Independent/Reader-funded/Infrastructure, not tokens
DeAINEWS

AI you control — open models, private inference, and the networks that run them.

Industry & Business

Hyperbolic Inference Pricing in 2026: Every Model Priced, Verified

Hyperbolic's relaunched API prices 11 open models as of 2026-10-21, from $0.06/$0.18 per million tokens (DeepSeek V4 Flash) to $3.00/$15.00 (Kimi K3).

DeAI is powered by Morpheus (mor.org). We cover competing providers on the same terms — see our methodology.

A single GPU server rack in a colocation aisle under cool white light with one amber status LED — the physical hardware behind Hyperbolic's relaunched per-token rate card. Illustration: DeAI
A single GPU server rack in a colocation aisle under cool white light with one amber status LED — the physical hardware behind Hyperbolic's relaunched per-token rate card. Illustration: DeAI

Hyperbolic Inference — the relaunched per-token API from GPU marketplace Hyperbolic — lists 11 open text models as of 2026-10-21, from $0.06 input / $0.18 output per million tokens on DeepSeek V4 Flash 0731 to $3.00/$15.00 on Kimi K3. It returned on October 15 after a three-week outage in which the company's previous serverless API was retired with no notice.

Key facts

  • The card is 11 open models, input / cached input / output: DeepSeek V4 Flash 0731 at $0.06 / $0.015 / $0.18, MiMo-V2.6-Flash at $0.14 / $0.0028 / $0.28, GLM-5.3 Flash at $0.15 / $0.03 / $0.50, Qwen3.8 27B at $0.20 / $0.02 / $2.00, MiniMax M3 at $0.30 / $0.06 / $1.20, DeepSeek V4.1 Flash at $0.30 / $0.006 / $1.20, Nemotron 3 Ultra at $0.50 / $0.10 / $2.20, Hy4 Preview at $0.834 / $0.042 / $2.501, Step 5 Preview at $1.00 / $0.10 / $2.70, GLM-5.3 at $1.40 / $0.26 / $4.40, Kimi K3 at $3.00 / $0.30 / $15.00 (Hyperbolic inference page, accessed 2026-11-17).
  • The previous API died without a date: the standalone serverless inference page began returning HTTP 404 on August 11, 2026; the API and playground were found retired on September 24, 2026, with no shutdown date, notice period, or release date for a replacement (UsagePricing pricing history).
  • The relaunch is a new product, not a restore: Hyperbolic Inference went live October 15, 2026 at base URL https://api.hyperbolic.ai/inference/v1, text models only, and does not accept the old endpoint, keys, or model IDs (Hyperbolic Flux blog post).
  • Access is paywalled: Free-tier accounts can call Inference only after phone verification (which adds a $1 credit); a one-time $5 deposit auto-promotes an account to Pro — the same deposit that unlocks GPU rental. Credits never expire (UsagePricing).
  • Rate limits are 300 RPM per key, with 600-second request caps on glm-5.3-flash and qwen3.8-27b and 300 seconds elsewhere; no tokens-per-minute table is published (Hyperbolic inference page; UsagePricing).
  • Footing: Hyperbolic announced a $30 million Series A1 on October 27, 2026, led by Walden Catalyst (the firm of Intel CEO Lip-Bu Tan), bringing total funding to $50 million; the company says more than 250,000 builders use the platform and that the platform manages about $200 million a year in compute (Hyperbolic announcement; AI Funding Tracker).

What Hyperbolic is

Hyperbolic (Hyperbolic Labs, founded 2023, San Francisco) began as a DePIN-style GPU marketplace: rather than owning datacenters, it aggregates underused GPU capacity from third-party datacenters and operators and resells it, refreshing per-hour rates weekly as supply shifts. Its 2026 identity has shifted toward orchestration — the Forge control plane standardizes mixed bare-metal clusters, managed Kubernetes and the inference service sit on top, and the announced Global Compute Grid would let suppliers and customers buy, sell, and resell capacity (Vertex Holdings, 2026-11-06).

Funding, from primary announcements and coverage: roughly a $7 million seed (July 2024), a $12 million Series A led by Variant and Polychain (December 2024), and the $30 million Series A1 announced October 27, 2026 — led by Walden Catalyst with Dell Technologies Capital, General Catalyst, Vista Equity Partners, Cambium Capital, and Salience Capital participating, bringing total funding to $50 million (Hyperbolic; AI Funding Tracker). The company says over 250,000 builders use the platform, with teams at Modal, Higgsfield, Daytona, HeyGen, and Pika cited as customers. AI Funding Tracker reports a company target of $150 million ARR by end of 2026 — a company claim we have not verified.

The full card, priced (2026-10-21, verified 2026-11-17)

Every model on the relaunched serverless card, first-party list rates (Hyperbolic inference page):

ModelInput $/MCached input $/MOutput $/M
DeepSeek V4 Flash 0731$0.06$0.015$0.18
MiMo-V2.6-Flash$0.14$0.0028$0.28
GLM-5.3 Flash$0.15$0.03$0.50
Qwen3.8 27B$0.20$0.02$2.00
MiniMax M3$0.30$0.06$1.20
DeepSeek V4.1 Flash$0.30$0.006$1.20
Nemotron 3 Ultra 550B A55B$0.50$0.10$2.20
Hy4 Preview (Tencent)$0.834$0.042$2.501
Step 5 Preview (StepFun)$1.00$0.10$2.70
GLM-5.3$1.40$0.26$4.40
Kimi K3$3.00$0.30$15.00

DeAI cross-checked this table against the October 21, 2026 rate card captured by UsagePricing: every model and rate matches, so the card has been stable for nearly four weeks at press time. The page itself carries no "as of" date, only the standing caveat that models and prices change over time — so date-stamp every figure before you commit spend.

Two structural notes. First, the Flash tiers own the price floor: DeepSeek V4 Flash at $0.06/$0.18 and MiMo-V2.6-Flash at $0.14/$0.28 undercut every row we've priced this quarter at the floor end — DeepInfra's V4 Flash card sits at $0.06/$0.18 with cached input at $0.015, essentially identical, while Novita and Fireworks price DeepSeek V4.1 Flash identically at $0.30/$0.006/$1.20. Second, the cache schedule is uneven: DeepSeek V4.1 Flash's $0.006 cached input is the deepest published discount on the card (98% off), MiMo-V2.6-Flash's $0.0028 is close behind, but Qwen3.8 27B's $2.00 output is the worst output rate on the card relative to its intelligence tier — 6.7x GLM-5.3 Flash's output price for a much smaller model.

The relaunch story: dark for three weeks, no notice

This page exists because the pricing history is the story. Between August 11 and October 15, 2026, Hyperbolic's inference product went through a two-step retirement with no dated announcement:

  1. August 11 — the standalone Serverless Inference marketing page began returning 404, and the live docs rate card shrank to two $0.40/M text SKUs, with image models and several text and vision-language models flagged "Sunset."
  2. September 24 — the API itself was found retired: docs redirected to a four-question FAQ, the billing docs' per-API-call meter disappeared, and requests to old models failed even with valid keys. The FAQ said old keys and code would "keep failing until the new product launches," with no release date given.

For roughly three weeks the company's only inference guidance was to rent an on-demand GPU and run your own serving stack — which shifts utilisation risk entirely onto the customer. A former API customer paying $0.40/M blended faced, in the worst case, a dedicated H100 SXM at $3.19/hour (about $2,329/month) billed whether or not requests arrived (UsagePricing).

The relaunch on October 15 is a different product rather than a restoration: the new card splits input, cached input, and output (the old card charged one blended rate per model), the floor dropped ($0.06/M input versus $0.10/M before) while the ceiling rose ($15.00/M on Kimi K3 versus $4.00/M on the old Llama-3.1-405B row), image and audio models were dropped, and the Free tier — which could call the old API — cannot call the new one without phone verification. Credits survived the gap untouched, because the prepaid wallet was never tied to the inference SKU.

For builders, the durable lesson is about sequencing, not spite: the pricing page disappeared first, the API went next, and no deprecation window separated "Sunset" flags from shutdown. If you build on any single host's serverless card, keep the migration path — an OpenAI-compatible base-URL swap — rehearsed.

Cache economics and the GPU alternative

The card's cached-input schedule rewards agentic workloads with long stable system prompts: DeepSeek V4.1 Flash caches at $0.006 (98% off), MiMo-V2.6-Flash at $0.0028 (98% off), Kimi K3 at $0.30 (90% off), GLM-5.3 at $0.26 (81% off). No separate cache-write fee is published on the captured card.

Worked example, assumptions labeled: an agentic request with a 20k-token system prompt (95% cache hit) and a 2k-token response on DeepSeek V4.1 Flash bills about $0.0003 uncached input + $0.000114 cached + $0.0024 output ≈ $0.0028 per request. That matches the figure we published for Together AI's identical card rate — the two hosts price this model identically, so caching, not host choice, moves the bill.

The alternative is the GPU side of the same platform. Hyperbolic's on-demand rates — H100 SXM at $3.19/GPU-hour, H200 at $3.99, B200 at $5.99 — have held since August 4, 2026, after a volatile year in which the H100 rate moved $1.50 → $2.89 → $3.19 across June, July, and August (+112.7% in seven weeks). The per-instance price lock means a running instance keeps its launch rate for life, but new instances inherit whatever the weekly supplier refresh produces. A dedicated instance costs about $2,329/month; at the card's prices that buys roughly 4.7 billion output tokens on GLM-5.3 Flash or 12.9 billion on DeepSeek V4 Flash — so steadily busy workloads at those scales are the crossover zone, and bursty traffic is what the per-token meter is for.

Rate limits, exactly

The published limits are per-key, not per-model: 300 requests per minute, with a 600-second request time limit on glm-5.3-flash and qwen3.8-27b and 300 seconds on other models. Account tiers add separate platform-level RPM limits — 60 RPM on Free (usable for Inference only after phone verification), 600 RPM on Pro, custom limits on Enterprise. No token-per-minute table exists, which puts Hyperbolic with Together and against Fireworks and Groq, whose published adaptive-limit tables we've priced in earlier cards.

Privacy: the FAQ says discard, policy says treat it as a claim

Hyperbolic's privacy FAQ states that inference input is collected "solely for the purpose of processing your request," is "not stored or retained," and is "immediately discarded" after completion; it also states the company does not "use, retain, or train our models on any user-submitted data" (Privacy FAQ). DeAI's standing rule applies: this is a self-published policy statement, not an independently verified attestation — the same framing we apply to every host in this series, including hosts whose statements are stronger on their face. No third-party audit, TEE-attestation, or compliance certification appears in the provider's public materials we reviewed.

How to choose

Hyperbolic Inference is the fit when you want DeepSeek V4 Flash or MiMo-V2.6-Flash at or near the cheapest published per-token rates in the field, can live with a three-week-old API and a credit wallet, and don't need image or audio endpoints. It is not the fit when you need a published token-per-minute table, TEE-verifiable privacy, or a provider with a longer uninterrupted inference track record — the September outage is recent. For the rest of the priced field, our host cards cover DeepInfra, Together AI, Fireworks, Novita, Groq, Cerebras, and SambaNova; the cross-host floor is tracked weekly in the Price Index. Prices move — treat every figure here as dated to 2026-10-21 (verified stable 2026-11-17) and re-verify against Hyperbolic's inference page before committing spend.

Questions

How much does Hyperbolic Inference cost per million tokens?
As of 2026-10-21, the card runs 11 open text models. DeepSeek V4 Flash 0731 is the floor at $0.06 input / $0.18 output; MiMo-V2.6-Flash is $0.14/$0.28 with cached input at $0.0028; GLM-5.3 Flash is $0.15/$0.50; DeepSeek V4.1 Flash is $0.30/$1.20 with cached input at $0.006; Kimi K3 is the ceiling at $3.00/$15.00. Reasoning tokens bill as output.
What happened to Hyperbolic's old serverless inference API?
It was retired. The marketing page 404ed on August 11, 2026; the API itself was found retired on September 24, 2026, with no shutdown date or notice period announced. Per-token inference returned October 15, 2026 as Hyperbolic Inference on a new base URL — api.hyperbolic.ai/inference/v1 — which does not accept the old endpoint, keys, or model IDs. Image and audio models were not carried over.
Does Hyperbolic have a free tier for the Inference API?
Effectively no. Free-tier accounts can call Inference only after verifying a phone number, which adds a $1 credit, and a one-time $5 deposit auto-promotes the account to Pro. Hyperbolic sells no subscription: everything draws from prepaid credits that never expire, with a $5 minimum per purchase.
What are Hyperbolic's Inference API rate limits?
300 requests per minute per API key, with a 600-second request time limit on glm-5.3-flash and qwen3.8-27b and 300 seconds on other models, per the docs. Account tiers add separate platform RPM limits — 60 RPM on Free after phone verification, 600 RPM on Pro. No token-per-minute table is published.
Does Hyperbolic train on your prompts?
Hyperbolic's privacy FAQ states that inference input is used solely to complete the request and is not stored or retained, and that the company does not use or train on user-submitted data. DeAI treats this as a self-published policy statement, not an independently verified attestation.
Is Hyperbolic Inference cheaper than renting a GPU?
At Hyperbolic's $3.19/GPU-hour H100 rate, a dedicated instance costs roughly $2,329 per month. At the $0.50/M output price on GLM-5.3 Flash, that buys about 4.7 billion output tokens; at $0.18/M on DeepSeek V4 Flash, about 12.9 billion. Below those volumes the per-token API is the cheaper route; above them, a rented GPU and your own serving stack can win.

Sources

  1. Serverless AI Inference API | Leading Models — Hyperbolic — Hyperbolic
  2. Hyperbolic Pricing — UsagePricing (rate card dated October 21, 2026; pricing history 2024–2026) — UsagePricing
  3. Introducing Flux: Next-Generation Inference Powered by Multi-Cloud Orchestration — Hyperbolic blog
  4. Privacy FAQ — Hyperbolic
  5. Announcing Hyperbolic's $30M Series A1 Funding — Hyperbolic blog
  6. Hyperbolic raises US$30 million to turn fragmented GPU capacity into a single market for AI compute — Vertex Holdings (investor newsroom)
  7. Hyperbolic Raises $30M Series A-1 led by Intel CEO Lip-Bu Tan's Walden Catalyst — AI Funding Tracker
  8. NVIDIA H200 Price in 2026 — Hyperbolic blog — Hyperbolic blog

About DeAI

DeAI is an independent publication covering open-weight AI models, private inference, and decentralized infrastructure — the tools for running AI you actually control. We test providers on price, privacy, and refusal behavior and publish the numbers, not the vibes. DeAI is powered by Morpheus (mor.org), a decentralized inference marketplace, and covers it on the same terms as every other provider.

Powered by Morpheus and StrandCMS

Morpheus is a decentralized inference marketplace, covered on the same terms as every other provider — we rank it wherever the data lands. StrandCMS is the open-source, agent-first framework this site is built on.

Learn more about the Morpheus Inference API →

Sponsor disclosure — not editorial

Powered by Morpheus and StrandCMS. Morpheus is a decentralized inference marketplace, covered on the same terms as every other provider. StrandCMS is the open-source, agent-first framework this site is built on.

Learn more →