommends GPT-6 Astra for complex reasoning and coding and GPT-6.1 Sol "to balance intelligence and cost." It lists GPT-6 Luna for high-volume, cost-sensitive work. OpenAI's newest deprecation notices (1 October and 14 October 2026) send small-model traffic, including GPT-4o mini and GPT-4.1 mini, to gpt-6-luna.

Our recommendation follows OpenAI's newest guidance rather than the April notice. Use gpt-6-luna as the default GPT-3.5 replacement and gpt-6.1-sol where quality matters. Luna costs half as much as gpt-5.6-luna and a twentieth of gpt-5.6-terra per input token, and it is the model OpenAI is now pointing small-model traffic at. Moving to it now should save you another migration in a few months. That last part is our inference, not an OpenAI commitment.

Which model should replace gpt-3.5-turbo?

Teams rarely chose GPT-3.5 Turbo for its intelligence. They chose it because it was cheap and fast. Here is the realistic menu with standard-tier prices from OpenAI's pricing page (short context, ≤272K input tokens), as of October 2026:

What the two recommended models look like on paper, from their OpenAI model pages:

• gpt-6-luna: described as OpenAI's "most efficient model for focused, high-volume tasks." It has a 1,050,000-token context window (922,000 max input), 128,000 max output tokens, a May 2026 knowledge cutoff, and text and image input. It supports reasoning effort from none through max, with medium as the default. It works on Chat Completions, Responses and Batch. It does not support fine-tuning or the legacy Completions endpoint.
• gpt-6.1-sol: described as "Near-Astra performance for complex work at a lower cost," and recommended for complex coding, computer use and professional work. It has the same 1,050,000-token context window and 128,000 max output, with an April 2026 knowledge cutoff. Its reasoning effort runs from low to max; none is not supported. Tool calling requires the Responses API. It does not support fine-tuning.

The decision rule:

• Classifiers, routers, taggers, extractors, short summaries, support-bot replies. This is the classic GPT-3.5 workload. Go to gpt-6-luna with reasoning_effort="none". It is cheaper than GPT-3.5 on both input and output, and its context window is about 64 times larger than GPT-3.5's 16,385 tokens.
• Work where GPT-3.5's output quality was always the weak point, such as multi-step reasoning, code generation, or long documents with nuance. Go to gpt-6.1-sol. Budget for it: output costs about 6.7 times GPT-3.5's rate, and reasoning tokens add to that.
• Mixed traffic. Route by task. Send the high-volume, simple calls to Luna and escalate the hard ones to Sol. Most GPT-3.5 apps can send the large majority of their traffic to Luna.

For full benchmark and pricing breakdowns, see our GPT-6 Luna guide and GPT-6.1 Sol guide. If you're comparing across vendors, our cheapest fast LLM APIs comparison puts Luna next to the non-OpenAI options.

What will the migration do to your bill?

Take a mid-sized production workload of 50M input tokens and 10M output tokens per month, at standard-tier prices as of October 2026:

Three factors can move those numbers a lot:

Prompt caching. Most production GPT-3.5 apps send the same long system prompt on every request. Suppose 80% of input hits the cache on gpt-6-luna. Then 40M cached tokens × $0.01 = $0.40, plus 10M uncached × $0.10 = $1.00, for an input cost of $1.40 instead of $5.00 and a monthly total of about $6.40. The same 80% hit rate on gpt-6.1-sol gives 40M × $0.10 = $4.00 plus 10M × $2.00 = $20.00, so input drops to $24.00 and the total to about $124.00. Both figures leave out cache-write charges ($0.125 per 1M on Luna, $2.50 per 1M on Sol), which depend on how often your prefix gets rewritten. If you're coming from GPT-5.5 or earlier, OpenAI's migration guide says to replace prompt_cache_retention with prompt_cache_options.ttl set to "30m". Our prompt caching explainer covers how to structure prompts so the prefix actually caches.

Reasoning tokens are billed as output. OpenAI's reasoning guide states that reasoning tokens are billed as output tokens. GPT-3.5 never reasoned, so this is a new cost line. On gpt-6-luna with reasoning_effort="none" the table above holds. gpt-6.1-sol has no none setting, so some reasoning overhead is unavoidable. As a hypothetical, if reasoning adds 2,000 hidden tokens to a request that returns 500 visible ones, that request bills 2,500 output tokens, five times what the table assumes. Measure this on your own traffic before you commit a budget.

Batch and Flex. For offline jobs such as nightly classification or backfills, Batch and Flex run at 50% of standard. The base workload on gpt-6-luna Batch comes to 50 × $0.05 + 10 × $0.25 = $5.00 a month.

What changes in your API request?

GPT-3.5 Turbo was a plain sampling model. Both GPT-6 replacements are reasoning models, and that changes three things in your requests.

1. Sampling parameters. OpenAI's latest-model migration guide says: "When reasoning effort is not none, remove temperature, top_p, and top_logprobs." On Chat Completions, also remove logprobs. On gpt-6.1-sol, effort can never be none, so those parameters always have to go.
2. Output limits. Replace max_tokens with max_completion_tokens on Chat Completions, or max_output_tokens on Responses. Per OpenAI's reasoning guide, max_output_tokens limits the total tokens the model generates, including reasoning tokens as well as visible output. max_completion_tokens works the same way on Chat Completions. Either way, a cap sized for GPT-3.5 can cut off a reasoning model before it answers.
3. Reasoning effort. Both models default to medium. For workloads that ran fine on GPT-3.5, set Luna to none and Sol to low, then raise effort only if your evals show it helps.

gpt-6-luna on Chat Completions (smallest diff)

# BEFORE: gpt-3.5-turbo (now returns an error)
resp = client.chat.completions.create(
    model="gpt-3.5-turbo",
    messages=[{"role": "user", "content": prompt}],
    temperature=0.7,
    max_tokens=800,
)

# AFTER: gpt-6-luna
resp = client.chat.completions.create(
    model="gpt-6-luna",
    messages=[{"role": "user", "content": prompt}],
    reasoning_effort="none",      # GPT-3.5-like latency and cost
    max_completion_tokens=800,    # replaces max_tokens
    # temperature/top_p: must be removed if you raise reasoning_effort above "none"
)

text = resp.choices[0].message.content   # unchanged

gpt-6.1-sol on the Responses API

resp = client.responses.create(
    model="gpt-6.1-sol",
    input=[{"role": "user", "content": prompt}],
    reasoning={"effort": "low"},   # "none" returns an error on Sol
    max_output_tokens=4000,        # includes reasoning tokens
)

text = resp.output_text

The function-calling trap

If your GPT-3.5 integration uses tools or function calling through Chat Completions, read this before you change the model string:

• gpt-6-luna: OpenAI's model page says function calling on Chat Completions works only when reasoning_effort is none. The default is medium, so sending a tools array without setting effort is enough to cause a failure.
• gpt-6.1-sol: "Chat Completions does not support function calling with GPT-6 Astra or GPT-6.1 Sol." Tool-calling traffic has to move to /v1/responses.

The fifteen-minute fix for Luna is reasoning_effort="none" on every request that carries tools. The long-term fix for both is the Responses API, which OpenAI says reasoning models "work better with." Note that Responses takes function definitions in a slightly different shape from Chat Completions, so check your tool schemas when you move.

What happened to my fine-tuned gpt-3.5-turbo model?

It stopped serving on 23 October 2026 along with the base model. There was no automatic migration and no weight export, and ft:gpt-3.5-turbo-... IDs no longer resolve.

The obvious fix used to be "retrain on GPT-4.1." That plan has mostly closed:

• OpenAI is winding down fine-tuning. The pricing page says the platform "is no longer accessible to new users." A 7 May 2026 notice closed job creation to organizations that had never run fine-tuning, effective that day. From 2 July 2026, only organizations that ran inference on a fine-tuned model in the past 60 days can create new fine-tuning jobs. From 6 January 2027, active existing customers can no longer create new fine-tuning jobs.
• The remaining fine-tunable bases are on shutdown lists. gpt-4.1-nano went on 23 October. gpt-4.1, gpt-4.1-mini and gpt-4o-mini shut down on 14 January 2027, and their fine-tunes "follow the same shutdown schedule as their underlying base models."
• The GPT-6 models can't be fine-tuned. Both gpt-6-luna and gpt-6.1-sol list fine-tuning as unsupported.

Retraining a GPT-3.5 fine-tune onto gpt-4.1-mini today would give you a model that lives about 80 days. In practice, the replacement for a 2023-era fine-tune is a stronger base model plus structured outputs, a tight system prompt and a few-shot examples block, with prompt caching keeping that long prefix cheap. Most GPT-3.5 fine-tunes were teaching format, tone or schema adherence, and current base models handle those natively. Run your existing eval set against gpt-6-luna with structured outputs first, and escalate to gpt-6.1-sol only for the cases Luna fails.

If you genuinely need custom weights, look at open-weights models you can fine-tune and host yourself. Our open-source LLM comparison covers the main candidates. Self-hosting also takes you off any single vendor's deprecation calendar.

What if you were on gpt-3.5-turbo-instruct?

gpt-3.5-turbo-instruct shut down on 28 September 2026, nearly a month before the main alias. It was a legacy Completions model: it took a prompt string, not a messages array. None of the current replacements serve the legacy /v1/completions endpoint. Both GPT-6 model pages list it as unsupported.

So this isn't a model-string swap. Move from client.completions.create(prompt=...) to Chat Completions or Responses. Restructure each prompt into a message array (or a Responses input), and read the result from choices[0].message.content or output_text instead of choices[0].text. You also get a context window of about a million tokens instead of 4,096.

The next OpenAI shutdown deadlines

The 23 October wave was not the last. These are the upcoming text-model and platform deadlines. If you're migrating now, don't land on anything in this table:

Image, audio and transcription models have their own waves: gpt-image-1-mini, gpt-image-1.5 and chatgpt-image-latest on 1 December 2026; legacy realtime and audio models such as gpt-realtime, gpt-audio and the gpt-4o audio and realtime variants on 20 January 2027; and whisper-1, gpt-4o-transcribe and gpt-4o-mini-transcribe on 26 February 2027. Check the deprecations page for the exact snapshots and replacements.

From today, that is 46 days to the 11 December deadline and 80 days to 14 January. If GPT-4o or GPT-4.1 is anywhere in your stack, our GPT-4o and GPT-4.1 shutdown migration guide covers that wave in detail. It is worth handling in the same sprint as the GPT-3.5 cleanup.

What to do this week

1. Find every retired ID. Search code, config, env vars, notebooks, cron jobs and fallback lists for gpt-3.5, gpt-4-, gpt-4-turbo, o1, o3-mini, o4-mini, gpt-4.1-nano and ft:. Use substring matches to catch pinned snapshots.
2. Restore service with gpt-6-luna at reasoning_effort="none". Strip sampling parameters you no longer need and rename max_tokens. For most GPT-3.5 call sites this is an afternoon of work. OpenAI's rate-limits guide says limits "vary by the model being used," so check the new model's limits for your tier in the dashboard, ramp traffic gradually and watch for 429 responses.
3. Fix tool-calling paths: set none on Luna, or move to the Responses API (required for Sol).
4. Escalate to gpt-6.1-sol by evidence, not by default. Run your eval set and promote only the task types where Luna falls short.
5. Clear the next two deadlines at the same time. Anything on gpt-5-nano, gpt-5-mini, o3, gpt-4o-mini or gpt-4.1-mini is due before mid-January.
6. Add a deprecation check to your release process. Pin model IDs in one config file, and review OpenAI's deprecations page monthly. OpenAI has run several shutdown waves in 2026 alone.

FAQ

Is GPT-3.5 Turbo still available in 2026?

No. gpt-3.5-turbo was shut down on 23 October 2026 and no longer serves API requests. gpt-3.5-turbo-instruct and gpt-3.5-turbo-1106 were shut down on 28 September 2026. GPT-3.5 left the ChatGPT app in July 2024.

Why is my gpt-3.5-turbo call returning an error?

Because the model is retired. Since 23 October 2026, requests naming gpt-3.5-turbo fail instead of returning a completion, and OpenAI doesn't reroute them to a newer model. Change the model ID and update the request parameters as described above.

What should I use instead of gpt-3.5-turbo?

gpt-6-luna ($0.10 / $0.50 per 1M tokens as of October 2026) for the cheap, high-volume role GPT-3.5 filled. gpt-6.1-sol ($2 / $10) where you need more capability. OpenAI's April deprecation notice named gpt-5.6-terra, which still works, but the newer GPT-6 models are cheaper for most workloads.

Is gpt-4o-mini a safe replacement?

No, even though OpenAI's old gpt-3.5-turbo page still suggests it. gpt-4o-mini-2024-07-18 was deprecated on 14 October 2026 and shuts down on 14 January 2027, and the notice names gpt-6-luna as its replacement. Migrating to it now means migrating again in under three months.

Can I just swap the model string to gpt-6-luna?

Almost. Set reasoning_effort="none" and replace max_tokens with max_completion_tokens. If you raise the effort above none, remove temperature, top_p, top_logprobs and (on Chat Completions) logprobs. Function calling on Chat Completions only works at none.

Can I retrain my GPT-3.5 fine-tune?

Only onto models that are themselves retiring. The GPT-6 models aren't fine-tunable. GPT-4.1, GPT-4.1 mini and GPT-4o mini, along with their fine-tunes, shut down on 14 January 2027. New fine-tuning jobs end for existing customers on 6 January 2027. Test gpt-6-luna with structured outputs against your eval set instead.

Are o4-mini and o3-mini shut down too?

Yes. Both were retired on 23 October 2026 in the same wave as GPT-3.5, along with o1, o1-pro, gpt-4, gpt-4-turbo and gpt-4.1-nano. o3 and o3-pro follow on 11 December 2026.

Sources

• OpenAI: Deprecations
• OpenAI: API pricing
• OpenAI: gpt-3.5-turbo model page
• OpenAI: gpt-6-luna model page
• OpenAI: gpt-6.1-sol model page
• OpenAI: Models overview
• OpenAI: Using the latest model (migration guide)
• OpenAI: Reasoning models guide
• OpenAI: Rate limits guide

If a deprecation migration like this is sitting in your backlog with nobody free to take it, Codersera can help you hire vetted remote developers who have shipped LLM integrations and can take it end to end.