Fireworks AI Pricing 2026: Inference API Cost vs GPU Rental

Fireworks AI prices its serverless API by the token. It also rents you the underlying GPU directly, on-demand, by the hour, if you want to skip the API layer entirely. Put those two numbers next to each other and something jumps out: Fireworks charges $7.00/hr for an on-demand H100, while Spheron's live H100 PCIe on-demand rate sits at $2.01/hr, about 3.5x cheaper for the same GPU class (Fireworks AI pricing, Spheron GPU offers API, fetched 17 Jul 2026). That gap changes the math on every crossover calculation in this post. The real alternative to Fireworks' per-token bill is renting a GPU somewhere that isn't marked up 3.5x.

This is a pricing breakdown, not a features roundup. If you want the fuller field of ten Fireworks competitors ranked on catalog breadth, fine-tune portability, and GPU control, we already covered that in Fireworks AI alternatives. Here, we're doing the token math: every serverless tier, the two discount programs that change what you actually pay, fine-tuning rates, and the exact point where renting your own hardware beats the API bill.

Fireworks AI Pricing by Model Size: Serverless Token Rates

Fireworks doesn't have one price. It has a tier system based on parameter count, plus a separate list of flagship-model rates for models big enough to get individual pricing. Knowing which bucket your model falls into is the first step to estimating a real bill.

Size-Tier Pricing (Sub-4B, 4B-16B, 16B+, MoE Tiers)

For any model not individually listed, Fireworks bills by parameter count band:

Source: Fireworks AI serverless pricing docs. The MoE split matters because a lot of the open-weight models teams actually deploy, Mixtral-class and DeepSeek-class architectures, are MoE, and lumping them into the 16B+ dense rate would overstate the bill. A 56B-total-parameter MoE model at $0.50/1M is roughly half the price of a 70B dense model at $0.90/1M, even though the MoE model might have comparable active-parameter compute cost per token.

Flagship Model Rates (DeepSeek V4 Pro, Kimi K2.6/K2.7, GLM 5.2, Qwen 3.7 Plus)

Above the tier system, Fireworks prices flagship models individually because they're expensive enough, or popular enough, to warrant their own line item:

Source: Fireworks AI serverless pricing docs, Standard tier rates. Output pricing runs roughly 2x to a bit over 4x input pricing across every model in this table, which is the normal shape for a transformer API: generation is the expensive half of the request. GLM 5.2 carries the highest output rate in this set at $4.40/1M, with Kimi K2.6 and K2.7 Code close behind at $4.00 each, while DeepSeek V4 Flash is priced almost an order of magnitude below its own Pro sibling for teams that don't need the larger model's reasoning depth.

Standard vs Priority Serving: What the ~50% Latency Premium Buys

Every flagship model except Qwen 3.7 Plus also has a Priority tier, priced at roughly 1.25x to 1.5x the Standard rate. DeepSeek V4 Pro Priority runs $2.61 input / $0.218 cached / $5.22 output per 1M tokens, against $1.74 / $0.145 / $3.48 on Standard, a 50% premium across all three token types (Fireworks AI serverless pricing docs). The other flagship models fall in roughly the same range, so budget accordingly before switching a workload from Standard to Priority.

What you're paying for is reserved capacity ahead of shared-tenancy contention, the same logic behind AWS Capacity Blocks or committed-use discounts elsewhere in GPU cloud, just inverted: here you pay more to jump the queue instead of paying less to commit to a term. If your workload has a hard P99 latency SLO and you're on a popular model during peak hours, Priority is the lever Fireworks gives you before you'd otherwise need to move to dedicated hardware. If your latency tolerance is loose, Standard is free money left on the table by not switching.

Cached and Batch Discounts Explained

Fireworks runs two separate discount programs on top of the base rates above, and they compound differently depending on your traffic pattern. Getting the two mixed up is an easy way to underestimate savings on flagship models specifically.

The 50% Batch Inference Discount (Input and Output)

Batch inference on Fireworks bills at a flat 50% of serverless pricing, on both input and output tokens, in exchange for asynchronous processing instead of a synchronous response (Fireworks AI pricing). For a 70B-class model at the standard $0.90/1M rate, batch brings that to $0.45/1M. This is the right lever for anything that doesn't need a live response: nightly re-ranking, offline evaluation runs, bulk classification, dataset labeling. It's a flat discount with no model-specific variance, so the math is simple: whatever you'd pay on Standard, halve it.

Cached Input Isn't a Flat 50% Off: It's 80-92% Off on Flagship Models

Fireworks' own pricing summary describes cache discounts loosely as "reduced rates," and it's tempting to assume that means something close to the 50% batch number. It doesn't, and undershooting this number means underselling your own savings on a repeated-prompt workload. Looking at the actual per-model cached rates against standard input:

Source: calculated from Fireworks AI serverless pricing docs. For any workload with a stable system prompt, a long few-shot template, or a repeated document context, and RAG pipelines are the obvious case, cache hit rate on the prompt prefix is the single biggest lever on your bill, well ahead of picking a cheaper model. A RAG pipeline re-sending the same 4,000-token system prompt and retrieved-document context on every call, with only the final user question changing, could be paying 90%+ less on that fixed prefix than the sticker input rate implies once caching kicks in.

Fireworks Fine-Tuning Pricing (Per-1M-Training-Token Tiers by Model Size)

Fireworks bills fine-tuning per 1M training tokens processed, in four size bands, with rates that scale by both model size and method:

Source: Fireworks AI pricing. Full-parameter methods run 2x the LoRA rate at every tier, and DPO runs 2x SFT within each method, so the two multipliers stack: Full Param DPO on an 80B-300B model, at $24.00/1M training tokens, is 4x the LoRA SFT rate for the same size band. Reinforcement fine-tuning is billed separately, at Fireworks' on-demand GPU rate rather than a per-token training rate, which is worth flagging since it's the one fine-tuning path that inherits the 3.5x GPU markup covered below. We've broken down the full fine-tuning-vs-renting math, including Together AI's tiers alongside Fireworks', in LLM fine-tuning cost 2026: API vs renting your own GPUs, so we won't re-derive it here.

The Crossover Point: When Renting Your Own H100/H200 Beats Per-Token Billing

Every per-token API has a volume where the math flips in favor of a dedicated GPU, because the GPU's cost is fixed per hour while the API bill scales linearly with tokens. The question is where that line sits, and it moves depending on which GPU you're comparing against, not just which model you're serving.

The Math: Fireworks $0.90/1M (70B Tier) vs a Rented H100 at $2.01/hr

Take the 16B+ dense tier at $0.90 per 1M tokens and a Spheron H100 PCIe on-demand instance at $2.01/hr, live rate as of 17 Jul 2026 (Spheron GPU offers API). Running that instance flat-out for 24 hours costs $48.24/day regardless of throughput. At a conservative 400 tok/s baseline for a 70B model on a single H100 PCIe under moderate concurrency, that's 34.56M tokens/day of capacity, worth $31.10/day at Fireworks' rate, still cheaper than the idle-tolerant GPU cost at that throughput ceiling.

The crossover shows up once you push throughput with real serving optimizations. vLLM's continuous batching, PagedAttention, and chunked prefill together push a Llama 3.3 70B FP8 deployment to 2,200-2,400 tok/s at 128+ concurrent requests on H100 SXM5, per our continuous batching and PagedAttention benchmarks. At even a more conservative mid-range 1,200-1,500 tok/s with batching, break-even against Fireworks' $0.90/1M rate lands around 53.6M tokens/day, roughly 10 hours of full GPU utilization. Past that volume, the dedicated GPU is cheaper, and the gap compounds daily since the API bill keeps climbing while the GPU cost stays flat.

Worked Example: Low-Volume Agent vs Sustained RAG Pipeline

Low-volume agent, 2M tokens/day. On Fireworks' 16B+ tier: $0.90 x 2 = $1.80/day, or about $54/month. Running a dedicated H100 PCIe continuously to serve that same load costs $48.24/day, nearly 27x more expensive, because the GPU sits idle almost all day waiting for a trickle of agent calls. This is exactly the workload Fireworks' per-token model is built for, and it isn't close.

Sustained RAG pipeline, 60M tokens/day. That volume needs the batched throughput from the section above, since a naive 400 tok/s deployment tops out at 34.56M tokens/day of capacity on a single H100. At Fireworks' $0.90/1M rate: $54.00/day, or $1,620/month. A Spheron H100 PCIe running vLLM continuous batching at $2.01/hr costs $48.24/day flat, or $1,447.20/month, cheaper even before accounting for the cached-input discount a RAG pipeline's repeated system prompt and document context would actually earn. Apply that cache discount to the Fireworks side, and the gap narrows again, so the real crossover for a caching-heavy RAG workload sits higher than the raw $0.90/1M math implies. For the underlying utilization framework behind sizing this decision across any billing model, see our AI inference cost economics playbook.

The Twist: Fireworks' Own On-Demand GPUs Cost 3.5x More Than Renting Elsewhere

Here's the part that changes the shape of this whole comparison. Fireworks sells more than tokens. It also rents the raw GPU directly: H100 80GB and H200 141GB both at $7.00/hr on-demand, B200 180GB at $10.00/hr, B300 288GB at $12.00/hr, billed per GPU-second with no separate start-up charge (Fireworks AI pricing).

Spheron rates are live on-demand minimums fetched from the Spheron GPU offers API on 17 Jul 2026. Every one of the crossover calculations above assumed you'd be renting the H100 at Spheron's rate, not Fireworks' own dedicated-GPU rate. If you priced the same "rent a GPU instead" decision against Fireworks' own $7.00/hr H100, the crossover point against its own serverless tier would sit dramatically higher, because you'd be paying $168/day just to keep the GPU on, versus $48.24/day for the same GPU class rented through Spheron's marketplace. That's the practical takeaway: don't run the rental side of this math against the API provider's own hardware markup. Current H100 pricing on Spheron moves with live availability, so run your own numbers before committing.

Pricing fluctuates based on GPU availability. The prices above are based on 17 Jul 2026 and may have changed. Check current GPU pricing → for live rates.

Fireworks AI vs Spheron: Side-by-Side

Fireworks rates from Fireworks AI serverless pricing docs and fireworks.ai/pricing. Spheron rates from the live GPU offers API, both fetched 17 Jul 2026. Both platforms update rates with availability, so treat these as a snapshot.

The company behind these numbers has grown fast enough that the pricing pressure is real, not theoretical. By October 2025, Fireworks said it was processing more than 10 trillion tokens per day for over 10,000 enterprise customers, including Cursor, Perplexity, Notion, Uber, DoorDash, Shopify, Samsung, and Vercel (GreyJournal). Revenue kept climbing from there: Fireworks hit $315 million in annualized revenue by February 2026, up 416% year over year, putting it in talks for a $15 billion valuation at the time, up 3.75x from the $4 billion mark set just seven months earlier at its October 2025 Series C (GreyJournal). Those talks closed on 15 Jul 2026, when Fireworks announced a $1.505 billion Series D at a $17.5 billion valuation, led by Atreides Management, Index Ventures, and TCV, with Nvidia among the participants. The company said it had surpassed $1 billion in annualized revenue run rate and was processing more than 40 trillion tokens a day, with over 95% of those tokens coming from models customers had specialized on their own proprietary data (Fireworks AI Series D announcement). That scale is exactly why the per-token tiers and the on-demand GPU markup both matter: Fireworks isn't a small shop absorbing infrastructure cost as a loss leader, it's pricing a mature, billion-dollar-ARR product, and the 3.5x gap on H100 on-demand is a real margin decision, not an oversight.

Which One Should You Use?

Stay on Fireworks' serverless tier if you're under roughly 50-60M tokens/day on a 70B-class model, running a bursty agent or low-volume API where the GPU would otherwise sit idle most of the day. The per-token model is genuinely the right tool there, and the batch and cache discounts make it cheaper still if your traffic pattern fits either one: async workloads should always route through batch, and any pipeline with a repeated prompt prefix should be checking its cache hit rate before it checks its model choice.

Once your volume clears that crossover, or you need to serve a fine-tuned checkpoint, or your latency SLO can't tolerate shared-tenancy variance, the math moves to dedicated hardware. Just don't price that decision against Fireworks' own $7.00/hr H100. Bare-metal H100 instances on Spheron start at $2.01/hr on-demand with per-minute billing and no contracts, and your existing OpenAI-compatible client code works after a base-URL swap to a self-hosted vLLM endpoint. Spheron's docs cover that deployment end to end.

If your Fireworks bill is climbing past the crossover, the GPU-rental side of that comparison shouldn't be priced against Fireworks' own hardware markup.

Get started on Spheron →