Short answer: if you only compare token prices on the same open model, DeepInfra is often lowest, Together and Fireworks often match each other, and the gap is usually small. The bigger differences are the discount mechanics: Fireworks has batch, spot and priority modes; DeepInfra has standard, priority and flex service tiers; Together has serverless, provisioned throughput and GPU clusters. Pick the platform whose discount fits your traffic shape.
Pricing model at a glance
| Provider | Default | Discount / premium modes | Reserved or dedicated | GPU rental |
|---|---|---|---|---|
| Together AI | Serverless per token | Batch API tab; cached input rates; provisioned throughput | Dedicated inference, PTUs, GPU clusters | H100 from $3.99/hr; B200 $9.29/hr |
| Fireworks | Serverless per token | Batch 50% off; Spot 30% off; Priority premium; cached input half price or less | Reserved throughput with SLA; dedicated GPUs | H100 $8/hr; B200 $13/hr |
| DeepInfra | Serverless per token | Priority 1.5×; Flex 0.8×; promotional model discounts | Custom LLMs; dedicated clusters | A100 $1.49/hr; H100 $2.90/hr; B200 $4.89/hr |
Sources: Together AI pricing, Fireworks pricing, DeepInfra pricing, all retrieved 19 Nov 2026. Prices and promotions change.
Representative token prices
USD per 1M input / output tokens. The same model name does not always mean the same version, quantisation or context limit; check each provider’s model page before comparing.
| Model | Together AI | Fireworks | DeepInfra |
|---|---|---|---|
| DeepSeek V4.1 Flash | $0.30 / $1.20 cached $0.006 | $0.30 / $1.20 cached $0.006 | $0.20 / $0.60 promo $0.098 / $0.294 |
| GLM 5.3 | $1.40 / $4.40 cached $0.26 | $1.40 / $4.40 cached $0.26 | $0.90 / $4.00 promo $0.5625 / $2.50 |
| Kimi K3 | $3.00 / $15.00 cached $0.30 | $3.00 / $15.00 cached $0.30 | $2.85 / $14.25 cached $0.285 |
Other models diverge more. DeepInfra lists Llama 3.3 70B at $0.10/$0.32 and Qwen3.8 Flash at $0.113/$0.382. Together lists gpt-oss-120B at $0.15/$0.60 and Qwen3.8 Flash at $0.15/$0.47. Fireworks lists gpt-oss-120B at $0.15/$0.60 and GLM 5.3 Flash at $0.15/$0.50. Always compare the exact model version you will call.
Cache, batch and spot: the real differentiators
Fireworks has the most explicit mode menu. Cached input is billed at half price or less with no write fee. Batch is 50% off input and output. Spot is 30% off Standard on best-effort capacity, and throughput can drop to zero, so it needs a retry or fallback. Priority costs more but schedules ahead of standard traffic. This is a good fit when you can classify traffic into interactive, batch and background lanes.
DeepInfra uses service tiers instead: Standard at 1×, Priority at 1.5× and Flex at 0.8×. Flex is for non-production or asynchronous work that can tolerate slower responses. Promotional model discounts can be large — the DeepSeek V4.1 Flash listing shows 51% off — but promotions expire.
Together AI has a Batch API tab, cached-input rates and provisioned throughput (PTUs). The PTU calculator is the right tool when you have steady traffic and want to convert tokens into reserved capacity. GPU clusters are available when you need dedicated hardware rather than serverless.
GPU rates
These are different products: Together’s GPU cluster on-demand rate, Fireworks’ on-demand GPU rate and DeepInfra’s custom-LLM GPU rate. Do not compare them without checking storage, networking, egress and utilization assumptions.
| GPU | Together AI | Fireworks | DeepInfra |
|---|---|---|---|
| H100 80 GB | $3.99/hr | $8.00/hr | $2.90/hr |
| H200 141 GB | $5.99/hr | $8.00/hr | $3.49/hr |
| B200 180 GB | $9.29/hr | $13.00/hr | $4.89/hr |
| B300 | $11.49/hr | $15.00/hr | $5.89/hr |
If your utilization is low, serverless is usually cheaper than renting a GPU. If utilization is high and steady, the GPU rate matters more than the token rate. Run the break-even on your own tokens per second, not the vendor’s benchmark.
Which one is cheapest?
For the three representative models above, DeepInfra has the lowest list price. Together and Fireworks match each other. But the ranking can flip after discounts:
- Batch-heavy: Fireworks Batch at 50% off can beat a lower sticker price.
- Background or non-urgent: DeepInfra Flex at 0.8× or Fireworks Spot at 30% off can win.
- Interactive, latency-sensitive: Fireworks Priority or DeepInfra Priority may be worth the premium; the cheapest lane may miss your p95.
- Steady high volume: Together PTUs or a dedicated GPU can beat all per-token prices.
Do not compare one provider’s discounted lane with another’s standard price. Use the same mode, the same model version and the same p95 token counts.
When a flat plan wins
A flat plan wins when the monthly per-token bill is above the plan price and the workload fits the plan’s RPM, concurrency and daily caps. Plugsky’s paid plans are $8–$120/month with 39 models and published caps. The LLM API cost calculator shows the break-even; the cheapest LLM API table gives the list-price baseline.
How to compare without a spreadsheet error
monthly_cost =
(requests × input_tokens ÷ 1,000,000) × input_price
+ (requests × output_tokens ÷ 1,000,000) × output_price
+ retry_surcharge
+ cache_write_cost
− batch_or_flex_discount
+ reserved_capacity_or_gpu_cost
Use the same request count and token split for every provider. Count the prompt scaffolding, tool schemas and retrieved context. Then apply the discount to the share of traffic that actually qualifies. A batch discount on 20% of traffic is not a 50% discount on the bill.
FAQ
Which is cheaper: Together AI, Fireworks or DeepInfra?
For the models in this snapshot, DeepInfra often has the lowest list price, especially on DeepSeek, GLM and Kimi. Together AI and Fireworks match each other on several open models. The cheapest sticker price is not always the cheapest bill: cache discounts, batch/spot/flex modes, reserved capacity and GPU rentals change the total.
Do Together AI and Fireworks charge the same for the same model?
Often they do on highlighted open models. For example, both list DeepSeek V4.1 Flash at $0.30 input and $1.20 output per 1M tokens, and GLM 5.3 at $1.40 / $4.40. DeepInfra lists the same models lower, but prices and promotions change.
What is the difference between serverless and reserved capacity?
Serverless bills per token and scales with traffic. Reserved capacity bills for guaranteed throughput or dedicated GPUs, usually with an SLA, and is better for steady high utilization. If you reserve and do not use it, you still pay.
When does a flat-rate API beat all three?
When your monthly per-token bill is above the flat plan price and the workload fits the plan’s RPM, concurrency and daily caps. A flat plan removes the per-token variance; it is not automatically cheaper at low volume.
Price your workload across all three
The calculator turns requests, tokens, cache and batch into a monthly bill, then shows when a flat plan wins.
Open the calculator → See plans