Groq Dropped Llama From the Free Tier on Aug 16 — the Official Quickstart Now Breaks (2026 gpt-oss Migration Guide)

If you pasted Groq’s official quickstart snippet into a project in the last two weeks and it stopped working, you did not mistype anything.

On 2026-08-16, Groq stopped serving llama-3.1-8b-instant and llama-3.3-70b-versatile — the two model IDs that appear in nearly every Groq tutorial — to free and developer tier users. And as of our direct read on 2026-08-27, the official Quickstart page still ships llama-3.3-70b-versatile in its sample code.

This is not a piece about whether Groq is worth using. It is the operational side: what the free tier actually runs now, how to get a key, the one line you have to change, and how to see your remaining quota before you hit the wall.

The short version

• The free tier has no Llama chat models anymore. The deprecation page states the change applies to free and developer tier usage; enterprise customers with a committed-spend contract are unaffected
• Official replacements: llama-3.1-8b-instant → openai/gpt-oss-20b; llama-3.3-70b-versatile → openai/gpt-oss-120b or qwen/qwen3.6-27b
• For most projects the fix is one line: the model ID
• The free tier still requires no credit card, and it still runs on the same LPUs. What was removed is model choice, not speed

1. What the free tier actually runs today

All figures below come from a direct read of the Free plan table at console.groq.com/docs/rate-limits on 2026-08-27. RPM = requests per minute, RPD = requests per day, TPM/TPD = tokens per minute/day.

Three things worth noticing before you pick a model:

1. Not one Llama chat model is on this table. The two llama-prompt-guard entries are safety classifiers, not general chat models.
2. qwen/qwen3.8-27b gets 2,000,000 tokens per day, ten times the rest of its group. If your workload is few requests with long inputs — transcript cleanup, long-document summaries — it is the best value on the free tier.
3. TPM will stop you long before RPD does. A loop with a long system prompt eats 8,000 tokens per minute in a handful of turns, while 1,000 requests a day is comparatively generous.

2. Check whether this affects you (30 seconds)

Search your project for these two strings:

llama-3.3-70b-versatile
llama-3.1-8b-instant

If you get hits and you are on a free key or the developer plan, your requests have been failing since Aug 16. No hits means this change does not touch you — though the quota table above is still worth a look.

One trap: both models are still visible in Groq’s model catalogue, labelled “Enterprise”. Visible in the catalogue is not the same as callable with your key.

3. Get a free key (three steps, no card)

1. Sign up and log in at console.groq.com
2. Go to API Keys (console.groq.com/keys) and create a key
3. Groq recommends storing it in an environment variable rather than hardcoding it:

export GROQ_API_KEY=your-key-here

The key is typically shown in full only once, so save it before closing the dialog.

4. The one-line switch

Here is the official Python quickstart, with the stale model ID it still ships:

from groq import Groq
import os

client = Groq(api_key=os.environ.get("GROQ_API_KEY"))

chat_completion = client.chat.completions.create(
    messages=[{"role": "user", "content": "Your prompt here"}],
    model="llama-3.3-70b-versatile"   # no longer served on the free tier
)

print(chat_completion.choices[0].message.content)

Following the mapping on the deprecation page, change that one line:

    model="openai/gpt-oss-120b"       # was llama-3.3-70b-versatile

And for the small model:

    model="openai/gpt-oss-20b"        # was llama-3.1-8b-instant

Installation is unchanged: pip install groq.

gpt-oss-120b or qwen3.6-27b? Groq lists both as replacements for the 70B model and does not say which is better. Their free-tier limits are identical (30 RPM / 1,000 RPD / 8,000 TPM / 200,000 TPD), so run your own real prompts through both and compare. That beats any benchmark for your specific use.

5. Point your existing OpenAI tooling at Groq

Groq exposes an OpenAI-compatible endpoint, so anything that lets you set a custom base URL — desktop clients, self-hosted front ends, your own code on the OpenAI SDK — can be redirected without an architecture change:

• Base URL: https://api.groq.com/openai/v1
• API key: your Groq key
• Model: an ID from the table above, e.g. openai/gpt-oss-120b

Groq documents several OpenAI features that are not supported. Grep for these before you migrate:

• logprobs, logit_bias, top_logprobs, and the messages[].name field
• The N parameter (if supplied, it must equal 1)
• temperature cannot be 0 — a zero is converted to 1e-8. Groq asks for float32 values greater than 0 and up to 2
• Audio output formats vtt and srt

The temperature=0 case is the quiet one. Plenty of people set it for deterministic output; on Groq it does not error, it just silently becomes something else.

6. Read your quota instead of waiting for a 429

Every response carries these headers, so you never have to track usage yourself:

• x-ratelimit-limit-requests / x-ratelimit-remaining-requests
• x-ratelimit-limit-tokens / x-ratelimit-remaining-tokens
• x-ratelimit-reset-requests / x-ratelimit-reset-tokens
• retry-after

Exceeding a limit returns HTTP 429 Too Many Requests. Groq notes that retry-after is only set when you actually hit a 429 — so the correct pattern is to throttle proactively on x-ratelimit-remaining-*, and fall back to retry-after only once you are already blocked. A fixed sleep is the worst of both worlds.

7. Free-tier gotchas

• Listed is not callable. The Llama entries carry an “Enterprise” label in the catalogue; a free key will not reach them
• Tutorials and sample code go stale. Most Groq guides online — including the official Quickstart — still use Llama model IDs. Check any snippet against the free-tier table above before pasting
• Groq has no GPT-5 and no Claude. It serves open-weight models; the selling point is speed, not the model roster
• Limits are per model. Switching models does not share an RPD pool, but do not treat model-hopping as an unlimited workaround — each has its own ceiling
• Still no credit card required (per our 2026-08-09 verification record); that did not change with this deprecation

What we could not verify (stated plainly)

• Whether cached tokens count against the limits. Our 2026-08-09 record says cache hits do not count, but that sentence was not present on the rate-limits page we read this time. It may be a page revision or a rule change; unverified this round, so we do not present it as current fact.
• Daily token ceiling for the groq/compound family. The official table leaves that column blank.
• Which error code a Llama call returns after Aug 16. The deprecation page says only that requests are no longer served. We have no paid key to compare against, so we do not guess.
• Whether free-tier data is used for training. The official documentation we read does not address this for the free tier specifically.

Sources

Rules and figures were read directly from Groq’s official documentation domain on 2026-08-27:

1. Model Deprecations — the 2026-08-16 shutdown of both Llama models, scope (free and developer tier, enterprise unaffected), and the official replacement mapping
2. Rate Limits — the per-model Free plan RPM/RPD/TPM/TPD table, x-ratelimit-* headers, 429 and retry-after
3. Quickstart — key creation, environment variable, and the Python sample that still names a deprecated model
4. OpenAI Compatibility — base URL https://api.groq.com/openai/v1 and the unsupported-parameter list
5. Models — the current “Enterprise” labelling of both Llama entries

“No credit card required on the free tier” comes from our own existing verification record for Groq (2026-08-09).

Quotas and model catalogues change without notice; the official pages are always the authority. For full plans and pricing see our Groq tool page. For two other routes to running models without a card, see OpenRouter’s free-model daily limits and Google AI Studio’s free tier.