NotCheap MAKE EVERY CREDIT COUNT
Analyze Your Costs

Cloudflare, Groq and Together: A Serverless LLM Budget

Share this guide

Token pricing looks refreshingly simple until you push a billion tokens through production and the invoice shows up. On our standard GPT-OSS 120B workload, Groq and Together both land at $240 of model usage per billion total tokens, while Cloudflare Workers AI works out to $430 before its daily allocation and any Workers plan fee. The cheapest advertised number isn't the whole story, so start with models you can actually deploy self-serve today. Leaning on old pricing for legacy Llama checkpoints is a reliable way to get an unpleasant surprise at month-end.

I checked the official sources on 25 October 2026: Cloudflare Workers AI, Groq's supported models and Together's inference pricing. All figures are in USD before tax. Every calculation assumes standard, uncached inference, with no dedicated instance reservations and no separately metered tooling calls.

Compare a current common model

Groq's active public catalog now lists Llama 3.1 8B and Llama 3.3 70B under Enterprise / Contact Sales tiers. I didn't want to quote old promotional pricing that new self-serve accounts can't get, so I compared the widely supported GPT-OSS 120B open-weights model across all three providers.

Provider / current model identifierInput per 1MOutput per 1M
Groq: openai/gpt-oss-120b$0.15$0.60
Together: gpt-oss-120B catalog offer$0.15$0.60
Cloudflare: @cf/openai/gpt-oss-120b$0.350$0.750

Running the same model gives us a fair financial baseline. What to keep in mind: identical model names across vendors don't guarantee identical hardware kernels, KV-cache quantization, time-to-first-token, or downstream accuracy. Run sample generations on your real production prompts before you rework your architecture around a price sheet.

A fixed 80/20 workload at two volumes

Most enterprise retrieval-augmented generation (RAG) and document-agent workloads lean heavily on input context, so I modeled an 80% input / 20% output split. That means 80M input + 20M output at the 100M token mark, and 800M input + 200M output at 1B tokens.

Model usage only100M total tokens1B total tokens
Groq GPT-OSS 120B$24.00$240.00
Together GPT-OSS 120B$24.00$240.00
Cloudflare GPT-OSS 120B, token equivalent$43.00$430.00

The math is simple: input millions × input rate + output millions × output rate. For Groq and Together at 1B tokens, 800 × $0.15 + 200 × $0.60 equals $240.00. For Cloudflare, 800 × $0.350 + 200 × $0.750 equals $430.00. These numbers are pure metered inference, with no pre-funded wallet minimums included.

Your workload mix can move the gap quite a bit. In generation-heavy flows like code generation or synthetic data, where output roughly equals input (a 50/50 split), 1B tokens cost $375.00 on Groq and Together ($75 input + $300 output), compared to $550.00 on Cloudflare ($175 input + $375 output). Also budget for reasoning and chain-of-thought tokens, because hidden reasoning traces count as billable output tokens on every major provider.

Cloudflare's daily allocation is not a monthly coupon

Cloudflare meters Workers AI in its own unit, Neurons, at $0.011 per 1,000 Neurons, and it publishes token-equivalent conversions per model. Free accounts get 10,000 Neurons per day, resetting at 00:00 UTC. Once you go past the free limits, you need a paid Workers subscription, which starts from $5/month. If your stack already pays for Workers, you don't need to count that base fee twice.

Over a 30-day billing cycle with active traffic every day, that 10,000-Neuron allowance is worth at most 30 × $0.11 = $3.30 in total credits. Add the token equivalents and a single $5 plan, and your net monthly bill comes out to approximately $44.70 for 100M tokens or $431.70 for 1B tokens. Treat these as planning figures and reconcile against actual billed Neurons, rather than trusting converted token rates down to the penny.

Burst patterns behave differently. If your 1B-token pipeline runs as a single 24-hour batch job, you only get that one day's $0.11 credit, and your bill lands at roughly $434.89 with the $5 Workers plan. The credits from the other 29 days vanish at midnight UTC, like a gym membership you forgot to use. The exact billing formula is sum of max(0, each day's Neurons − 10,000) × $0.011 / 1,000 + applicable plan fee.

If your app goes idle, inference drops to $0.00, but your active Workers plan fee sticks around. External storage, KV lookups, and Vectorize indexes belong in their own infrastructure budgets.

Model variants and capacity still matter

Model size and quantization are another handy lever. For lightweight tasks, Cloudflare lists Llama 3.1 8B at $0.282 input and $0.827 output, alongside an FP8 Fast variant at just $0.045 input and $0.384 output per million tokens. On our 80/20 1B-token workload, the standard 8B variant totals $391.00, while the FP8 Fast edition drops to $112.80 before daily credits and base fees. These are distinct model weights, so don't assume FP8 quantization matches full precision on complex logic.

Pacing matters for stability too. Processing one billion tokens over 30 days averages roughly 386 tokens/second, or 23,148 tokens/minute. Real traffic rarely arrives that politely, though. Make sure your account's rate limits (TPM and RPM) can absorb your peak bursts. Hitting 1B monthly tokens through smooth, sustained queues rarely requires custom enterprise contracts, but sharp spikes might run into API walls on standard tiers.

My practical rule of thumb: if your stack already lives on Cloudflare, the platform convenience can justify paying more per token. If you're running a standalone, input-heavy pipeline on GPT-OSS 120B, the Groq and Together price gap is hard to ignore at this volume.

If raw latency and specialized LPU throughput are your top priorities, Groq is exceptionally fast and cost-effective. If you want zero-cold-start inference right next to your Workers and KV cache, Workers AI simplifies your architecture, as long as you account for the Neuron conversion. For broad open-weights support and dedicated endpoints, Together AI offers reliable performance with predictable token billing.

PASS IT ON

Useful for your team?

Share this guide with a colleague.