Together AI Pricing 2026: Dedicated Endpoints vs a Fixed-Price Plan

Together AI runs two separate hourly-GPU products, and they don't bill the same way. Dedicated Endpoints lease you a single GPU with a managed serving stack; GPU Clusters rent you multi-GPU nodes on-demand or on a reserved term. Together AI pricing 2026 for either product means tracking per-minute billing floors, idle time, and a reserved-term ladder that only shows up on the GPU Clusters side, none of it covered by the serverless per-token rate cards most comparisons stop at. Here's exactly how both products bill, the published rates for each, and how either one measures up against a fixed-price inference plan that charges by counted tokens instead of GPU-hours.

TL;DR: Together AI Pricing 2026 for Dedicated Endpoints vs a Fixed Plan

Together AI's Dedicated Endpoints bill $5.49/hr for H100 80GB and $8.99/hr for B200 180GB per minute, idle or busy, as of 5 Nov 2026.

Hourly billing suits model breadth and compute control; a fixed token plan suits avoiding idle-time risk. Compare current GPU pricing.

How Together AI's Dedicated Endpoints Are Billed: Per Minute, Idle or Not

A Dedicated Endpoint isn't metered by requests or tokens. Together AI's own pricing documentation states plainly that "a deployment bills for as long as it runs, not per token or per request," and that it's billed by the minute regardless of how much traffic actually hits it. Leave an endpoint running overnight with zero requests and the meter keeps counting exactly as if it were saturated.

The lever that matters here is the scaling configuration, not the traffic pattern. minReplicas sets a continuous billing floor, the number of GPUs you're paying for at all times no matter how quiet the endpoint gets, while maxReplicas caps the billing ceiling, the most you can be charged for even under a traffic spike. Set minReplicas to 1 and you're always paying for at least one GPU; set it to 0 and the deployment can scale all the way down.

A replica stops billing the moment it scales down, and a deployment that's scaled to zero or explicitly stopped costs nothing. Together's docs also describe an inactivity timeout that can trigger that scale-to-zero automatically, so a genuinely idle endpoint with minReplicas at 0 doesn't need to be babysat to stop the bill. That's the real idle-time trap worth knowing before deploying: Together doesn't bill idle GPUs unconditionally, but the minReplicas floor you set decides whether idle time is free or not, and the default most teams reach for, keeping at least one replica warm for latency, is also the default that keeps the meter running.

Dedicated Endpoint Rates: H100 and B200 (H200 Is Quote-Only)

Together AI publishes exactly two Dedicated Endpoint rates and gates the third behind a sales conversation. Here's the full rate card as it stands today:

Together AI's own pricing documentation lists $5.49/hr for a single H100 80GB and $8.99/hr for a single B200 180GB, with H200 141GB marked as requiring a sales conversation rather than carrying a self-serve number. That's a real gap for anyone trying to compare GPUs on this product head to head: you can rate-shop H100 against B200 directly, but H200 only shows up once someone from Together quotes you a number, which makes it the one tier you can't put in a spreadsheet before talking to sales.

Both published rates are per GPU, per hour, and bill on the per-minute, idle-or-not basis described above, not on request volume or token count. Together's own framing for this product is that it's built for teams that, as they scale, "need guaranteed performance, more control and customizability over their deployment, and support for custom models," which is the tradeoff a rate card alone doesn't show: you're paying for a dedicated GPU and a managed serving stack together, not just raw compute.

Together AI Pricing 2026 for GPU Clusters: the Reserved-Term Ladder Down to $3.19/hr

GPU Clusters are a separate product from Dedicated Endpoints, and the pricing structure is different too: instead of one rate and a sales-quote exception, GPU Clusters run a full reserved-term ladder that gets cheaper the longer you commit.

Together AI's pricing page lists H100 at $3.99/hr on-demand, stepping down to $3.69/hr for a 3-30 day term, $3.45/hr for 31-90 days, and $3.19/hr for 91-180 days, the lowest published self-serve rate on this product. H200 and B200 follow the same shape at different price points: H200 runs $5.99/hr on-demand down to $3.99/hr at 91-180 days, and B200 runs $8.19/hr on-demand down to $6.79/hr at the same term length. Commit past 180 days on any GPU and the rate goes quote-only, the same as H200 on Dedicated Endpoints.

Together also sells preemptible capacity on GPU Clusters at roughly half the on-demand rate: $1.99/hr for H100, $2.99/hr for H200, and $4.09/hr for B200. That's real savings if your workload can tolerate reclaim, but it's a different risk profile from the Dedicated Endpoint's steady per-minute billing and from the reserved ladder's guaranteed capacity for the term.

The two products also aren't priced on the same scale. A single H100 Dedicated Endpoint costs $5.49/hr against GPU Clusters' $3.99/hr on-demand rate for what's functionally the same chip. That gap is the managed serving layer Dedicated Endpoints wraps around the hardware: you're not paying more for the GPU, you're paying for Together running the inference stack instead of you.

Together AI's rates above are current as of 5 Nov 2026, per Together's own published pricing and documentation pages. GPU pricing changes without notice, so check Together's pricing page before budgeting off any figure here. This post quotes no Spheron GPU-hour rate directly, since the comparison below is against a token-based fixed-price plan, not an hourly rental rate.

Matching GPU-Hour Billing to Counted Tokens on a Fixed Plan

Every rate above is priced in GPU-hours: you're buying a slice of compute time, and the bill runs whether or not a token gets generated. Spheron's fixed-price inference plans bill a different unit entirely: a monthly allowance of counted tokens, spent against a prepaid, fixed price rather than metered by the minute. Every plan serves the same three models, GLM-5.3, GLM-5.3-Flash, and DeepSeek V4.1 Flash, with Dedicated plans able to add a custom or fine-tuned open model on top. Starter runs month-to-month on a 5B-token monthly allowance with 8 concurrent requests; Growth holds its rate for a 3-month term on a 20B-token allowance with 32 concurrent requests; Dedicated is sized for 50B+ tokens a month on single-tenant, region-pinned nodes in the EU or US. The GPUs behind every tier are reserved, never spot, which is the opposite trade from the preemptible discount on Together's own GPU Clusters.

Counting isn't 1:1 with raw tokens moved. Output tokens count 4x a fresh input token, and cached input counts 0.25x, dropping to 0.03x specifically on DeepSeek V4.1 Flash. A workload leaning on repeat agent context, the same system prompt and tool definitions sent on every call, drains the allowance far slower than its raw token throughput would suggest, which is a different way of solving the same problem Together's own minReplicas floor is aimed at: both products are ways of pricing in advance for a workload's actual shape rather than billing it fully linearly.

Like Together's own reserved tiers past 180 days, Spheron's plan price is quoted rather than published, so there's still no public number to drop into the tables above. Usage past the monthly allowance keeps serving at the plan's rate rather than cutting off, and unused tokens expire at the end of the month rather than rolling over.

When Dedicated-by-the-Hour Beats a Fixed Token Plan, and When It Doesn't

The two products are solving different problems, and the billing mechanics above point to where each one actually wins.

Together's Dedicated Endpoints and GPU Clusters win when you need model breadth or raw compute control a fixed plan doesn't offer. Dedicated Endpoints deploy a much wider range of open models than Spheron's three-model lineup, on your own schedule; GPU Clusters give you the multi-GPU hardware directly with no serving stack imposed on top, which matters if you're fine-tuning or running your own inference engine. And if your traffic is large and steady enough to hold a 91-180 day commitment, GPU Clusters' $3.19/hr H100 rate (or $3.99/hr H200, $6.79/hr B200) is a real discount nothing on the Dedicated Endpoint side matches.

A fixed-price plan wins when the GPU-hour billing itself is the risk you're trying to remove. There's no minReplicas floor to size against your own idle-time tolerance, no inactivity timeout to configure, and no distinction between a quiet hour and a busy one: the plan's price holds regardless of how traffic lands within the month. Concurrency is stated as a number, 8 or 32 requests at once, rather than something you derive from a GPU count and a serving framework's throughput. The three-model catalog is the honest tradeoff for that simplicity: if GLM-5.3, GLM-5.3-Flash, or DeepSeek V4.1 Flash cover your workload, that's the whole decision; if they don't, Together's broader catalog is the only one of the two options that can serve you at all today.

Running the same traffic pattern through Azure OpenAI's PTU pricing turns up a near-identical idle-billing structure under a different name, and AWS Bedrock's Provisioned Throughput is a third hyperscaler running the same fixed-hourly-unit model. For the fuller field of GPU clouds and inference platforms beyond Together, Together AI alternatives covers ten more options, and Spheron's own head-to-head against Together's rental and inference pricing runs the GPU-rental math this post doesn't repeat. The counted-token weighting described above gets a fuller worked example in DeepSeek V4.1 Flash pricing: API vs a fixed-price plan.

Together's own rate card bills by the GPU-hour whether you use it or not; a Spheron inference plan bills by the token and holds its price for the term.

Request a Spheron inference plan →