gpt-oss-120b on Groq vs Together vs Fireworks: Same Weights, Same Price, a 4x Latency Gap

By Promptster Team · 2026-11-08

If you run gpt-oss-120b, which host should you call? Groq, Together and Fireworks all serve the same open weights at the same list price, $0.15 per million input tokens and $0.60 per million output. We sent each one the same 20 graded calls.

On our four harder tasks, the bill was almost the same on all three hosts, between $0.0069 and $0.0076 for 20 calls. The difference was latency: the median call took 1.0 s on Groq, 1.6 s on Fireworks and 4.0 s on Together. Groq and Fireworks passed 20/20. Together passed 19/20; it dropped a prefix from one JSON value. Cerebras would have been the fourth host, but it could not be tested: our key was still unfunded.

In October we compared Kimi K3 on Together and Fireworks. This post looks at a cheaper model with an extra host, on a harder battery.

How we tested

Four script-graded tasks, with no LLM judge. Each task has a trap:

We ran each task five times per host (n=5) with temperature 0.7 and max_tokens 4000, on 2026-10-23, through Promptster's own provider handlers. Costs are cost_usd at list rates from shared/pricing.ts. Latency is wall-clock time from one machine. It is not a host SLA, and five runs per task is a small sample.

The results

Groq Fireworks Together Cerebras
List price (in/out per 1M) $0.15 / $0.60 $0.15 / $0.60 $0.15 / $0.60 $0.35 / $0.75
Passed 20/20 20/20 19/20 blocked (HTTP 402)
Cost, all 20 calls $0.0069 $0.0076 $0.0072 —
Median latency, all 20 calls 1.0 s 1.6 s 4.0 s —
Median: json-order / stock-floor / taglines / launch-notes 1.0 / 2.7 / 0.9 / 1.1 s 1.4 / 3.8 / 1.5 / 1.5 s 2.9 / 9.5 / 5.2 / 3.1 s —
Fastest / slowest single call 0.82 s / 3.4 s 1.1 s / 6.4 s 2.1 s / 9.9 s —
Output tokens, all 20 calls 10,586 11,714 11,038 —

Groq had the fastest median on every task, and every Groq call outside stock-floor finished in under 1.2 s. Together's slowest calls were all on stock-floor, between 7.6 s and 9.9 s. Groq answered the same prompt in 2.2 to 3.4 s.

No host returned an error, timed out or hit a rate limit. Every call finished with finish_reason: stop.

Why the costs differ

All three hosts charge the same rates, so the only thing that can change the bill is the number of tokens. Input was almost the same everywhere: between 3,595 and 3,655 prompt tokens across the 20 calls. The small differences come from each host's chat template. Output varied more. Fireworks produced 11,714 output tokens, Together 11,038 and Groq 10,586, and most of those were reasoning tokens (9,665, 9,252 and 8,605). That spread puts Fireworks a few hundredths of a cent above Groq for the 20 calls.

A spread that size comes from five samples of a model that reasons at temperature 0.7, not from the host. On Groq, the stock-floor runs alone ranged from 913 to 1,484 output tokens. At these rates, choose a host for speed and reliability, not price.

Together reported cached prompt tokens on all 20 calls, and Fireworks on 16. Groq reported none. cost_usd prices cached tokens at the full input rate, because shared/pricing.ts lists no cached rate for gpt-oss-120b on any of the three hosts. We did not check whether either host discounts them, and input is a small share of these bills either way.

The one miss: Together dropped the SKU prefix

The json-order task expects "skus":["SKU-A07","SKU-B12"]. Together's second run returned this:

{"po":null,"skus":["A07","B12"],"units":4,"total":168}

The JSON is valid. The PO is correctly null, the cancelled line is gone, and both the units and the discounted total are right. Only the prefix is missing. The other four Together runs, and all ten json-order runs on Groq and Fireworks, kept it.

With n=5 this is one response, so we don't read it as a pattern. Tasks like this one are exactly what the battery is for, though. A schema validator that checks only that skus is an array of strings would have accepted this response. A downstream lookup on A07 would then have failed. If you parse identifiers out of model output, check their format as well as their type.

Cerebras: still the missing fourth host

shared/pricing.ts lists Cerebras' gpt-oss-120b at $0.35/$0.75, more than double the other three hosts' input rate. On the first call it returned HTTP 402:

{"message":"Payment required to access this resource. Visit your billing tab.","type":"payment_required_error",...}

Our harness skipped the other 19 calls instead of billing a dead key. That is the second batch in a row with Cerebras blocked. We have no Cerebras latency or cost to report, and we have not estimated them.

How this compares with October

In October's battery, gpt-oss-120b on Groq passed 15/15 with a 0.9 s median. Those tasks were easier and used a lower max_tokens (1500). On the harder tasks Groq still passed everything and stayed under a second and a half on three of the four tasks. Only the ten-step calculation took longer.

The Kimi K3 post found Together faster than Fireworks. Here, on gpt-oss-120b, it was the other way round, and by a wide margin: Fireworks had the faster median on every task. One host's speed on one model tells you little about its speed on another, so measure the model you plan to ship. Earlier host comparisons pointed the same way: Together vs Fireworks in August and the Llama host gap in June.

What we'd do

Pick the host on latency. At $0.15/$0.60 the price is a tie, and accuracy differed by one call in 60. For an interactive feature, Groq's 1.0 s median and 3.4 s worst case is the strongest result here. Fireworks is a close second. Together is usable for batch work, where its slower median doesn't matter.

Validate the content of identifiers, not just their type. A regex such as ^SKU-[A-Z]\d{2}$ would have caught the one miss.

Measure again before you commit. Host rankings have changed between our batches and between models. A rerun costs less than a cent.

For the wider open-weight picture, see the original open-source vs closed-source benchmark.

Run this comparison yourself — no account, no API key

To run the same open-weight model on several hosts and compare latency on your own prompts, use Promptster.


Run 2026-10-23 through Promptster's provider handlers: openai/gpt-oss-120b on Groq and Together, accounts/fireworks/models/gpt-oss-120b on Fireworks, gpt-oss-120b on Cerebras (HTTP 402 on the first call, the remaining 19 skipped). Four script-graded tasks, five runs each, temperature 0.7, max_tokens 4000. Costs are cost_usd from list rates in shared/pricing.ts, with no cache discount. Latency is wall-clock from one machine. Raw responses and usage objects for every call are in content/blog/data/q4b-2026-battery/.