Fireworks AI vs Together AI (2026): Managed Platform or Neocloud?
Fireworks AI is a managed platform that serves and fine-tunes open-weight models; Together AI is a neocloud that also rents raw NVIDIA GPU clusters. Fireworks reports more than 40 trillion tokens served a day, while Together lists 200+ models. Their shared serverless models carry identical per-token prices as of November 2026.
The short version
Price will not decide this one: on 15 November 2026 Fireworks and Together list identical per-token rates on all eight shared headline models. Pick Fireworks to serve and tune a specialized model without running hardware. Pick Together if you need GPU clusters yourself, from $1.99 an hour preemptible, or many modalities on one bill.
Not sure which one fits you? Describe your situation and Smart Match names three tools.Find my AI tools
In this guide15 sections
On 15 November 2026, Fireworks AI and Together AI list the same per-token price, to the cent, on all eight headline open models they both serve. In August the match held on one model. Now it holds across the shared catalog.
That settles the question most teams open with. For a startup deciding where to route open-model inference for the next year, the price list no longer separates Fireworks from Together. What separates them is the kind of company each one is: a managed training-and-inference platform on one side, a GPU cloud that also sells tokens on the other.
Hokai comparison of what separates Fireworks AI and Together AI, from both vendors' pricing and docs pages as read on 15 Nov 2026.
Who should pick which: the verdict
Pick Fireworks if the job is serving, and increasingly tuning, one or two specialized models at high volume without ever touching a GPU cluster. Pick Together if the job needs raw GPU capacity you run yourself, for training from a checkpoint or for workloads no managed endpoint covers, or if you want chat, image, video, audio and rerank on one invoice.
That split is not a hedge. A four-person team shipping one specialized support model to production, the kind of workload Fireworks' own case studies describe, has no reason to pay for a cluster. A team that wants SSH-level control of H100s for a week of training is paying for exactly the business Together raised $800 million to expand.
- Fireworks AI
- managed serving and training, no bare GPU clusters
- Together AI
- per-token API plus rented GPU clusters
- Shared models
- eight, at identical per-token prices
- Written uptime SLA
- Together only, on Provisioned Throughput
What changed since 7 November 2026
Both vendors' pricing pages were read again on 15 November 2026. The verdict holds; Together's newest GPUs got dearer.
- Together raised its Blackwell rates: a B200 now costs $9.29 an hour on demand and a B300 $11.49. H100 and H200 rates did not move.
- MiniMax M3 left Together's serverless list, which now carries MiniMax M3.1 Flash instead, so M3 is Fireworks-only.
- Together's reservations now top out at 91 days or more, its deepest discount.
On 7 November, Kimi K3.1 Pro, DeepSeek V4.1 Flash 1017 and Muse Spark Open joined the shared list, Qwen 3.8 Max became Fireworks-only and Fireworks added a Spot tier at 30% off. The 17 October review recorded the bigger shift. In August the price match held on a single model. Since then Fireworks has raised its H100 rate to $8.00 an hour on its pricing page and begun selling training, while Together moved fine-tuning to per-model prices and added preemptible GPUs.
Is Fireworks AI a neocloud?
No, not in the usual sense. The term describes GPU-as-a-service providers: Cisco's explainer says these companies "are built from the ground up for GPU-as-a-service" and put raw hardware access ahead of a broad service catalog. Fireworks sells something different. Its product is a model served or trained on hardware it runs for you, billed per token, per training token, or per GPU-hour of a managed deployment.
Is Together AI a neocloud?
Yes, by the usual definition. Its pricing page, read on 7 November 2026, rents bare NVIDIA GPU clusters by the hour, from $3.99 for an H100 on demand, next to its per-token API. TechCrunch's July coverage of its Series C described it as "an AI neocloud that rents out Nvidia GPU clusters." Together itself prefers the label "AI Native Cloud" on its homepage.
| Fireworks AI | Together AI | |
|---|---|---|
| What you rent | A served or trained model | A model, or the GPUs themselves |
| Bare GPU clusters | Not offered | H100 from $3.99/hr on demand |
| Usual label | Inference platform | Neocloud, per TechCrunch |
| Sells by the token | Yes | Yes |
The distinction matters for a buyer because it predicts what each company will build next. A neocloud's margin comes from keeping GPUs busy. A managed platform's margin comes from making each token cheaper to produce than the rate it charges. Fireworks says its $1.505 billion Series D will help it "expand our compute infrastructure", so the line is not fixed, but on 7 November its pricing page still lists no bare-cluster rental: its GPU rates buy managed deployments and training time.
Pure GPU clouds such as CoreWeave and Crusoe sit at the far end of that spectrum. They rent capacity and leave the serving stack to you.
The price that stopped being the differentiator
Fireworks' serverless pricing overview as captured on 23 Aug 2026. The Standard and Priority layout is unchanged; the model rows have since moved, so use the table below for current rates.
Fireworks' serverless pricing docs and Together's pricing page, both read on 15 November 2026, list these standard per-million-token rates for the models they share:
| Model | Input | Cached input | Output |
|---|---|---|---|
| Kimi K3 | $3.00 | $0.30 | $15.00 |
| Kimi K3.1 Pro | $2.00 | $0.20 | $10.00 |
| GLM-5.3 | $1.40 | $0.26 | $4.40 |
| GLM-5.3 Flash | $0.15 | $0.03 | $0.50 |
| Muse Spark Open | $1.25 | $0.15 | $4.25 |
| DeepSeek V4.1 Flash | $0.30 | $0.006 | $1.20 |
| DeepSeek V4.1 Flash 1017 | $0.30 | $0.006 | $1.20 |
| gpt-oss-120B | $0.15 | Fireworks only | $0.60 |
Every row is identical on both pages. That is not two vendors copying each other by accident. It is the market rate for hosting each model, and neither company is trying to win on it. The labs behind those rows, Moonshot AI and Z.ai among them, ship the weights; both platforms resell the serving. MiniMax's MiniMax M3 is now the exception: Fireworks still serves it at $0.30 in and $1.20 out, and Together dropped it from its serverless list by 15 November.
Models only one of them lists
The price gaps that remain sit in each catalog's exclusives:
| Listed on one side only | Platform | Input | Output |
|---|---|---|---|
| MiniMax M3.1 Flash | Together | $0.15 | $0.60 |
| DeepSeek V4 Pro 0813 | Together | $1.32 | $3.96 |
| Qwen3.7-Max | Together | $2.50 | $7.50 |
| Qwen 3.8 Max | Fireworks | $2.00 | $6.00 |
| GLM-5.3 (US region) | Fireworks | $2.10 | $6.60 |
| Ember-1 | Fireworks | $3.00 | $15.00 |
MiniMax M3.1 Flash on Together costs half the M3 rate on both input and output, which matters for bulk classification or extraction jobs where a smaller model is good enough. Fireworks' "(US)" variants run at a 50% premium over the standard rows, the price of keeping serving inside the United States. Fireworks also lists two NVIDIA Nemotron chat models that Together's chat table does not show.
Together AI's serverless pricing page as captured on 23 Aug 2026. The MiniMax M3 and GLM-5.2 rows shown are unchanged as of 7 Nov 2026; the DeepSeek V4 Flash row has since left the serverless table.
For a wider view across more than these two vendors, the cheapest LLM API guide compares per-token rates by job, and the model leaderboard shows how these open models score against closed ones on price and benchmarks.
Fine-tuning, and the self-host escape route
Fine-tuning is where the two pricing models diverge in shape, not just level. Fireworks still prices LoRA supervised fine-tuning in three model-size bands. Together now prices each model separately, with a minimum charge per job.
| Job | Fireworks | Together |
|---|---|---|
| LoRA SFT, small model | $0.50 (up to 16B) | $0.34 (Qwen3.5 9B) |
| LoRA SFT, 70B class | $3.00 (16.1B-80B) | $2.03 (Llama 3.3 70B) |
| LoRA SFT, gpt-oss-120B | $6.00 (80B-300B) | $2.50 |
| Minimum per job | None listed | $4.00 to $60.00 |
All figures are per million training tokens, from each vendor's pricing page on 7 November 2026. Together is cheaper per token on every comparable row. Its minimums bite on small jobs and on large models: a GLM-5.2 tune costs $40 per million tokens with a $60 floor. Fireworks is running a promotion that makes supervised tuning of its Muse Glimmer 30B model free until 30 November, and says fine-tuned models serve at the same price as their base models.
The cheaper route past both is Hugging Face's Inference Endpoints or a plain GPU rental, self-hosting the same open weights without either company's markup. That saves real money for a team with the engineering time to run its own serving stack. It is also the exact chore both Fireworks and Together exist to remove for everyone else.
What actually still differs
Reliability is documented differently. Together publishes a 99% uptime SLA on its Provisioned Throughput tier, on its own homepage. Fireworks does not publish an uptime number. TokenMix, a routing service that resells both, measured 99.8% availability for Fireworks and 99.7% for Together in the first quarter of 2026. Those are TokenMix's figures, not either vendor's.
Catalog breadth points toward Together. Its models page shows "200+ models" across text, image, video, code and audio. Fireworks does not publish a running total, though its live catalog covers the same model families plus FLUX image models.
Raw compute: where the gap is largest
The GPU price gap follows directly from the neocloud question above:
| GPU, per hour | Fireworks on-demand | Together on-demand | Together preemptible |
|---|---|---|---|
| H100 | $8.00 | $3.99 | $1.99 |
| H200 | $8.00 | $5.99 | $2.99 |
| B200 | $13.00 | $9.29 | $4.65 |
| B300 | $15.00 | $11.49 | $5.75 |
Fireworks' hourly price buys a managed deployment that serves a model; Together's buys a cluster node. They are not the same product, which is why the Fireworks rate is roughly double.
Source: each vendor's pricing page, read 15 November 2026.
Together also sells managed dedicated endpoints, a product much closer to Fireworks' deployments, at $5.49 an hour for an H100. Longer cluster commitments cut its rates further, with the deepest discount on reservations of 91 days or more.
Much of that hardware, on both platforms, comes from NVIDIA and is racked in servers from builders such as Supermicro. Neither platform matches the raw speed of purpose-built silicon: Groq's LPU chips outrun both on tokens per second for the models Groq supports, a trade-off covered in the Groq vs Fireworks AI guide.
Neither vendor discloses profit, only scale. Fireworks says it passed $1 billion in annualized revenue run rate and serves more than 40 trillion tokens a day, in its Series D announcement of 15 July 2026. Together reported annual bookings above $1.15 billion in its most recent quarter, per TechCrunch. Bookings and run rate measure different things, so which company is larger cannot be settled from the numbers either has published.
What a month actually costs
List prices are easier to judge as bills. The table below applies each vendor's 7 November rates to four common monthly jobs. The arithmetic is ours; the rates are the vendors'.
| Monthly job | Fireworks | Together |
|---|---|---|
| 1 billion output tokens on GLM-5.3 | $4,400 | $4,400 |
| One H100 running all month, 730 hours | $5,840 | $2,913 |
| Same H100, preemptible | Not offered | $1,453 |
| 50M-token LoRA tune, 70B-class model | $150 | $101.50 |
The first row is the whole story of serverless inference between these two: the same model, the same bill. The second row is the neocloud gap made concrete. A Fireworks H100 deployment costs about twice a Together cluster node, but the Fireworks figure includes a tuned serving stack, autoscaling and someone else's on-call rotation. A team that would otherwise need half an engineer to run the cluster can come out ahead on Fireworks despite the higher rate.
The preemptible row only suits work that can be interrupted, such as batch evaluation or checkpointed training runs. The fine-tuning row assumes Together's Llama 3.3 70B rate against Fireworks' 16.1B to 80B band, and both sit well below the cost of the GPU time a team would rent to do the same job itself. Input tokens, cached tokens and storage are left out of every row, so treat these as floors rather than quotes.
Data handling, security and compliance
Both platforms clear the certifications most procurement checklists ask for, but they handle prompt data with different defaults, and that default is what a team actually lives with.
| Fireworks AI | Together AI | |
|---|---|---|
| Prompt retention default | Zero retention for open models | Set by the account owner |
| Zero-retention control | Admin policy blocks storing jobs | Privacy setting enables it |
| SOC 2 Type II | Yes | Yes, since July 2025 |
| HIPAA | Yes | Yes, with BAAs |
Fireworks' data handling docs say it "does not log or store prompt or generation data for any open models" without an explicit opt-in. Prompts live in volatile memory for the length of the request, or a few minutes when prompt caching is on. The exception is its Responses API, which stores conversations for 30 days when a request sets store to true, the default, unless the call opts out. Its security page lists ISO 27001, ISO 27701 and ISO 42001, plus TLS 1.2 in transit and AES-256 at rest.
Together's privacy policy says it does not train on customer data unless the customer opts in. Zero data retention exists, but it is a choice: the account owner answers "No" to storing prompts in the Privacy and Security settings, and only then is submitted content not retained. Together announced its SOC 2 Type 2 examination in July 2025 and says it signs business associate agreements for HIPAA workloads.
For a regulated team the practical difference is one setting. On Fireworks the safe state is the starting state, apart from the Responses API. On Together someone has to switch it on, and a new project that skips that step stores prompts by default.
Where Fireworks wins
Cursor is a named customer of both, which makes it a useful reference rather than a tiebreaker. On Fireworks' homepage Cursor's CPO says Fireworks helps it "train and serve the models behind Cursor at scale", including the reinforcement-learning workloads behind its Composer model. That is the strongest evidence yet that Fireworks' managed training is production-grade, not a side feature.
Notion's AI lead is quoted cutting latency from about two seconds to 350 milliseconds after fine-tuning with Fireworks, and Quora reports a 3x speedup in response time. Gumloop's product chief describes moving its main internal agent from Opus 4.8 to GLM-5.2 on Fireworks with no user-visible change. Sourcegraph and Harvey are named customers too.
If the whole job is keeping one or two specialized models fast and reliable in production, and nobody on the team wants to run a cluster, Fireworks is built for exactly that, and its case studies back it with latency numbers rather than adjectives.
Where Together wins
Together wins whenever the job needs the GPU itself. A team training from a base checkpoint, running a custom research pipeline, or bursting a week of experiments onto preemptible H100s at $1.99 an hour has no equivalent on Fireworks. Together's customers named in coverage of its round include Cursor, Cognition, the maker of Devin, and Decagon.
Teams that also need image or video generation, transcription, a code sandbox or rerank alongside chat get all of it on one bill at published unit rates: Whisper Large v3 transcription at $0.0015 a minute, a 60-minute code interpreter session at $0.03. A team juggling Google AI Studio for one modality and Replicate for another can consolidate onto Together AI's invoice instead.
Modal covers similar serverless-GPU ground for teams that want more direct control of the compute layer than either managed endpoint gives. Baseten, the third name searchers most often attach to this comparison, competes with Fireworks on managed model serving rather than with Together on clusters.
The turn: both sides are converging
The best argument against this verdict is that both companies are walking toward each other. Fireworks now sells training by the GPU-hour and says its new money funds more compute. Together, per Pulse 2.0's report on its round, expects its capacity and infrastructure footprint to grow about 50-fold over five years and already sells managed dedicated endpoints that look a lot like Fireworks' product.
That is a real risk to the split, and it is worth watching. It does not change the answer today, because the products a buyer can sign up for on 7 November still differ in kind. Together hands you a cluster if you ask for one. Fireworks does not, and its pitch, in its Series D post, is that "companies are no longer renting general intelligence" but owning specialized models. One company sells GPUs that can run models. The other sells models that happen to need GPUs.
If the choice still depends on budget, team size or workload, describe the workload and get a matched shortlist of providers, including ones beyond these two.
Switching cost: what actually moves
Both platforms expose an OpenAI-compatible chat completions endpoint, and OpenRouter routes to both, so a team on a plain serverless chat model can switch by changing a base URL and an API key. With the shared prices identical, that switch is now a reliability or feature decision, not a cost one. Stripe's deal to buy OpenRouter is covered in AIML API vs OpenRouter, for teams that route through it.
What does not move automatically: fine-tuned weights are not portable between the two without re-running the job, and the per-job economics differ, as the fine-tuning table shows. Platform-specific pieces, such as Fireworks' Training API and RL tooling or Together's GPU clusters, sandbox and rerank endpoint, have no drop-in equivalent on the other side. For a team with a fine-tuned model or one of those features in the critical path, budget a real sprint, not a config change.
What would change the verdict
Three public signals would overturn this split, and each can be checked on a vendor page.
- Fireworks lists bare GPU clusters. If its pricing page ever sells cluster nodes rather than model deployments, the neocloud answer changes and the training case for Together weakens.
- The shared prices stop matching. The eight-model match is a market rate today. If either vendor undercuts the other on the models most teams run, price becomes the deciding factor again.
- Together's managed endpoints catch up on latency. Fireworks' strongest evidence is customer latency numbers. If Together's dedicated endpoints publish comparable figures, the serving case for Fireworks narrows.
The number worth checking again in six months is not either price list, since those now move in lockstep. It is whether Fireworks starts renting bare clusters, or Together's managed endpoints catch Fireworks on the latency its customers quote. Either one would redraw this split.
Frequently asked questions
Is Fireworks AI cheaper than Together AI?
Not on shared serverless models. On 15 November 2026 both list identical rates on all eight headline models they share, for example GLM-5.3 at $1.40 input and $4.40 output per million tokens. Together is cheaper for raw GPUs and per-token fine-tuning, while Fireworks charges no per-job fine-tuning minimum.
Which one should a small startup team pick?
It depends on whether the team wants to run hardware. A team serving or tuning one or two specialized models with no cluster to manage fits Fireworks, whose named users include Notion, Quora and Gumloop. A team that needs GPUs it controls, or chat, image, audio and rerank on one invoice, gets more from Together.
Do Fireworks and Together AI support the same models?
Largely, since both host the main open-weight families: DeepSeek, GLM, Kimi, Qwen, MiniMax and gpt-oss. Together counts more than 200 models on its catalog page, while Fireworks gives no total. Each also carries a few the other lacks: Qwen 3.8 Max and MiniMax M3 are now Fireworks-only, while MiniMax M3.1 Flash is sold only by Together.
How does GPU pricing compare for training versus serving?
Together rents bare GPU cluster nodes at $3.99 an hour for an H100 on demand and $1.99 preemptible. Fireworks charges $8.00 an hour for an H100 deployment, but that price buys a managed model deployment, not a cluster. For serving through the per-token API, the shared models cost the same on both.
Which platform has better uptime?
Only Together puts a number in writing: a 99% uptime SLA on Provisioned Throughput. Fireworks offers no published SLA. Independent monitoring by a reseller put both platforms above 99.5% in early 2026, Fireworks marginally ahead, but a third party's measurement is not a contractual commitment from either vendor.
Covered in this guide
- Fireworks: Enterprise inference platform for open models: serverless tokens from $0.10 per 1M, managed fine-tuning per training token, and dedicated NVIDIA GPUs from $8.00/hr. $1 free credits, self-serve, no cold starts.
- Together: The AI Native Cloud: a full-stack platform for training, fine-tuning, and deploying open-source AI models
- Baseten: Baseten is an AI inference platform for custom and open-weight models, with a no-fee Basic plan, per-minute GPU billing, SOC 2 Type II and HIPAA.
- CoreWeave: Bare-metal NVIDIA GPU cloud with managed Kubernetes and Slurm, 59 availability zones across the US, Canada and Europe, and free data egress.
- Crusoe: GPU cloud with openly published hourly rates, free network egress, Managed Kubernetes and serverless inference and fine-tuning.
- Cursor: Cursor is an AI code editor built on VS Code, widely adopted across large enterprise engineering teams, with Agent Mode, Tab completion, and Cloud Agents.
- Decagon: Decagon is a $4.5B-valued AI concierge that automates chat, voice, email and SMS support for enterprises, with contracts typically starting near $95K/year.
- DeepSeek V4 Pro 0813: DeepSeek-V4-Pro-0813 is DeepSeek's flagship 1.6T-parameter mixture-of-experts model, generally available since August 2026 with MIT-licensed open weights.
- DeepSeek V4.1 Flash: The MIT-licensed vision-and-text model, launched in September 2026 and refreshed with an October snapshot.
- DeepSeek V4.1 Flash 1017: DeepSeek's October 2026 update to V4.1-Flash: MIT-licensed weights, image input and a 1M-token context, tuned for coding agents and long task loops.
- GLM-5.2: GLM-5.2 is Z.ai's June 2026 open-source 744B MoE model with a 1M-token context window, 85.6% GPQA Diamond (Vals AI), and MIT license at $1.40/1M input tokens.
- GLM-5.3: GLM-5.3 is Z.ai's text-only flagship for agentic coding and security work, with a 1M-token context, a live API and open weights.
- GLM-5.3 Flash: Open-weight, natively multimodal MoE model from Z.ai with a 1M-token context window and MIT license, released in August 2026.
- Google AI Studio: Free web-based IDE for building, testing, and deploying generative AI applications with Google's Gemini models
- Groq's: Fast open-model inference, now built on NVIDIA-licensed LPU hardware
- Harvey: Legal AI platform serving 60 of AmLaw 100 firms. Document review 80x faster with citations, contracts, due diligence, and workflow agents. Enterprise-only.
- Hugging Face's: The AI community building the future. Platform for discovering, sharing and collaborating on machine learning models, datasets and applications.
- Kimi K3: 2.8T-parameter open-weight MoE model from Moonshot AI (July 2026) with a 1M-token context window and the top published open-weight result on GPQA Diamond.
- Kimi K3.1 Pro: Moonshot AI's October 2026 reasoning model: a 1M-token window, always-on thinking, text and image input, and open weights on Hugging Face.
- MiniMax: Listed in Hong Kong since January 2026 and controlled by founder Yan Junjie, MiniMax builds the MiniMax M3 language model and the open-weight H3 video model from its Shanghai base.
- MiniMax M3: MiniMax's June 2026 open-weight flagship, built around a 1M-token context window, native multimodal input, and frontier-level agentic coding performance.
- MiniMax M3.1 Flash: MiniMax's October 2026 Flash model: 1M-token context, text, image and video input, and always-on reasoning with five effort levels.
- Modal: AI infrastructure that developers love: serverless compute for ML inference, training, and batch processing
- Moonshot AI: Beijing AI lab founded in 2023 by Yang Zhilin, maker of Kimi and the open-weight Kimi K3 models, valued at $20B at its May 2026 round.
- Muse Spark Open: Meta's open-weight model for coding and agents: 816B total parameters, 42B active, released October 2026 under the Muse Model License.
- Notion's: Your AI everything app: All your tools and work in one unified workspace.
- NVIDIA: Santa Clara chipmaker that supplies the GPUs, networking, CUDA software and open Nemotron models behind most large AI training and inference clusters.
- OpenRouter: Single API endpoint for hundreds of AI models from dozens of providers, billed at provider list prices with one bill and no vendor lock-in.
- Qwen3.7-Max: Qwen3.7-Max is Alibaba Cloud's text-only flagship agent model from May 2026, called through Model Studio's OpenAI-compatible API with a million-token context window.
- Qwen 3.8 Max: Qwen3.8-Max is Alibaba's flagship MoE model, launched Aug 2026, with a 1M-token context, native multimodal input, and a text-only open-weight checkpoint released weeks later.
- Sourcegraph: Enterprise code search and AI Deep Search for multi-repo codebases. Its Cody assistant is being retired in January 2027.
- Supermicro: Supermicro designs AI server hardware from edge systems (65W TDP) to 8-GPU data center nodes, with $12.68B quarterly revenue and NASDAQ: SMCI listing.
- Together AI's: Together AI develops infrastructure for training and deploying open-source language models at scale. The company emphasizes cost-effective model training, inference optimization, and community-driven AI development.
- Z.ai: Z.ai (formerly Zhipu AI) is a Beijing lab behind the GLM model family, founded 2019 and listed on the Hong Kong Stock Exchange (2513) since January 2026.
Sources
Still deciding?
This guide covers a handful of options. Smart Match checks every listing in the directory against how you actually work and what you can spend, then hands you the shortlist and the reason behind each pick.
Find your tools

