Fireworks AI vs Together AI Pricing 2026: Same Models Side by Side, Cached Input, Batch, GPUs, Fine-Tuning and Cost per Conversation

Published 29 October 2026. Every price below was read on 29 October 2026 from Fireworks' pricing page and serverless pricing docs, and from Together AI's pricing page and model catalog, and every price is as of that date.

Fireworks AI and Together AI charge the same list price per million tokens for every open model they both serve, as of 29 October 2026. Kimi K3 costs $3.00 input, $0.30 cached input and $15.00 output on both; GLM 5.3 costs $1.40, $0.26 and $4.40 on both; DeepSeek V4.1 Flash costs $0.30, $0.006 and $1.20 on both; and Muse Spark Open, MiniMax M3, GLM 5.3 Flash and the 2.4-trillion-parameter Qwen3.8 model match line for line on both pricing pages. The one shared model where the per-token price differs is GPT-OSS 120B: both charge $0.15 input and $0.60 output, but Fireworks bills cached input at $0.015 and Together lists no cached rate, so every repeated token costs ten times as much on Together.

So the bill is decided by everything around the token price. Fireworks takes 50% off every serverless model through its Batch API, while Together's batch documentation names one discounted model, Whisper Large v3, and says the rest "run at standard rates". A US-only request costs 1.5 times list on Fireworks and 1.25 times on Together. A dedicated H100 costs $8.00 an hour on Fireworks and $5.49 on Together, and fine-tuning gpt-oss-120B with LoRA costs $6.00 per million training tokens on Fireworks against $2.50 on Together. Fireworks gives a $1 starter credit; Together has no free trial and a $15 minimum first purchase. In the worked example below, a 40-turn conversation on GPT-OSS 120B costs $0.018 on Fireworks with a warm cache and $0.080 on Together, and on every other shared model the two cost the same.

Most pages that rank for "fireworks vs together pricing" say Fireworks is 10% to 20% cheaper. Those figures come from Llama 2, Llama 4 and Mixtral rows priced in 2025, and the third-party price trackers mix US-only rows with standard ones; one shows Fireworks' Kimi K3 at $4.50 input and $22.50 output, which matches Fireworks' US-only and Fast rates, not its $3.00 standard price. On the models both providers sell today, neither one undercuts the other on list price.

Fireworks vs Together price per million tokens, same model side by side

US dollars per million tokens, standard serverless tier, as of 29 October 2026. Sources: Fireworks serverless pricing and model pages; Together pricing page and chat model catalog. Fireworks' Qwen 3.8 Max model page links to the Qwen/Qwen3.8-2.4T-A95B weights, which is the model Together sells under that name.

Two rows carry a caveat. Together's pricing page lists Qwen3.8 2.4T-A95B at $0.25 cached, and its documentation catalog lists the same model at $0.50 cached, so check the console before you size a cache-heavy workload on it. MiniMax M3 appears on Together's pricing page but not in its documentation catalog, which lists MiniMax M3.1 Flash ($0.15, $0.015 and $0.60) instead.

The same pattern holds beyond these two. GPT-OSS 120B at $0.15 and $0.60 is also what Groq lists for the same model, where cached input costs $0.075, and our Kimi K3 pricing guide sets the Together and Fireworks rates for K3 against Moonshot's own platform. Hosts of the same open weights converge on one list price and compete on caching, batch, tiers and hardware instead.

Models only one of them serves

Fireworks alone lists a per-model price for Ember-1 ($3.00, $0.30 and $15.00), NVIDIA Nemotron 3.5 Lightning 30B A3B ($0.05, $0.01 and $0.20), NVIDIA Nemotron 3 Ultra in preview ($0.60, $0.12 and $2.40) and the Rune 26B-A4B decision model, which bills input only at $0.042. For any other text model you deploy, Fireworks prices by size: $0.10 per million tokens under 4B parameters, $0.20 from 4B to 16B, $0.90 above 16B, and $0.50 or $1.20 for mixture-of-experts models up to 56B or up to 176B, with the same rate for input and output.

Together alone lists a per-model price for DeepSeek V4 Pro 0813 ($1.32, $0.13 and $3.96, with a 1,048,576-token context), MiniMax M3.1 Flash, NVIDIA Nemotron 3.5 Super VL ($0.15, $0.02 and $0.50), Reflection's Beam 1.0 ($0.40, $0.04 and $1.20), Muse Glimmer 30B ($0.35, $0.04 and $1.50), several proprietary Qwen models, and two models at no charge: Ternary Bonsai 27B, free on input and output, and Tev1 4B Experimental, free on output. Together's pricing page and docs disagree on Qwen3.7 Max ($1.50 input and $4.50 output on the page, $2.50 and $7.50 in the docs), so treat that row as unsettled.

Fireworks also sells embeddings at $0.008 per million tokens for models up to 150M parameters, $0.016 up to 350M, and $0.10 for Qwen3 8B.

Cached input: where the per-token bills actually differ

Both providers cache prompt prefixes automatically and charge nothing to write to the cache, and both only match an exact prefix, so a timestamp or user ID at the top of the system prompt breaks the cache for everything after it. The differences are in the rate and in where the cache lives.

Fireworks bills cached tokens at a model-specific rate and, where a model lists none, at 50% of input, according to its serverless overview. On the shared models the listed rates run from 2% of input on DeepSeek V4.1 Flash to 20% on GLM 5.3 Flash and MiniMax M3. The cache is replica-local, so Fireworks asks you to send a stable x-session-affinity header or user field per conversation to land repeat requests on the same replica, and its prompt caching guide says serverless deployments "maintain separate caches for each Fireworks account". Entries usually last "at least several minutes" and up to several hours.

Together bills cached tokens only on models that have a value in the cached column of its catalog, and GPT-OSS 120B, Qwen3.8 Flash, Qwen3.7 Plus and the other models without one bill every input token at full rate. Its prompt caching page says entries "typically last about an hour" in a cache "shared across the fleet", offers a prompt_cache_key field to route related requests together, and notes that cached requests "still count toward your rate limits".

For a chat product on GPT-OSS 120B, this one line is the whole comparison. Every turn after the first resends the full history, so most of the input on a long conversation is cached input, and $0.015 against $0.15 is a tenfold gap on most of the tokens you send.

Batch API: 50% off on Fireworks, mostly full price on Together

Fireworks' Batch API runs at "50% off" serverless per-token prices, for any base model that supports on-demand deployment, and the discount stacks with caching: Fireworks says batch jobs use prompt caching "for additional 50% cost savings on cached tokens". GPT-OSS 120B in a batch costs $0.075 input, $0.0075 cached and $0.30 output. Input files can reach 80 GiB, you choose an expiry of 12, 24, 48 or 72 hours, and completed rows are billed and saved if the job expires.

Together's batch story depends on which page you read. Its batch product page advertises "up to half the cost of our real-time API for most serverless models", and its batch documentation, read the same day, lists exactly one discounted model, openai/whisper-large-v3, and says "Models not listed run at standard rates." Together's pricing page has a "Batch API price" toggle, and on 29 October its page source carried no batch price for any chat model. Plan a Together batch job at full serverless price until the console shows a discount for your model. The job still buys you a separate rate-limit pool, up to 50,000 requests per file, a 24-hour target window, and no charge for failed requests.

Priority, Fast, reserved and provisioned throughput

Fireworks sells three serverless modes, described on its serverless modes page. Standard is the list price. Priority, set with service_tier: "priority", is "less likely to be load shed" and costs 20% to 50% more depending on the model: GPT-OSS 120B rises to $0.18, $0.018 and $0.72, Kimi K3 and GLM 5.3 rise by 25%, and Qwen 3.8 Max, MiniMax M3 and Muse Spark Open rise by 50%. Fast, a separate model ID aiming at 100 or more output tokens per second, costs $4.50 and $22.50 on Kimi K3 and $2.10 and $6.60 on GLM 5.3. Above that, Reserved Throughput is bought as a dollar-per-minute amount covered by an SLA, use-it-or-lose-it each minute, with overage billed at list.

Together has no priced priority tier. Its rate limits page calls serverless performance "best-effort" and points teams that need guarantees to provisioned throughput, sold in throughput units through sales, with an estimator on the pricing page and no published unit price.

US-only inference: 1.5x on Fireworks, 1.25x on Together

Input / cached / output per million tokens, as of 29 October 2026, from Fireworks' serverless pricing and Together's US-only table. Together serves these through us-api.together.ai.

For a team that must keep inference in the United States, Together is 16.7% cheaper than Fireworks on every line of these four models, because 1.25 is five-sixths of 1.5. Muse Spark Open is the exception that favours Fireworks, since it runs only on US capacity there at the standard price. Fireworks also charges a 1.5x premium for region-restricted on-demand GPU deployments.

What a 40-turn conversation costs on Fireworks and Together

Take a chat assistant with a 2,000-token system prompt, 150-token user messages and 350-token replies, the same assumptions as our Groq API pricing, DeepSeek API pricing and Kimi K3 guides. Turn 1 sends 2,150 input tokens and turn 40 sends 21,650, because each request carries every exchange before it, so 40 turns consume 476,000 input tokens and 14,000 output tokens, and 100 turns consume 2,690,000 and 35,000. "Cache warm" means that from turn 2, the previous request's prompt is read at the cached rate and the 500 new tokens bill at full input price. Reasoning tokens are left out.

Our arithmetic on the rates in the tables above, as of 29 October 2026.

At 10,000 forty-turn conversations a month on GPT-OSS 120B, the bill is $185 on Fireworks with a warm cache and $798 on Together, and the same model on Groq, at $0.075 cached, comes to about $457. On Kimi K3 the same volume costs $4,113 on either provider with caching, and $5,141 against $6,169 if it has to stay in the United States. At the cheap end, GLM 5.3 Flash ($0.024) and DeepSeek V4.1 Flash ($0.026) come close to GPT-OSS 120B on Fireworks ($0.018) per cached conversation, which makes GLM 5.3 Flash the cheapest shared model on Together for long chats.

The table also shows the shape of the problem. Going from 40 to 100 turns multiplies turns by 2.5 and the uncached bill by between 5.2 and 5.4 on every row, because every turn carries every turn before it. Caching cuts the price of the repeated tokens; it does not stop them being sent, and a cache that lasts an hour does nothing for a user who comes back tomorrow. The other lever is to stop resending the whole transcript. We build Maximem Synap, a memory layer that returns a ranked set of facts about the user, trimmed to a token budget you set, in place of the full history. When that beats caching, and when it does not, is worked through in how to reduce LLM token costs in long conversations.

Dedicated GPUs: Together is cheaper per hour

Price per GPU per hour, as of 29 October 2026, from Fireworks' pricing page, Together's pricing page and its dedicated inference pricing docs.

The like-for-like comparison is the first two columns, since both are managed inference endpoints on single-tenant GPUs. An H100 running all month (720 hours) costs $5,760 on Fireworks and $3,952.80 on Together, and a B200 costs $9,360 against $6,472.80, so Together is 31% cheaper on both. Together's GPU clusters are raw compute for your own serving stack: H100s at $3.99 on demand, $1.99 preemptible, and $3.69 to $3.19 on 3-day to 180-day reservations.

Billing granularity is close. Fireworks bills per GPU second "with no extra charges for start-up times", scales a deployment to zero after an hour without traffic by default, and gives a new account a quota of 16 GPUs each of H100, H200, B200 and B300. Together bills per minute per replica, only while a replica is ready, so provisioning and cold starts are free, and a deployment scaled to zero costs nothing.

When a dedicated GPU beats serverless depends on how busy it stays. Together's own guidance is that dedicated inference is "usually cheaper when a replica would stay busy most of the day" and serverless is cheaper when traffic is "low or bursty". In spend terms, a Fireworks H100 month equals about 311,000 cached 40-turn GPT-OSS 120B conversations on Fireworks serverless, and a Together H100 month equals about 49,500 uncached ones on Together serverless; the throughput one H100 sustains on your model is the number to measure before you switch.

Fine-tuning prices: Together is cheaper per training token

Price per million training tokens, as of 29 October 2026. Both count training tokens as dataset tokens times epochs; Together adds validation tokens, per its fine-tuning pricing docs.

A LoRA SFT run over a 10-million-token dataset for 3 epochs (30 million training tokens) costs $75 on gpt-oss-120B at Together and $180 at Fireworks, $31.50 against $90 on a 27B model, and $10.20 against $15.00 on an 8B model. The order flips at the top. Fireworks never charges more than $10 per million for LoRA SFT, while Together charges $15 for Kimi K2.6 and $40 for GLM-5.1 and GLM-5.2, so for those models Fireworks is cheaper if its fine-tuning catalog supports them. Together refunds failed jobs in full, charges only completed steps on cancelled ones, and sets minimum charges from $4.00 to $60.00 per job depending on the model.

Two short-lived offers change the arithmetic this quarter. Fireworks says Muse Glimmer 30B SFT is free through its UI and Serverless Training API "until 11/30", and Together's post-training B200s are $9.00 an hour against a list $11.00 until 14 December 2026. Fireworks also sells per-token RL and LoRA training through a Serverless Training API, priced per million tokens prefilled, sampled and trained (Qwen 3.8 27B: $1.86 prefill, $0.372 cached prefill, $5.595 sample, $4.103 train).

Serving the result is where the training bill stops being the main cost. Fireworks' pricing page says it will "serve fine-tuned models for the same price as base models", and its LoRA deployment guide says trained LoRA models "can only be deployed to on-demand (dedicated) deployments", so in practice you pay GPU hours. Together also serves fine-tuned models on dedicated endpoints billed per minute. One H100 for a month is $5,760 on Fireworks and $3,952.80 on Together, which is larger than most training runs above; Fireworks' multi-LoRA option, which shares one deployment across many adapters, is the way to keep that to one GPU bill.

Free credits, minimums and how billing works

Fireworks gives new accounts a $1 starter credit, according to its billing FAQ. It is prepaid: you buy credits, usage draws them down, and an account without a payment method is suspended when the dollar runs out and limited to 10 requests per minute meanwhile. With a payment method and credits the account-wide cap is 6,000 requests per minute. Spend tiers step up at $50, $500 and $5,000 of spend or credit purchases, and the higher tiers raise serverless rate-limit ceilings. A monthly spend limit warns at 80% and pauses every request at 100%, and Fireworks warns that recorded spend "can lag live traffic by a day or more".

Together "does not currently offer free trials", according to its credits page. The first purchase must be at least $15 and later ones at least $5, the platform is "fully prepaid", API access stops at a zero balance, and purchased credits do not expire. Auto-recharge works only with a card as the default payment method. The free models, Ternary Bonsai 27B and the output side of Tev1 4B Experimental, are the only way to call Together without paying, and they still need the account to exist.

Rate limits on serverless

Fireworks publishes its limits. They are adaptive, per account and per model, measured in tokens per minute on three meters, and grow by 25% at most once every 15 minutes while you use more than half of the current limit, according to its rate limits page. The ceilings depend on model size: models under 800B parameters, which include GPT-OSS 120B, GLM 5.3 and DeepSeek V4.1 Flash, top out at 216 million total prompt tokens, 54 million uncached prompt tokens and 2.16 million generated tokens per minute, and Kimi K3 and Qwen 3.8 Max, in the largest tier, at 21.6 million, 5.4 million and 216,000. The full ceilings need spend Tier 3.

Together publishes no numbers. Its rate limits page says most users "can expect to use serverless inference without encountering rate limits", that it may limit requests when demand is high with a 429 or 503, and that it may apply custom limits to a model in high demand. For capacity planning, that leaves provisioned throughput or a dedicated endpoint.

Which is cheaper for your workload

For a long-running chat product on GPT-OSS 120B, Fireworks is cheaper by about four times on a warm cache, because it is the only one of the two with a cached rate on that model. On every other shared model the token bill is identical, and the choice turns on the items below.

For offline work such as evaluations, back-fills, labelling and synthetic data, Fireworks is cheaper by half through its Batch API, with caching stacked on top, unless Together's console shows a batch discount for your model.

For inference that must stay in the United States, Together is cheaper by 16.7% on Kimi K3, GLM 5.3, GLM 5.3 Flash and DeepSeek V4.1 Flash, and Fireworks is cheaper on Muse Spark Open.

For dedicated GPUs and for fine-tuning most models, Together is cheaper: 31% less per H100 and B200 hour for managed inference, and roughly 30% to 65% less per LoRA training token. Fireworks is cheaper for fine-tuning Kimi K2.6 and the GLM-5.x models, and it is the one that sells priced Priority and Fast tiers and SLA-backed reserved throughput without a sales call.

For a first prototype, Fireworks costs nothing to start with its $1 credit, and Together costs $15 up front. A $15 difference should not decide a platform. Run your own evaluation traffic on both before committing, since the list prices match and neither pricing page tells you the latency your prompts will see.

Frequently asked questions

How much does Fireworks AI cost per million tokens?

As of 29 October 2026, Fireworks charges $0.15 input, $0.015 cached and $0.60 output for GPT-OSS 120B, $0.30, $0.006 and $1.20 for DeepSeek V4.1 Flash, $1.40, $0.26 and $4.40 for GLM 5.3, and $3.00, $0.30 and $15.00 for Kimi K3 on the standard serverless tier. Batch requests cost 50% less, and Priority costs 20% to 50% more.

How much does Together AI cost per million tokens?

As of 29 October 2026, Together charges $0.15 input and $0.60 output for GPT-OSS 120B with no cached rate, $0.30, $0.006 and $1.20 for DeepSeek V4.1 Flash, $1.40, $0.26 and $4.40 for GLM 5.3, and $3.00, $0.30 and $15.00 for Kimi K3. US-only requests through us-api.together.ai cost 1.25 times those rates.

Is Fireworks or Together AI cheaper?

On list price per token they are the same for every model both serve, except GPT-OSS 120B, where Fireworks charges $0.015 for cached input and Together has no cached rate. Fireworks is cheaper for batch jobs and for long chats on GPT-OSS 120B; Together is cheaper for US-only inference, dedicated GPUs and most fine-tuning.

What is the Fireworks AI pricing model?

Fireworks bills serverless inference per token on three meters (input, cached input and output) from prepaid credits, on-demand deployments per GPU second, and fine-tuning per million training tokens or per GPU hour for dedicated training. Batch runs at half the serverless price.

What is the Together AI pricing model?

Together bills serverless inference per token, dedicated model inference per minute per running replica, GPU clusters per GPU hour, and fine-tuning per million tokens processed with a minimum charge per job. Everything draws on prepaid credits, with a $15 minimum first purchase.

Does Together AI have a free tier or free credits?

Together says it "does not currently offer free trials" and requires a $15 minimum first credit purchase. Two models are free to call once you have an account: Ternary Bonsai 27B and, on output, Tev1 4B Experimental.

Does Fireworks AI give free credits?

Fireworks gives new accounts a $1 starter credit. Without a payment method the account is limited to 10 requests per minute and is suspended when the credit runs out.

Does Together AI have a batch discount?

Its batch product page advertises up to half off for most serverless models, but its batch documentation on 29 October 2026 lists only Whisper Large v3 as discounted and says other models run at standard rates. Fireworks takes 50% off every batch-eligible model.

How much is an H100 on Fireworks or Together?

An H100 on-demand deployment costs $8.00 an hour on Fireworks. On Together, an H100 for dedicated model inference costs $5.49 an hour, and an H100 in a GPU cluster costs $3.99 on demand or $1.99 preemptible.

How much does fine-tuning cost on Fireworks and Together?

LoRA SFT costs $0.50 per million training tokens for models up to 16B on Fireworks and $0.34 for an 8B or 9B model on Together; for gpt-oss-120B it costs $6.00 on Fireworks and $2.50 on Together, with a $6 minimum. Serving the fine-tuned model needs a dedicated GPU on both.

Sources

All retrieved 29 October 2026.

• Fireworks: pricing, serverless pricing, serverless overview, serverless modes, Batch API, prompt caching, account quotas, serverless rate limits, reserved throughput, on-demand deployments, deploying LoRAs, credits FAQ, billing FAQ, and model pages for GPT-OSS 120B, Kimi K3, Qwen 3.8 Max and GLM 5.3
• Together AI: pricing, batch inference, serverless models, serverless overview, rate limits, prompt caching, batch overview, dedicated inference pricing, fine-tuning pricing, fine-tuned model deployment and credits
• Groq: GPT-OSS 120B model page

Worked examples are our own arithmetic on those rates.