CloudAPIcredits

Groq API Review 2026 for Developers Who Need Speed on a Budget

Fast LPU inference with a no-card free plan and 50% batch pricing. Independent review by the CloudAPIcredits editorial team, updated November 2, 2026.

4.1/ 5
Editor rating

Our verdict

Groq is the right pick for latency-sensitive apps built on open models and the wrong one for anyone who needs closed frontier models or startup credits. We rate it 4.1 out of 5. The Free plan is real: no card, 200,000 tokens a day on gpt-oss-120b, and it resets every morning, which is enough to ship a prototype. Developer pricing of $0.15 and $0.60 per million tokens with 50% off batch and cached input keeps it cheap once you upgrade. The catalog is the weak spot. Llama has moved to contact sales, and the startup credits page has simply disappeared.

Groq credits and programs (4) Visit Groq ↗

What we like

  • Free plan needs no card and resets every day
  • gpt-oss-120b at $0.15 input and $0.60 output per million
  • Batch jobs priced 50% below synchronous calls
  • Prompt caching halves cached input with no extra fee
  • Spend limits with alerts at 50%, 75% and 90%

What to watch

  • No signup credit, trial balance or promo codes
  • Llama 3.3 70B and 3.1 8B now listed as contact sales
  • Free plan cap of 8,000 tokens per minute is tight
  • Startup and partner program pages redirect to the homepage

What Groq actually sells you

Groq is a Mountain View chip company founded in 2016 that sells inference on its own Language Processing Units through GroqCloud, an OpenAI-compatible API launched in 2024. Groq now describes itself as a neocloud for fast inference, points to a $350 million Series A, and positions its LPX hardware to work alongside NVIDIA GPUs. The product you buy is tokens per second on a curated list of open models.

The catalog is deliberately small. Production models are gpt-oss-120b, gpt-oss-20b, Llama 3.3 70B, Llama 3.1 8B and two Whisper models, with qwen3.8-27b, Orpheus text to speech, Prompt Guard and a few enterprise-only models in preview. If your stack needs Claude or GPT-5, Groq is a second provider for the open-model slice, not a replacement.

What free gets you and what it doesn’t

There are no credits. The three tiers are Free at $0 for building and testing with community support, Developer as pay per token with higher limits, chat support, Flex, batch, spend limits and 7-day audit logs, and Enterprise with custom capacity, dedicated support, LoRA inference, SSO and SCIM. You start on Free with an API key and no payment method.

The Free plan is defined by its rate limits. gpt-oss-120b, gpt-oss-20b and qwen3.8-27b each allow 30 requests per minute, 1,000 requests per day, 8,000 tokens per minute and 200,000 tokens per day. Whisper allows 20 requests per minute and 7,200 audio seconds per hour. The limits view under Settings is the exact table for your organization, and select workloads get exceptions. In practice 8,000 tokens per minute is the wall you hit.

Promotional credits exist but aren’t self-serve. Groq offers them occasionally at hackathons and special events. Official pages for a $10,000 Groq for Startups program and a Groq Partner Program still turn up in search, but both URLs redirect to the homepage, so neither is a program a founder can apply to today.

How much room the paid plan gives you

Upgrading to Developer multiplies the ceilings. gpt-oss-120b goes to 1,000 requests per minute, 500,000 requests per day and 250,000 tokens per minute with no daily token cap, and whisper-large-v3-turbo to 400 requests per minute with 4 million audio seconds per day. Prompt Guard models get 100 requests per minute and 30,000 tokens per minute. That’s roughly a 30x jump in tokens per minute for adding a card you won’t be charged on until you’ve used $1.

Two further tiers sit on top. Flex processing, available to paid accounts by setting service_tier to flex, provides roughly 10x the on-demand limits at identical per-token prices but fails fast with a 498 capacity_exceeded error when capacity is short. The Performance tier is Enterprise only, sold as provisioned input and output capacity with a 99.9% availability SLA and a 99% latency guarantee, and requires uncached context under 8,192 tokens.

What you pay once you upgrade

gpt-oss-120b costs $0.15 per million input tokens and $0.60 per million output tokens, gpt-oss-20b and gpt-oss-safeguard-20b cost $0.075 and $0.30, and qwen3.8-27b costs $0.80 and $4.00. Whisper large v3 is $0.111 per audio hour and v3 turbo is $0.04. Orpheus English text to speech is $22 per million characters and the Arabic model is $40. Llama 3.3 70B and 3.1 8B are listed as contact sales.

Batch processing is 50% cheaper than synchronous calls with a 24 hour to 7 day completion window, and prompt caching takes 50% off cached input on the GPT-OSS models with no cache-write fee. The two discounts don’t stack. Batch files are limited to 50,000 lines and 200 MB, results are kept for 30 days, and batch limits sit outside your per-model table.

Billing is in arrears. Nothing is charged at upgrade, invoices fire when lifetime usage crosses $1, $10, $100, $500 and $1,000, then monthly, and totals under $0.50 aren’t billed at all. Spend limits under Settings, Billing, Limits cap the whole organization and block new requests with a 400 blocked_api_access error, with a 10 to 15 minute tracking lag. Downgrading to Free triggers a final invoice.

Where the startup credits went

Groq once advertised $10,000 in inference credits for hand-selected startups and VC portfolio companies, expiring 90 days after award, and a $1,000 credit for members of the Vercel Open Source Program. The Groq for Startups page, its application page and the Partner Program page all redirect to the Groq homepage now, and no replacement or closure notice has been published. If a VC partner mentions Groq credits, ask them for a current contact, because the public door is shut.

There’s no education program, no student tier and no research credits on groq.com. The only academic footprint is hardware access at Argonne’s ALCF AI Testbed for open-science researchers, which isn’t an API credit. Students get the same Free plan as everyone else, which is enough for coursework if prompts stay short.

Which models earn their place

The two GPT-OSS models carry most workloads: 131,072 token context, prompt caching and the lowest prices in the catalog. gpt-oss-safeguard-20b, released October 29, 2025, is a policy-following classifier priced like the 20B base. Qwen3.8-27b is in preview and priced well above gpt-oss-120b. MiniMax M2.5 and Qwen3-VL 32B have been restricted to Enterprise customers since April 18, 2026.

Whisper is the quiet bargain. At $0.04 per audio hour, whisper-large-v3-turbo transcribes a 60-minute podcast for 4 cents, and the batch endpoint accepts audio transcriptions and translations at the same 50% discount. Text to speech through Orpheus is new and still rate-limited to 100 requests a day on the Free plan.

What can go wrong on Groq

Free accounts get community support; Developer adds chat support and Enterprise adds a dedicated team. The main operational risk is capacity, not billing. On-demand requests can queue at peak times, Flex requests fail outright with a 498, and the Performance tier that guarantees latency is contract only. Build retries with jittered backoff from day one.

Billing risks are small because spend is capped by your own limit and invoices are tied to usage already consumed. Payment can fail and be retried twice before suspension. The bigger product risk is the catalog: models move between preview, production and contact sales without notice, as Llama did, so pin model IDs and watch the changelog.

Who should build on Groq

Anyone whose application is built on open models and treats speed as a feature. The Free plan is enough to ship a prototype without a card, and the Developer plan costs $0.15 and $0.60 per million tokens on gpt-oss-120b with structural discounts of 50% for batch and cached input. Not anyone who was counting on startup credits or needs Llama at a published price. Both have moved behind contact sales or vanished from the site.

Best for

  • Prototypes that need a real API with no card on file
  • Latency-sensitive chat and agent loops on gpt-oss-120b
  • Transcription pipelines at $0.04 per audio hour
  • Batch classification and evaluation jobs that can wait a day

Not ideal for

  • Founders expecting startup credits
  • Apps that need Claude, GPT-5 or other closed models
  • Workloads that need Llama 70B at a published price

Groq FAQ

do i need a credit card for the groq api
No for the Free plan. You create an account, generate a key on the API Keys page and start calling models within the Free limits. A credit card, US bank account or SEPA debit account is required only when you upgrade to the Developer plan under Settings, Billing, and the upgrade itself triggers no charge.
what happens when i hit groq rate limits
The API returns a 429 and the request isn’t billed. Per-minute limits reset within the minute and per-day limits reset daily. Cached input tokens don’t count toward limits, batch jobs have their own separate limits, and the Developer plan lifts gpt-oss-120b to 1,000 requests and 250,000 tokens per minute with no daily token cap.
can i set a spending limit on groq
Yes on paid accounts. Under Settings, Billing, Limits click Add Limit and enter a monthly USD amount, then add Usage Alerts at thresholds such as 50%, 75% and 90% and click Save. When the cap is reached every key in the organization returns a 400 blocked_api_access error until you raise the limit or the month resets. Spend tracking lags 10 to 15 minutes.
how do i cancel or downgrade groq developer plan
Downgrade at any time from Settings under Billing. Groq issues a final invoice for any unbilled usage, which must be paid before the downgrade completes, and your account returns to Free plan limits. You lose Developer features such as Flex, batch and spend limits, but you can upgrade again later.
is groq batch api available on the free plan
No. Batch processing is a Developer plan feature, so you need a payment method on file. Once upgraded, batch jobs cost 50% less than synchronous calls, run within a 24 hour to 7 day window you choose, and only completed requests are charged. Batch limits are separate from the per-model rate limit table.
does groq train on my api data
Groq’s docs don’t describe training on customer prompts, and the Developer and Enterprise plans include audit logs with 7-day and 90-day retention respectively. For a contractual answer check the Terms of Use and Privacy Policy linked from the console footer, and ask the enterprise team for a data processing agreement if you handle regulated data.

How we review: ratings are written before we look at commission rates, and every fee and policy statement is checked against the platform's own help pages. Read our editorial policy.