OpenRouter vs Together vs Fireworks 2026 Compared

These three get compared as if they were interchangeable inference vendors, and they are not even the same kind of business. OpenRouter sells routing across other people's infrastructure. Together and Fireworks sell tokens off their own GPUs, plus dedicated capacity and fine-tuning. That difference decides your price, your failure modes, and your data story, so this comparison starts there, then puts verified same-day prices side by side. All figures checked against official pricing pages and docs on 2026-07-12.

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: automatic fallback across providers, price-floor and latency-based routing, and data-policy-aware routing^[4].
• Together AI is a direct host: a serverless per-token menu of open models, dedicated endpoints billed per GPU-hour (H100 $5.49/hr, B200 $8.99/hr), provisioned throughput, GPU clusters, and fine-tuning from $0.48 per million training tokens with a $4 job minimum^[5].
• Fireworks AI is a direct host with a three-speed serverless menu (Standard, plus Fast at roughly twice Standard and Priority at roughly twice input and +50% output), 50% discounts on cached input and batch, on-demand GPUs (H100 $7/hr, B200 $10/hr), and LoRA fine-tuning^[7].

One consequence is easy to miss: 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, Claude Fable 5 lists at its vendor price of $10/$50 per million tokens and Gemini 3.5 Flash at $1.50/$9, consistent with the zero-markup claim^[1]. If your product mixes a closed flagship with cheap open models, the router is the only single-key option among these three.

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

Per-token list prices, $ per million input / output tokens, all read on 2026-07-12. A dash means the platform did not publish a verifiable price for that model that day:

Model	OpenRouter (routed)	Together	Fireworks (Standard)
DeepSeek V4 Pro	—	$1.74 / $3.48[5]	$1.74 / $3.48[7]
MiniMax M3	—	$0.30 / $1.20[5]	$0.30 / $1.20[7]
Kimi K2.7 Code	$0.72 / $3.50[1]	—	$0.95 / $4.00[7]
Qwen 3.7 Plus	$0.32 / $1.28[1]	—	$0.40 / $1.60[7]
DeepSeek V4 Flash	—	—	$0.14 / $0.28[7]
GPT-OSS-120B	—	—	$0.15 / $0.60[7]
Claude Fable 5	$10 / $50 (vendor list)[1]	not hosted	not hosted
Gemini 3.5 Flash	$1.50 / $9 (vendor list)[1]	not hosted	not hosted

Two patterns in the numbers. Where the model's own maker sets the price, hosts converge: DeepSeek V4 Pro and MiniMax M3 cost exactly the same on Together and Fireworks. Where hosts price independently, routing finds the gap: Kimi K2.7 Code routed through OpenRouter listed about 24% below Fireworks' Standard input rate the same day, because the router sends the request to whichever host currently serves it cheapest. Never assume parity; 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 does each layer add?

Fee dimension	OpenRouter	Together	Fireworks
Inference markup	None (passthrough)[1]	Own menu (no separate fee)[5]	Own menu (no separate fee)[7]
Money-in fees	Stripe 5.5% ($0.80 min); crypto 5%[1]	None published[5]	None published[7]
Discounts	1% for opt-in prompt logging[4]	Batch and committed capacity options[5]	Cached input 50%; batch 50%[7]
Speed tiers	Routing sort: price floor or latency[4]	Serverless vs dedicated ($/GPU-hr)[5]	Standard / Fast (~2×) / Priority (~2× in, +50% out)[7]
Rate limits	Free variants: 20 req/min, 50 or 1,000 req/day by lifetime credits; paid models uncapped by OpenRouter[2]	Dynamic per-org, no published tiers; grows with successful traffic[6]	10 req/min with no payment method; up to 6,000 req/min with credits; spend tiers $50 to $50k[7]
Fine-tuning	Not offered[1]	$0.48–$8.00/M tokens, $4 min[5]	LoRA $0.50–$40/M tokens; bring your own model[7]

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, roughly $5 goes to payment processing before a single token is bought, and the credits then buy tokens at the providers' posted rates. For repetitive workloads, Fireworks' 50% cached-input and batch discounts are the largest single cost lever on this table^[7].

4. Why is everything prepaid now?

The 2026 change that affects operations more than price: postpaid self-serve inference is gone across all three.

• OpenRouter has always been prepaid credits^[1].
• Together switched to prepaid billing on April 22, 2026, in the same release that made rate limits dynamic per organization^[6].
• Fireworks migrated self-serve accounts from postpaid to prepaid on July 1, 2026^[8].

For a solo builder this cuts both ways. The upside is a hard spending cap by construction: an agent loop gone wrong can only burn the balance you loaded, a real protection when one bad deploy can loop a $50-per-million-token flagship overnight. The downside is availability risk: a drained balance stops production traffic. Treat the balance like infrastructure: auto-top-up on, alert threshold set, and a fallback key on a second provider, which is exactly the failover a router automates^[4].

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

Dimension	OpenRouter	Together	Fireworks
BYOK	Yes: 5% of equivalent cost, waived first 1M BYOK reqs/mo[3]	No (bring your own container, not keys)[5]	No (bring your own weights, not keys)[7]
Default retention	No prompt logging by default[4]	Zero data retention by default[6]	Zero data retention by default[7]
Exceptions	Opt-in logging (1% discount)[4]	Training use is opt-in[6]	Volatile prompt cache; Response API store=true keeps state 30 days (that endpoint's default)[7]
Extra control	data_collection: deny routes only to non-retaining hosts[4]	VPC / private networking options[6]	Spend-tier account controls[7]

The headline "zero data retention everywhere" is roughly true and the exceptions are where client trust lives. If you handle customer data, the two flags to check in your own integration are OpenRouter's routing policy (set data_collection to deny and accept a smaller provider pool) and Fireworks' Response API default (set store=false unless you want 30-day server-side state)^[4][7].

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, no fine-tuning: price it on both a direct host and the router; where the host's menu matches the routed price, direct removes a dependency, and where routing finds a cheaper host, it pays for itself.
• Fine-tuning or custom weights: Together or Fireworks; the router does not train models^[5][7].
• Repetitive or offline workloads: Fireworks, for the 50% cached-input and batch discounts^[7].
• Negotiated direct contracts you want to keep: OpenRouter BYOK, effectively free below 1M requests/mo^[3].
• Modality note: Fireworks deprecated its audio inference and image generation in June 2026, while OpenRouter added unified image and audio APIs this year; check your roadmap's modalities before committing^[7].

Inference list prices move monthly; re-verify before committing volume. For the cross-vendor frontier view see the cheapest LLM API ranking, and for what happens to your margin when a vendor repricing hits, the vendor lock-in switching math.

All pricing figures verified against official pricing pages, docs, and changelogs as of 2026-07-12.

Frequently asked questions

Is OpenRouter cheaper than going direct to Together or Fireworks?

Often, and never by much in the other direction. OpenRouter states it adds no markup on inference pricing: you pay the underlying provider's rate, and the platform earns on credit purchase fees (5.5% via card with a $0.80 minimum, 5% via crypto), verified 2026-07-12. Because it routes each request across competing hosts, the routed price can undercut a single host's menu: Kimi K2.7 Code routed through OpenRouter listed at $0.72/$3.50 per million tokens against $0.95/$4.00 on Fireworks Standard the same day. Where a model's host sets one price everywhere, like DeepSeek V4 Pro at $1.74/$3.48 on both Together and Fireworks, the router matches rather than beats it.

What does prepaid billing mean on Together and Fireworks in 2026?

You load a balance before you can run inference, instead of settling an invoice after the month ends. Together moved to prepaid billing alongside dynamic rate limits on April 22, 2026; Fireworks migrated self-serve accounts to prepaid on July 1, 2026; OpenRouter has always run on prepaid credits. For a solo builder the trade is straightforward: no surprise end-of-month bill, but a new operational duty to keep the balance funded, because a drained balance stops your app. Set an auto-top-up and an alert threshold on day one.

Which of OpenRouter, Together, and Fireworks supports BYOK?

Only OpenRouter. You attach your own provider API keys (say, a direct Anthropic or OpenAI key) and OpenRouter routes requests through them, charging 5% of what the same usage would have cost in OpenRouter credits, with the fee waived for the first 1M BYOK requests each month, per its docs on 2026-07-12. Together and Fireworks are hosts, not routers: their bring-your-own story is about custom containers and model weights, not keys. If your provider spend already sits on negotiated direct contracts, BYOK through a router is how you keep those rates and still get one API.

Do these providers retain or train on your prompts?

All three lead with a no-retention default, but the fine print differs. OpenRouter does not log prompts by default; opt-in logging earns a 1% discount, and a data_collection: deny setting routes only to providers that retain nothing. Together is zero-data-retention by default with training use strictly opt-in. Fireworks is zero-data-retention by default with two exceptions worth knowing: a volatile prompt cache, and its Response API stores conversation state for 30 days when store=true, which is the default for that endpoint. Verified against each vendor's privacy documentation on 2026-07-12.