# GPT-OSS-120B pricing, a clean two-to-one ladder > GPT-OSS-120B pricing on Groq is $0.15 per million input and $0.60 output, exactly double the 20B at every rate. So the only question is quality. URL: https://getculpa.com/openai-gpt-oss-120b-pricing Last reviewed: 2026-08-01 Rates effective: 2026-07-02 ## Answer GPT-OSS-120B pricing on Groq is $0.15 per million input tokens and $0.60 per million output tokens, exactly double the 20B model at input, output and cached rates alike. That leaves no pricing subtlety to argue about, only whether the larger model earns twice the money. Culpa, a local-first LLM cost, margin, and forecast ledger, prices both against the same workload so the comparison is measured. ## Why this happens Most model comparisons are muddied by uneven price shapes. One model is cheaper on input and dearer on output, another has a cached rate the first lacks, and the answer changes with your token mix. This pair removes all of it. Every rate on the 120B is exactly twice the 20B, so the cost ratio is 2.0 for any workload whatsoever, whatever your input and output split. The choice reduces to a quality question, and quality questions get answered with an evaluation rather than with a spreadsheet. ## What this usually looks like - The larger model was chosen once, during a quality push, and never re-tested. - Nobody has run the smaller model against current production traffic. - The decision is defended on general reputation rather than on your own task. - Both models are in use and nobody can say which requests go where or why. ## Common mistakes - Defaulting to the larger open-weight model for everything. Why it hurts: You pay exactly double on all traffic to fix the fraction of it the smaller model handles badly. Do instead: Route by difficulty. Send the easy majority to the 20B and escalate what fails a check. - Comparing the two on a benchmark rather than on your own data. Why it hurts: A public benchmark measures somebody else's task, and the gap on yours is what the money buys. Do instead: Run both over a sample of your real traffic and score the outputs you actually care about. - Ignoring the cached rate on both. Why it hurts: Each carries a cached rate at half its input price, which is a real saving on a repeated prefix. Do instead: Put the stable block at the front of the prompt and confirm cached tokens are actually billing. ## Self-check - Price last month's workload on both models. The answer is exactly 2.0 either way, which is the point. - Sample 200 real requests and score both models against them. That gap is what you're buying. - Check what share of your traffic genuinely needs the larger model. - Test a routing rule that escalates only on a failed check, and price the mix. - Confirm what share of your input bills at the cached rate on whichever model you keep. ## One workload, two models, a ratio that never moves (illustrative) A workload of 30 million input and 8 million output tokens in a month, priced on both models at real Groq rates. Volumes are modelled, and the ratio below holds at any volume. GPT-OSS-120B: (30 x $0.15) + (8 x $0.60) = $4.50 + $4.80 = $9.30 GPT-OSS-20B: (30 x $0.075) + (8 x $0.30) = $2.25 + $2.40 = $4.65 Ratio = 2.00, and it stays 2.00 for any input and output split you choose At ten times the volume: $93.00 against $46.50, still exactly 2.00 Routing 70% of traffic to the 20B: (0.3 x $9.30) + (0.7 x $4.65) = $2.79 + $3.26 = $6.05, a 35% cut There's no token mix that makes the larger model relatively cheaper, so no amount of prompt tuning changes the decision. Only routing does, and a 70% split to the smaller model takes 35% off. ## Cost figures Every figure carries its confidence and its source. No figure on this site is provider-reported. - $0.15 / $0.60 per million — GPT-OSS-120B input and output rates on Groq [calculated] Source: Price book row for groq/openai/gpt-oss-120b, $0.00015 and $0.0006 per 1k, effective 2026-07-02, re-verified against Groq's rendered pricing page 2026-08-01. - 2.00x — cost of the 120B against the 20B, identical at input, output and cached rates [calculated] Source: $0.15 divided by $0.075 input and $0.60 divided by $0.30 output, from both price-book rows. ## FAQ Q: How much does GPT-OSS-120B cost on Groq? A: $0.15 per million input tokens and $0.60 per million output, with a cached input rate at half the input price. Re-verified against Groq's pricing page on 2026-08-01. Q: Is GPT-OSS-120B worth twice the 20B? A: The pricing gives no clue, because it's exactly double at every rate for any token mix. That makes it purely a quality question on your own task, which is best answered by scoring both over a sample of real traffic. Q: Can I use both models together? A: Yes, and it's usually the cheapest arrangement. Send the easy majority to the 20B, apply a cheap validity check, and escalate only the failures. Routing 70% of traffic down takes about a third off the total. ## Sources - Groq pricing: https://groq.com/pricing Run the free Cost Leak Scan: https://app.getculpa.com/scan?source=pseo&slug=openai-gpt-oss-120b-pricing&cluster=model_pricing Machine-readable index of every guide: https://getculpa.com/api/pages Human-readable index of every guide: https://getculpa.com/guides Site overview: https://app.getculpa.com/llms.txt Privacy: Culpa runs on your infrastructure. Your prompts and responses never leave it. Culpa counts calls to run your plan, and it fails open, so if it ever breaks your app keeps running.