OpenRouter vs Fireworks vs Together AI Pricing 2026

These three get compared as if they were interchangeable inference vendors, and they are not the same kind of business. OpenRouter sells routing across other companies' infrastructure. Together and Fireworks sell tokens from their own GPUs, plus dedicated capacity and fine-tuning. That difference decides your price, your failure modes and your data story. All figures on this page were checked against official pricing pages, docs and OpenRouter's public endpoints API on 2026-10-11.

1. What is the actual difference between OpenRouter, Together and Fireworks?

• OpenRouter is an aggregating router. One key reaches hundreds of models across many hosts, at the underlying providers' posted prices with no markup on inference^[1]. It adds routing features no single host offers: fallback across providers, price-sorted (:floor) and throughput-sorted (:nitro) routing, and data-policy filtering^[1][4].
• Together AI is a direct host: a serverless per-token menu of open models, dedicated GPUs billed per hour (H100 $5.49, B200 $8.99), provisioned throughput, GPU clusters and LoRA fine-tuning from $0.34 per million training tokens with a $4 minimum^[5].
• Fireworks AI is a direct host with Standard and Priority serverless modes plus Fast variants on some models, batch at 50% of serverless, on-demand GPUs (H100 $8.00 an hour, B200 $13.00) and LoRA fine-tuning from $0.50 per million training tokens^[7][10].

The frontier closed models live on the router. Anthropic and Google flagships are not on Together's or Fireworks' serverless menus, which serve open-weights models. Through OpenRouter on 2026-10-18, Claude Fable 5.1 listed at $10 / $50 per million tokens and Gemini 4 Argon at $2 / $10, the same as Anthropic's and Google's own rates, in line with OpenRouter's no-markup policy^[1][3]. Not every listing matches: Gemini 3.8 Flash showed $0.825 / $4.125 against Google's $0.75 / $3.75, so check the router's price for the exact model you will call. If your product mixes a closed flagship with cheap open models, the router is the only single-key option of the three.

2. What do the same models cost on each in 2026?

List prices in dollars per million input / output tokens, read on 2026-10-11. The OpenRouter column shows the model maker's own endpoint and the lowest-priced endpoint listed that day, with its quantization:

Model	Fireworks (Standard)	Together	OpenRouter: maker's endpoint	OpenRouter: cheapest listed
DeepSeek V4 Pro (0813)	not listed	$1.32 / $3.96[5]	DeepSeek $0.66 / $1.98[3]	$0.22 / $0.66 (fp8)[3]
Kimi K3	$3.00 / $15.00[7]	$3.00 / $15.00 list[5]	Moonshot $3.00 / $15.00[3]	$0.80 / $13.50 (fp4)[3]
GLM 5.3	$1.40 / $4.40[7]	$1.40 / $4.40[5]	Z.AI $1.40 / $4.40[3]	$0.35 / $1.10 (fp8)[3]
MiniMax M3	$0.30 / $1.20[7]	$0.30 / $1.20[5]	MiniMax $0.30 / $1.20[3]	$0.23 / $0.96 (fp4)[3]
DeepSeek V4.1 Flash	$0.30 / $1.20[7]	$0.30 / $1.20[5]	DeepSeek $0.15 / $0.60[3]	$0.0348 / $0.1392 (fp8)[3]
GPT-OSS-120B	$0.15 / $0.60[7]	$0.15 / $0.60[5]	n/a (open weights)	$0.03 / $0.15 (int4)[3]

Two patterns. Fireworks and Together price shared models identically: Kimi K3, GLM 5.3, MiniMax M3, DeepSeek V4.1 Flash and GPT-OSS-120B all match to the cent^[5][7]. The router's advantage is the long tail of hosts: the model maker's own endpoint is sometimes half the US hosts' price, and the cheapest endpoint is often a quantized deployment^[3]. Together's Kimi K3 promotion ($2.70 / $13.50) ran until October 11, 2026^[5]. Compare your exact model at your token mix with the AI stack cost calculator, and stress-test the dependency with the model price drop stress test.

3. What fees, discounts and limits does each add?

Dimension	OpenRouter	Together	Fireworks
Inference markup	None (passthrough)[1]	Own menu[5]	Own menu[7]
Money-in fees	Card 5.5% ($0.80 min); crypto 5%[1]	$15 minimum first purchase[6]	None published[7]
Discounts	Batch variants on some models[1]	Batch API prices; cached-input rates[5]	Batch 50%; cached-input rates[7]
Speed tiers	Sort by price or throughput[1]	Serverless vs dedicated GPUs[5]	Priority at 1.25× Standard; Fast variants at 1.5×[7]
Rate limits	Free variants: 50 or 1,000 requests/day by credits[2]	No published tiers; may return 429 under load[5]	10 RPM without credits; 6,000 RPM with credits[9]
Fine-tuning	Not offered	LoRA SFT from $0.34/M, $4 min[5]	LoRA SFT $0.50-$10/M by size[10]

The often-misquoted number is OpenRouter's 5.5%: it is a card-processing fee on buying credits, not a markup on token prices^[1]. On a $100 card top-up, $5.50 goes to the fee and the credits then buy tokens at the providers' posted rates. Fireworks' Priority mode costs 25% more than Standard on both input and output, and its Fast variants 50% more, so price the mode you will actually run^[7].

4. Why is everything prepaid now?

• OpenRouter has always run on prepaid credits^[1].
• Together describes itself as fully prepaid: no free trial, a $15 minimum first purchase, and API access suspended at a zero balance unless you hold a Scale or Enterprise contract^[6].
• Fireworks migrated self-serve accounts from postpaid to prepaid on July 1, 2026^[8].

For a solo builder this cuts both ways. A prepaid balance is a hard spending cap: an agent loop gone wrong can only burn what you loaded. The cost is availability risk, because a drained balance stops production traffic. Treat the balance like infrastructure: auto-recharge on, an alert threshold set, and a fallback key on a second provider, which is the failover a router automates.

5. Who handles BYOK, and what happens to your data?

Dimension	OpenRouter	Together	Fireworks
BYOK	Yes: free to $25,000/mo of list-price usage, then 5%[1]	No (custom containers, not keys)[5]	No (custom weights, not keys)[7]
Default retention	Each provider's own policy[4]	Stores prompts unless ZDR is enabled[12]	Zero data retention for open models[11]
Training	Opt out of routing to providers that train[4]	Separate opt-in, off by default[12]	No logging without opt-in[11]
Exception to watch	Free and paid models have separate settings[4]	Passthrough models follow the third party's policy[12]	Response API keeps data 30 days when store=true[11]

"Zero data retention everywhere" is not true by default. If you handle customer data, check three settings: turn on zero data retention in Together's organization privacy settings, set store=false on Fireworks' Response API, and restrict OpenRouter routing to providers whose data policy you accept^[4][11][12].

6. Which layer does a solo builder actually need?

• Prototyping across many models, or mixing closed flagships with open models: OpenRouter. One key, no markup, failover included^[1].
• One open model at real volume: price it on a direct host and on the router. Where the maker's own endpoint is cheaper, routing pays for itself; where prices match, going direct removes a dependency.
• Fine-tuning or custom weights: Together or Fireworks; the router does not train models^[5][10].
• Offline workloads: batch pricing on Fireworks (50% of serverless) or Together's Batch API^[5][7].
• Negotiated direct contracts you want to keep: OpenRouter BYOK, free up to $25,000 a month of list-price usage^[1].

Inference list prices move monthly; re-verify before committing volume. For a same-model comparison that adds Groq, see Groq vs Together vs Fireworks. For the cross-vendor frontier view see the cheapest LLM API ranking, and for margin risk when a vendor reprices, the vendor lock-in switching math.

All pricing figures verified against official pricing pages, docs and OpenRouter's public endpoints API on 2026-10-11.

Frequently asked questions

Is OpenRouter cheaper than Fireworks?

For the same model, OpenRouter is never more expensive than the provider it routes to, because it passes through each provider's own price with no inference markup and earns on a 5.5% card fee when you buy credits. It is often cheaper, because it can route to a lower-priced host. On 2026-10-11, DeepSeek's own endpoint for DeepSeek V4 Pro on OpenRouter listed $0.66 / $1.98 per million tokens against $1.32 / $3.96 at Together. The cheapest OpenRouter endpoints are often 4-bit or 8-bit quantized, so check the quantization column before you route by price alone.

What does prepaid billing mean on Together and Fireworks?

You load a balance before you can run inference, instead of settling an invoice after the month ends. Together's docs describe it as fully prepaid, with a $15 minimum first purchase and API access suspended at a zero balance. Fireworks moved self-serve accounts to prepaid on July 1, 2026. OpenRouter has always run on prepaid credits. The trade is no surprise bill, but a drained balance stops your app, so set auto-recharge and an alert threshold on day one.

Which of OpenRouter, Together and Fireworks supports BYOK?

Only OpenRouter. You attach your own provider keys and OpenRouter routes requests through them. On the pay-as-you-go plan, the first $25,000 a month of list-price usage carries no BYOK fee; above that the fee is 5% of what the same model and provider would cost on OpenRouter. Together and Fireworks are hosts, not routers: their bring-your-own story is custom containers and model weights, not keys.

Do these providers retain or train on your prompts?

The defaults differ. Fireworks has zero data retention by default for open models, except its Response API, which keeps conversation data for 30 days when store=true, the default for that endpoint. Together stores prompts and responses by default and may use them for product improvement unless an organization turns on zero data retention; training on your data is a separate opt-in that is off by default. On OpenRouter, each provider's own policy applies, and you can opt out of routing to providers that train.