API model string
gpt-3.5-turboContext window
16K tokens (gpt-3.5-turbo-16k) / 4K tokens (base)
Max output 4K tokens
- Knowledge cutoff
- September 2021 (varies by snapshot)
- Released
- 2022-11
- Modalities
- text in; text out
Last verified July 10, 2026
Rate limits by tier
GPT-3.5 Turbo is a delisted legacy model — no official per-tier limits are published as of July 10, 2026. Historical limits were higher than GPT-4-class (legacy community data: ~3,500 RPM / 90K TPM at Tier 1), but current allocation for this legacy model is unknown. All figures must be verified in your OpenAI dashboard.
| Tier | Requirements | RPM | TPM | RPD | Concurrent | Notes |
|---|---|---|---|---|---|---|
| Free | No ongoing free production tier | not published | not published | not published | not published | No free production access; model is legacy/delisted. |
| Tier 1 | ~$5 cumulative spend | not published (legacy; verify in dashboard) | not published | not published | not published | Historical Tier 1 for GPT-3.5 Turbo was ~3,500 RPM / 90K TPM (pre-2026 third-party, verify); current allocation for legacy model unknown. |
| Tier 2–4 | $50–$250 cumulative spend | not published | not published | not published | not published | Verify in dashboard; legacy model may have non-standard limits. |
| Tier 5 | ~$1,000 cumulative + 30 days | not published | not published | not published | not published | Manual limit-increase request via Settings → Limits → Request increase; 3–10 business-day response. |
| Enterprise | Contact sales | not published | not published | not published | not published | not published |
Swipe the table sideways to see every limit column.
- 1.GPT-3.5 Turbo is delisted from OpenAI's current pricing table — no official per-tier limits published as of July 10, 2026.
- 2.Legacy Tier 1 from pre-2026 community data: ~3,500 RPM / 90K TPM (per third-party trackers — verify; may no longer apply).
- 3.Self-serve fine-tuning for GPT-3.5 Turbo blocked for new organizations since May 7, 2026; existing organizations until January 6, 2027.
- 4.OpenAI recommended replacements: GPT-4.1 nano ($0.10/$0.40/MTok, 1M context) or GPT-5.4 nano ($0.20/$1.25/MTok) — both cheaper AND more capable.
- 5.GPT-3.5 Turbo pricing is not published (delisted) as of July 10, 2026.
- 6.Historical third-party data (pre-delisting, not current first-party): approximately $0.50 input / $1.50 output per MTok — verify, not current pricing.
- 7.Batch API eligibility for this legacy model: not confirmed; verify on current OpenAI pricing page.
- 8.Recommended migration — GPT-4.1 nano: $0.10/$0.40/MTok, 1M context. GPT-5.4 nano: $0.20/$1.25/MTok.
Limits verified against the OpenAI docs, July 10, 2026.
gpt-3.5-turbo vs the alternatives
GPT-3.5 Turbo (delisted) compared to its current successors GPT-5.5 and legacy GPT-4o.
| Aspect | gpt-3.5-turbo | GPT-5 (GPT-5.5) | GPT-4o |
|---|---|---|---|
| Input price | not published (delisted; hist ~$0.50) | $5.00/MTok | $2.50/MTok (Azure ref) |
| Output price | not published (delisted; hist ~$1.50) | $30.00/MTok | $10.00/MTok (Azure ref) |
| Context window | 16K tokens | ~1M (922K) | 128K |
| Knowledge cutoff | Sep 2021 | Dec 2025 | Oct 2023 |
| Model status | legacy/delisted | ga current | legacy |
| Multimodal | text only | text+image | text+image |
| Fine-tuning | blocked new orgs (May 7, 2026) | being wound down | being wound down |
| Replacement pricing | not published | GPT-5.4 nano: $0.20/$1.25 | GPT-4.1 nano: $0.10/$0.40 |
Swipe the table sideways to see every model.
Hitting a 429? The playbook
The exact errors you'll see
429 Too Many Requests{"error": {"message": "Rate limit reached for gpt-3.5-turbo in organization org-xxx on requests per min (RPM): Limit 3500, Used 3500, Requested 1.", "type": "requests", "code": "rate_limit_exceeded"}}HTTP header Retry-Afterx-ratelimit-limit-requestsx-ratelimit-remaining-requestsx-ratelimit-reset-requestsx-ratelimit-limit-tokensx-ratelimit-remaining-tokensx-ratelimit-reset-tokensWhy it happens & how to fix it
RPM burst — historically high ceiling but may now be reduced for legacy model
Exponential backoff; implement request queue; verify current limit in dashboard — legacy allocation may differ from historical 3,500 RPM.
TPM exceeded on legacy 16K context (tight token budget per request)
Keep system prompts short (under 1K tokens); chunk user input; set conservative max_tokens — 16K window means every wasted token costs more relative to window size.
Shared org key hit by multiple services
Assign separate API keys per product; monitor per-key consumption in the dashboard.
Fine-tuned GPT-3.5 Turbo jobs blocked for new organizations since May 7, 2026
Migrate fine-tuned models to a supported model before January 6, 2027 for existing orgs.
Deprecated fine-tuned checkpoint no longer accessible
Use only GA base model string (gpt-3.5-turbo); confirm checkpoint availability in your dashboard.
Retry strategy
Honor the Retry-After header on every 429 response. Implement exponential backoff with jitter starting at 1 second, max 60 seconds. GPT-3.5 Turbo's historically high RPM ceiling means 429s often resolve in under 20 seconds — short backoff intervals are usually sufficient. Use rolling-window awareness: OpenAI windows roll every 60 seconds, not on a fixed clock.
1import OpenAI, { RateLimitError } from "openai";23const client = new OpenAI({ apiKey: process.env.OPENAI_API_KEY });45async function chatWithRetry(6 messages: OpenAI.Chat.ChatCompletionMessageParam[],7 maxRetries = 68): Promise<string> {9 let attempt = 0;10 while (attempt <= maxRetries) {11 try {12 const response = await client.chat.completions.create({13 model: "gpt-3.5-turbo", // migrate to "gpt-4.1-nano" or "gpt-5.4-nano" — same API shape14 messages,15 max_tokens: 512,16 });17 return response.choices[0].message.content ?? "";18 } catch (err) {19 if (err instanceof RateLimitError) {20 const retryAfter = Number(err.headers?.["retry-after"] ?? 0);21 const jitter = Math.random() * 1000;22 // GPT-3.5 Turbo: start with shorter backoff (historically high RPM)23 const backoff = retryAfter > 024 ? retryAfter * 100025 : Math.min(500 * 2 ** attempt, 30_000) + jitter;26 console.warn(`429 — retrying in ${(backoff / 1000).toFixed(1)}s (attempt ${attempt + 1}/${maxRetries})`);27 await new Promise(r => setTimeout(r, backoff));28 attempt++;29 } else {30 throw err;31 }32 }33 }34 throw new Error("Max retries reached");35}How to raise your limits
The ladder from the starter tier to enterprise — what each rung takes, and what it unlocks.
Legacy Tier 1
Automatic~$5 cumulative spend
Unlocks: Historical high RPM for GPT-3.5 Turbo (per third-party trackers — verify in dashboard; current legacy allocation may differ)
Tier 2–4
Automatic$50–$250 cumulative spend
Unlocks: Higher limits; check dashboard for current legacy-model allocations
Tier 5
Automatic; manual limit-increase request thereafter (3–10 business-day response)$1,000 cumulative + 30 days; then Settings → Limits → Request increase
Unlocks: Manual limit-increase form; 3–10 business-day response
Best scaling move — Migrate to GPT-4.1 nano
ImmediateChange model parameter to gpt-4.1-nano — same OpenAI API shape
Unlocks: $0.10/$0.40/MTok, 1M context, current-gen capacity allocation, no legacy constraints
Migrate to GPT-5.4 nano
ImmediateChange model parameter to gpt-5.4-nano
Unlocks: $0.20/$1.25/MTok, larger context, higher capability than GPT-3.5 Turbo for similar cost
Cut your token spend
Migrate to GPT-4.1 nano
80% cheaper on input ($0.10 vs ~$0.50/MTok historical), 73% cheaper on output, 1M context vs 16KChange model: "gpt-3.5-turbo" to model: "gpt-4.1-nano" — identical API request/response shape. Migration is the single highest-impact optimization available.
Keep system prompts under 1K tokens
16K context means every wasted system prompt token is proportionally expensiveCompress instructions; remove filler language; use bullet points. Every 1K tokens saved extends effective context by 6.25% of the total window.
Set max_tokens strictly to needed output length
Prevents verbose responses that waste tokens against the 4K output capProfile your actual average output length; set max_tokens to that value plus 20% buffer — avoid the default uncapped setting.
Batch non-realtime work
50% off if Batch API eligibility is confirmed for legacy modelVerify GPT-3.5 Turbo Batch eligibility on the current OpenAI pricing page (legacy models may be excluded). Use for offline classification or summarization.
Remove fine-tuned GPT-3.5 Turbo dependencies
Avoids hard deadline failure on January 6, 2027 for existing org fine-tuned modelsAudit all fine-tuned checkpoints; plan migration to a supported model; test replacements with GPT-4.1 nano which matches or exceeds GPT-3.5 Turbo capability.
Build a model-string abstraction
Allows instant swap to a supported model with one config changeStore the model string in an environment variable (MODEL_ID=gpt-3.5-turbo) rather than hardcoding; flip to gpt-4.1-nano or gpt-5.4-nano when OpenAI announces a sunset date.
Use rolling-window awareness for retry timing
GPT-3.5 Turbo's historically high RPM means 429s often resolve in under 20 secondsStart with a 500ms initial backoff (vs the standard 1s for GPT-4-class) since the high-RPM window usually clears within a few seconds.
Frequently asked questions
Is GPT-3.5 Turbo still available?
The gpt-3.5-turbo API endpoint remains callable as of July 2026, but the model is delisted from OpenAI's current pricing table. No official shutdown date has been announced. Fine-tuning is blocked for new organizations since May 7, 2026; existing orgs have until January 6, 2027.
What are the GPT-3.5 Turbo rate limits?
OpenAI does not publish per-tier rate limits for GPT-3.5 Turbo as of July 10, 2026. Pre-2026 community data showed approximately 3,500 RPM / 90,000 TPM at Tier 1 (per third-party trackers — verify; may no longer apply). Check your actual limits in the OpenAI dashboard.
What is the best replacement for GPT-3.5 Turbo?
GPT-4.1 nano ($0.10/$0.40/MTok, 1M context) is OpenAI's recommended replacement — approximately 80% cheaper on input than historical GPT-3.5 Turbo pricing and far more capable. GPT-5.4 nano ($0.20/$1.25/MTok) is the current-gen alternative. Both use the same API shape, so migration is a single model string change.
How do I fix GPT-3.5 Turbo 429 rate limit errors?
Check the Retry-After header and wait that many seconds. GPT-3.5 Turbo historically had a high RPM ceiling, so 429s often resolve in under 20 seconds — start with a 500ms backoff. Implement exponential backoff with jitter. Long-term: migrate to GPT-4.1 nano for current-gen capacity allocation.
When does GPT-3.5 Turbo fine-tuning end?
Self-serve fine-tuning was blocked for new organizations as of May 7, 2026. Existing organizations have until January 6, 2027 to complete fine-tuning. No deadline has been announced for the base chat API endpoint.
Is GPT-3.5 Turbo free?
No. There is no ongoing free production tier for GPT-3.5 Turbo. A first payment of roughly $5 is required to reach Tier 1. Current pricing is not published (model is delisted).
GPT-3.5 Turbo vs GPT-4o: which is better?
Both are legacy models as of July 2026, but GPT-4o is significantly more capable with 128K context and multimodal support (text+image), while GPT-3.5 Turbo is text-only with a 16K context cap. Neither is recommended for new projects — use GPT-4.1 nano or GPT-5.5 instead.
Can RapidDev help migrate our GPT-3.5 Turbo integration?
Yes. RapidDev engineers handle GPT-3.5 Turbo to GPT-4.1 nano or GPT-5.x migrations including fine-tuning migration planning before the January 2027 deadline. Book a free scoping call at rapidevelopers.com/contact.
We build AI apps that don't hit rate limits
- Retry, backoff & caching built in
- Multi-provider fallback routing
- Fixed price, you own the code
30-min call. No commitment.
