Skip to main content
Week of November 9, 2026

Updates

  • Legacy API keys no longer work for any API. API keys issued in the older format, which don’t start with sk_live_, were already rejected for inference. Now every Hyperbolic API, including GPU rentals and account management, rejects them with 401. They no longer appear on the Personal API Keys page. If you still use one, create a new key under Personal API Keys and use the sk_live_ key it gives you. See API keys.
Week of November 9, 2026

Updates

  • Node IDs now use your rental ID. Each node in a rental now has an instanceRef of the form <rentalId>-<node index>, such as <rentalId>-0, instead of an ID from the underlying cloud provider. Read it from instances[].instanceRef in GET /v2/on-demand/rentals/active or GET /v2/on-demand/rental, and pass it to POST /v2/on-demand/rentals/reboot-instance. This is a breaking change: node IDs saved before this change no longer work and return INVALID_INSTANCE_REF, so update any scripts that store them. On bare metal rentals, nodes can be rebooted only after every server in the rental has been provisioned. Until then, reboot requests for that rental return INVALID_INSTANCE_REF. See Rebooting an instance and Reboot a single instance within a rental.
  • Copy the rental ID from the instance details page. The page header now always shows the rental ID with a copy button. See Managing instances.
Week of November 2, 2026

New features

  • Choose your reserved term on a calendar. Reserving GPUs, switching an on-demand instance to reserved, and extending a reservation now use the same end-date calendar. When you rent or switch, select Reserved term, then pick the day your reservation ends, up to 30 days out, or use the 1 Week, 2 Weeks, or 1 Month quick picks. A summary shows your rate, total cost, and savings before you confirm. If your balance is short, you can add funds without leaving the dialog. For terms longer than 30 days, contact sales. See Reserve an instance and Extending a reservation.
  • See sold-out configurations on the GPUs page. The GPUs page now has All and Available tabs, each showing how many configurations match your filters. Available, the default, lists what you can rent now. All also lists configurations Hyperbolic offers that are sold out right now, each marked sold out. Filters now list every value Hyperbolic offers: under All, values with no free capacity are labeled Sold out, and under Available, they can’t be selected. If nothing is available for your filters but sold-out configurations match, click Show sold out to switch to All. On your personal account, click Notify me on a sold-out card, or Create alert below the list, to get an email when that configuration comes online. See Sold-out configurations on the GPUs page.

Bug fixes

  • Security updates. We patched third-party dependencies in Hyperbolic platform services to resolve known vulnerabilities. No action is needed on your part. See Security and compliance.
Week of November 2, 2026

Updates

  • Managed Kubernetes clusters show GPUs in use. The cluster Overview now shows GPUs in use, such as 8 / 16. This counts the GPUs requested by running pods out of the GPUs your nodes provide. Before, the card counted only GPUs above 5% utilization, so GPUs held by idle inference pods looked unused. Average utilization now appears below the count. See Monitoring.
  • New link preview for the dashboard. When you share a link to app.hyperbolic.ai, the preview card on social sites and chat apps now shows the current Hyperbolic branding instead of the old logo. The card also has alt text for screen readers on X (Twitter).

Bug fixes

  • Missing GPU utilization shows as unknown, not 0%. If a node doesn’t report GPU utilization, the cluster page now shows — or No data instead of 0%. Averages include only GPUs that report utilization and note when only some are reporting, such as “6 of 8 GPUs reporting”. Nodes without utilization data are no longer labeled idle and sort after nodes with readings. See Monitoring.
  • New virtual machines show their correct specs. New virtual machine rentals in us-east-1, us-east-2, and ca-east-3 could list 0 GB of RAM, 0 GB of storage, and 0 vCPUs. They now show the RAM, storage, and vCPU count of the machine size you picked. This fix applies to rentals created from now on. See Managing instances.
Week of November 2, 2026

New features

  • Reboot virtual machines in more regions. You can now reboot virtual machine rentals in us-east-1, us-east-2, and ca-east-3 from the instance’s Reboot action or POST /v2/on-demand/rentals/reboot-instance. These reboots are a hard power-cycle. Disk contents are preserved, but anything in memory is lost. While the reboot runs, the node shows Rebooting, then returns to Running. A reboot request is refused while one is already in progress. See Rebooting an instance and Reboot a single instance within a rental.

Updates

  • expectedReservedUntil is now currentReservedUntil. In POST /v2/on-demand/extend-reserved, the field for your reservation’s current end date is renamed to currentReservedUntil, so it’s no longer mistaken for the new end date. Send the extension quote’s top-level reservedUntil, not the option’s newReservedUntil. This is a breaking change: requests that still send expectedReservedUntil are rejected, so update any scripts that extend reservations through the API. Extending from the dashboard isn’t affected. See Extend a reserved rental.
  • Custom end dates work like the preset lengths. When you reserve GPUs, extend a reservation, or change an on-demand rental to reserved, the Custom end date card now always reads Any length up to 30 days. Your picked dates and price appear below the cards instead of on the card. When you reserve GPUs, a summary under the calendar shows your dates, the approximate end time in your timezone, the length, and the price per GPU-hour. If you get a discount, it also shows the discount and your savings. The discount hint is now one line, such as “Move to Nov 10 or later for 1% off”. Click the date to pick it. Hovering a day no longer shows a price. See Reserve an instance and Extending a reservation.

Bug fixes

  • Settings shows the sign-in prompt when you’re logged out. Opening Settings while logged out could show an error. You now see the sign-in prompt on the same page, and after you sign in you land on the settings page you opened. See Account management.
  • Clearer status for stopped virtual machines. In us-east-1, us-east-2, and ca-east-3, a virtual machine that’s off now shows Unknown instead of Failed when Hyperbolic can’t confirm whether a reboot is running. See Rebooting an instance.
Week of November 2, 2026

New features

  • Image input on vision models. glm-5.3-flash, step-5-preview, minimax-m3, kimi-k3, and mimo-v2.6-flash now accept images in chat completion requests. Send an image as an image_url part of a user message, as a base64 data URL, in the same format as OpenAI. All of them except glm-5.3-flash also accept public HTTPS URLs. In the Model Playground, these models open in a vision playground with an attachment button. Video input is experimental and available through the API only on step-5-preview, minimax-m3, kimi-k3, and mimo-v2.6-flash. glm-5.3-flash doesn’t accept video, and the playground can’t attach videos yet. See Image input and Image and video input.
  • Reboot bare metal instances. You can now reboot bare metal rentals in eu-west-3, eu-west-4, and uk-southeast-3 from the instance’s Reboot action or POST /v2/on-demand/rentals/reboot-instance. Bare metal reboots are a hard power-cycle and can take several minutes. Disk contents are preserved, but anything in memory is lost. The instance details page now lists each bare metal server so you can reboot one at a time. A reboot request is refused while that server is already rebooting. See Rebooting an instance and Reboot a single instance within a rental.

Updates

  • Custom reservation lengths go up to 30 days. Custom lengths now accept up to 30 days for new reservations, extensions, and conversions from on-demand to reserved. Requests over 30 days return BAD_REQUEST. To rent longer, extend again once your queued extension begins. See Reserve an instance and Extending a reservation.
  • Refreshed quick-start cards on Home. The On-Demand GPU Clusters, Model Playground, Managed Kubernetes, and Dedicated Model Hosting cards on the dashboard Home page have new illustrations that fill each card. The images are much smaller, so Home loads faster. See Quickstart.
  • Usage recordings now show fixed interface text. Hyperbolic’s product analytics session recordings used to hide all text in the app. They now show fixed interface text: navigation, headings, labels, table headers, tabs, and button text. Your name, email, and balance in the top navigation are therefore visible in recordings. All other text stays masked, including your Model Playground prompts and responses, API keys, IP addresses, SSH commands, and endpoints. Everything you type in form fields is still masked, and request headers and bodies are never recorded. Uploaded image file names, Kubernetes cluster IDs, endpoints, and event messages are left out of recordings entirely. See Security and compliance.

Bug fixes

  • VLM models are back on the Models page. glm-5.3-flash, step-5-preview, minimax-m3, kimi-k3, and mimo-v2.6-flash were missing from the Models page, search, the playground model picker, and recently used models. Their model pages also returned a 404. They now appear in a VLM Models section. A Text/VLM toggle next to the search bar switches between categories, and the playground model picker has a VLM Models group. See Models and pricing.
  • Bare metal servers show Rebooting during a reboot. Bare metal servers that are rebooting now show Rebooting instead of Pending. See Rebooting an instance.
  • Fewer 503 errors on inference requests. When authentication was slow, some inference requests failed with a 503 even though they would have been accepted given a little more time. Hyperbolic now allows up to 1.5 seconds for the authentication check, so these requests go through. If you still get a 503, retry with backoff. See Errors.
  • No more misleading termination date in the order summary. If you cleared Auto top-up with a balance too low for the instance, the order summary said it would be terminated around the current time. Now, if your balance doesn’t cover one hour of the instance, it says you’ll add credits when you launch. Start Building then prompts you to add funds. Accounts with enough balance still see the estimated termination time. See Auto Top-Up.
Week of November 2, 2026

Bug fixes

  • Billing history no longer points to a missing Usage page. The Billing History card in Settings > Billing used to say inference usage was summarized on a Usage page, which doesn’t exist. It now says API and inference usage is billed separately and isn’t itemized in this list. See Transaction history.
Week of October 26, 2026

New features

  • Inference on the Free Tier with a verified phone. Personal accounts on the Free Tier can now use inference once they verify a phone number. Verifying also adds $1 in credits. If you already verified your phone, you’re all set. Each phone number unlocks inference for one account only. Pro accounts and Organizations don’t need a verified phone. See Who can use inference and Phone verification.

Updates

  • New error code for unverified Free Tier accounts. Inference requests from a Free Tier account without a verified phone now return 403 with the code inference_requires_verified_phone. This replaces inference_requires_pro. Verify your phone in the dashboard, or upgrade under Settings > Plan. See Errors.
  • Playground prompts for phone verification. On a model’s page, Free Tier accounts without a verified phone see a notice with Verify phone and Upgrade to Pro buttons. The Upgrade to Pro dialog now compares the Free and Pro tiers. See Model Playground.

Bug fixes

  • Clearer playground errors. When a request is refused because your balance is empty or your phone isn’t verified, the playground now says so instead of showing a generic error. See Errors.
  • Easier-to-spot Recommended badge in auto top-up setup. In the auto top-up setup dialog, the Recommended badge on the 24-hour Add enough for option is now solid purple with white text. It sits centered on the option’s border. Before, its light fill blended into the selected option. The badge next to the Auto top-up checkbox in the order summary is unchanged. See Enabling Auto Top-Up.
Week of October 26, 2026

New features

  • Choose the card auto top-up charges. The auto top-up editor now has a Pay with menu that lists your saved cards with Primary and Backup labels, plus Add a new card. Expired cards can’t be selected. A note under the menu tells you how your primary and backup cards will change when you save. The card you pick becomes your primary card, because auto top-up always charges your primary card. Nothing changes until you click Turn on auto top-up or Save changes. The menu appears in Settings > Billing, the My Instances banner, and the setup step after you launch. See Auto Top-Up.

Updates

  • Auto top-up docs from the dashboard. A Learn more link to the Auto Top-Up guide now appears in the auto top-up editor, the order summary, and the My Instances banner. See How Auto Top-Up works.
  • Clearer auto top-up messages. Warnings now say what happens “if a charge fails”, and backup card notes say the backup is charged when your primary card is declined. Messages about running out of balance now say your on-demand instances are terminated and their data is deleted. Dates more than about 6 months away now include the year. See Backup card.

Bug fixes

  • Auto top-up stays on with a single card. If your only saved card was never set as your primary card, auto top-up could turn off at its first charge. Saving auto top-up settings now always leaves you with a primary card. See Enabling Auto Top-Up.
Week of October 26, 2026

New features

  • Rental Operator (beta). Rent on-demand GPUs from your own Kubernetes cluster. Install the open-source rental-operator Helm chart, store your API key in a Secret, and apply a GPURental. The operator creates the rental through the public API, puts its SSH details on the resource, and terminates the rental when you delete it. See Rental Operator.
  • Go SDK. hyperbolic-go, the official Go SDK for the Hyperbolic API, is now open source. It is generated from the published OpenAPI specification and includes an in-memory fake API for tests. Install it with go get github.com/HyperbolicLabs/hyperbolic-go. See the package reference.
  • Redesigned auto top-up setup. In Settings → Billing, the auto top-up switch now opens a setup dialog. Set Top up when my balance drops below, which defaults to 24 hours of your current spend. Then pick Add enough for 4 hours, 12 hours, 24 hours (recommended), or a custom amount, priced from your current on-demand spend. A plan summary shows when you’ll be charged and how much, including any charge due within 10 minutes. If you have no saved card, click Add card and turn on. Auto top-up turns on once your card is saved. When it’s on, a summary card shows your settings with Edit, and turning it off asks you to confirm. See Auto Top-Up.
  • Set up auto top-up when you launch. If you can manage billing, the order summary for on-demand instances now has an Auto top-up checkbox, marked Recommended and checked by default. If you clear it, the order summary warns that your on-demand instances are terminated when your balance hits $0 and, if you’re already being billed, estimates when that will be. Launching doesn’t save any auto top-up settings. If the box was checked, the You’re live! dialog offers Set Up Auto Top-Up. It opens the same editor as Settings → Billing, priced with your new instance included. Click Not now to skip. If auto top-up is already on, the order summary shows your current settings with Edit, or Update if they may be too low for the new instance. See Auto Top-Up.

Updates

  • Auto top-up status and recommendations. My Instances now shows a dismissible recommendation to turn on auto top-up, with when your instances would be terminated. When auto top-up is on, you see Your auto top-up settings may be low if your settings leave little time to fix a declined card or would charge your card often. The order summary shows the same warning when your settings are too low for the instance you’re launching. Under Current On-Demand GPU Spend, a status line shows when your next top-up is expected, or your estimated run time when auto top-up is off. See How Auto Top-Up works.
  • New auto top-up minimums. Your threshold and top-up amount must each be at least $30. Your top-up amount must also cover at least one hour of your current spend, so your card is charged no more than 24 times a day. Before, any threshold of $0 or more and any top-up of $5 or more was accepted. The dashboard shows these errors as you type. See Auto Top-Up.
  • End-of-term emails fit your reservation’s length. Reminders that a reserved instance is about to end, terminate, or switch to on-demand now depend on how long the reservation is, extensions included. 1-day reservations get reminders 12 hours and 1 hour before the end. 2-day reservations add one at 24 hours. Reservations of 3 days or more add one at 48 hours. Short reservations no longer get a “48 hours left” or “24 hours left” email the moment they start, and every reservation now gets a final 1-hour reminder. See What happens at the end of the term.

Bug fixes

  • Auto top-up turns on reliably after you add a card. Hyperbolic now turns on auto top-up as soon as Stripe confirms your card is saved, even if you don’t return to the dashboard. Repeated confirmations no longer overwrite amounts you’ve changed since. See Enabling Auto Top-Up.
  • Reminders resume after you extend a reservation. After an extension, you now get end-of-term reminders for the new end date. Before, a reminder could be skipped if you’d received the same one for the old end date recently. See Extending a reservation.
Week of October 26, 2026

New features

  • Turn off GPU health monitoring from the dashboard. When you rent a virtual machine in a supported region, the rent page now has a GPU health monitoring switch, on by default. The order summary shows whether it’s enabled or disabled. On a running rental, the same switch is in the instance’s settings card, and changes apply within a few minutes without a restart. Before, you could only opt out from inside the VM. Opting out doesn’t change your instance, its performance, or its price. See Telemetry Hyperbolic collects.
  • Opt out of monitoring through the API. Pass monitoringOptOut: true when you create a virtual machine rental with POST /v2/on-demand/rentals or POST /v2/on-demand/virtual-machine-rentals. To change a live rental, call POST /v2/on-demand/rentals/monitoring-opt-out with rentalId and monitoringOptOut. Rentals now return monitoringOptOut, which is null when a rental can’t opt out or its setting couldn’t be read. Only standalone virtual machine rentals in supported regions can opt out. Other rentals, including Managed Kubernetes node pools, get a 400. See Opting out.

Updates

  • Rent page shows one monitoring choice per option. For virtual machines in supported regions, the GPU health monitoring switch replaces the optional Monitoring field, which didn’t apply to them. Other options keep the Monitoring field unchanged. See GPU monitoring.
Week of October 26, 2026

Updates

  • Inference API returns 402 and 403 for account denials. Requests denied because your account or Organization has no credits left now return 402 with insufficient_balance. Requests from a Free Tier personal account now return 403 with inference_requires_pro. Every other denial, such as an invalid or revoked key, still returns 401. The JSON error body is unchanged, so clients that branch on code keep working. See Errors.
  • Model Playground reopens the right dialog mid-session. If your credits run out or your plan changes while you chat, the Model Playground now refreshes your account and opens the add credits or Upgrade to Pro dialog. Before, you only saw an error line. See Plan.

Bug fixes

  • Clearer credit and plan errors in the Model Playground. When the service doesn’t send a specific error, the Model Playground now says Your credits are used up. Add credits to continue. or Inference is not included in your plan. Upgrade to continue. instead of a generic failure. See Errors.
  • Editing a Managed Kubernetes cluster keeps its settings. Adding or removing nodes, changing pools, or renaming a cluster now keeps settings the dashboard doesn’t show. This includes the cleanup that removes a cluster’s machines when you terminate it. See Node pools.
  • Security updates. We patched third-party dependencies across the Hyperbolic dashboard and platform services to resolve known vulnerabilities. No action is needed on your part. See Security and compliance.
Week of October 19, 2026

New features

  • Upgrade to Pro from a model’s lock icon. On restricted models, Upgrade Now in the lock icon’s tooltip now opens the Upgrade to Pro dialog right where you are. This works on model cards on the Models page, in the featured carousel, and in a model page’s header. Before, it sent you to the Billing page. The dialog names the model you picked and shows a Free vs Pro comparison and the $5 payment form. After your upgrade goes through, the model unlocks. If you’re signed out, Upgrade Now still takes you to Settings > Plan. See Plan.
  • Convert on-demand rentals to reserved for 1 to 90 days. POST /v2/on-demand/convert-to-reserved now accepts any whole number of days from 1 to 90 in termPeriodDays, not just the standard terms. Conversions are discounted the same way as new reservations and extensions: 1 to 6 days get no discount, and longer lengths get the rate of the longest standard term they fully cover. See Convert a rental to a reserved term.
  • Pick a custom end date when you switch to reserved. The Change Your Duration dialog for on-demand instances now has a Custom end date option next to the standard terms. Pick any end date from 1 to 90 days out, review the total price, and click Switch to Reserved to pay from your credits. See Reserved overview.
  • Backup card for Auto Top-Up. You can now keep a backup card next to your primary card. If an Auto Top-Up charge on your primary card is declined, Hyperbolic charges your backup card for the same top-up, so your instances keep running. Under Billing > Saved Payment Methods, click Add backup, or open a saved card’s menu and select Use as backup. Select Make primary on your backup card to swap the two roles. See Auto Top-Up.

Updates

  • Quote any conversion length. GET /v2/on-demand/get-reserved-conversion-quote accepts an optional termPeriodDays parameter from 1 to 90. The response adds the price for that length as requestedOption, with termPeriodDays, discountPercentage, and totalPriceCents. Without the parameter, the response is unchanged. See Get a reserved conversion quote.
  • Conversions over 90 days return BAD_REQUEST. Conversion requests longer than 90 days now fail with BAD_REQUEST instead of INVALID_TERM_PERIOD_DAYS, matching new reservations and extensions.
  • Clearer messages when switching to reserved or extending. The Change Your Duration and Extend Reservation dialogs now show specific messages, such as Already scheduled to change when a reserved term is already pending, or Your quote was updated when the price changed before you confirmed. If options fail to load, click Try again. Changing your plan or end date clears an earlier error. The action buttons now stay in view while the dialog scrolls. See Extending a reservation.
  • See when a card was last declined. Each saved card now shows Last charge declined with the date of its most recent declined charge. The note clears after the card is charged successfully. See Auto Top-Up.
  • Removing a card promotes another one. When you remove your primary card, your backup card becomes the primary. Without a backup, another saved card becomes the primary. Removing your only card turns off Auto Top-Up. The removal dialog now tells you which of these will happen.
  • Card changes retry a failed top-up. Adding a card, changing your primary or backup card, or swapping them now retries a failed Auto Top-Up right away. Each failed top-up is retried this way up to 3 times.

Bug fixes

  • Model Playground code samples work with slow models. The cURL, Python, and TypeScript samples for text models now stream the response, so text prints as it arrives. Before, long generations from slower models, such as nemotron-3-ultra-550b-a55b, could time out and fail with a 520 error or a JSON parsing error. The Python and TypeScript samples now print the HTTP status and error body when a request fails, and they also report errors returned inside the stream. See Streaming and Time limits.
Week of October 19, 2026

New features

  • Multiple SSH keys in us-east-1, us-east-2, and ca-east-3. You can now attach more than one SSH key when you launch a virtual machine in these regions, so several people can connect with their own keys. On the rent page, click Add another key after you select a key. Through the API, pass several IDs in sshPublicKeyIds. These regions now report supportsMultipleSshKeys as true. Changing keys on a running instance isn’t available in these regions yet. See Attach multiple keys to an instance.
  • Multiple SSH keys in eu-west-3, eu-west-4, and uk-southeast-3. You can now attach more than one SSH key when you launch a bare metal instance in these regions. On the rent page, click Add another key after you select a key. Through the API, pass several IDs in sshPublicKeyIds. These regions now report supportsMultipleSshKeys as true. Changing keys on a running instance isn’t available in these regions yet. See Attach multiple keys to an instance and Available regions.

Updates

  • Balance and plan errors have their own status codes. An inference request refused for an empty balance now returns 402 with the code insufficient_balance, and one refused because the plan doesn’t include inference returns 403 with inference_requires_pro. Both used to return 401, which is now only for credential problems. The JSON body is unchanged, so clients that branch on code keep working; clients that branch on status should treat 402 and 403 like 401 and not retry. In the Model Playground, either answer now opens the add-credits or upgrade dialog. See Inference errors.
  • Tidier instance details layout. On wide screens, the half-width sections on the instance details page, such as SSH Access, Port Forwarding, Node Networking, and Custom Ports, now fill the two-column grid in order with no empty gaps. The page shows only the sections that apply to your instance. See Managing instances.
  • Refreshed sign-in and sign-up pages. The login and signup pages have a new tagline, “Liquidity AI Infra for Global Compute,” and a new animated graphic comparing Hyperbolic GPU costs with AWS. The graphic also appears on the organization invitation and onboarding pages. It’s hidden on smaller screens. If your device is set to reduce motion, you see a still image instead. See Quickstart.

Bug fixes

  • Your prompts stay out of usage recordings. Hyperbolic’s product analytics session recordings now mask all text and form fields in the app, and never record request headers or bodies. Your Model Playground prompts and responses, and anything you type, are never captured. Click and page view analytics are unchanged. See Security and compliance.
Week of October 19, 2026

Updates

  • Public API banner in the app. A banner at the top of the app now announces that you can rent, reserve, and extend GPUs from your own code with the Hyperbolic API. It links to the API reference and to API Keys, where you can create a key. In an Organization, the API key link only shows if you have permission to view the Organization’s API keys. The banner shows to every signed-in user until you dismiss it, and stays hidden after that. See API keys.
  • **$1 phone verification banner removed.** The banner offering $1 of free credit for verifying your phone number no longer appears at the top of the app. You can still verify your phone number and get the $1 credit during onboarding, from an invite link, or under Settings > Plan. See Plan.
  • SSH key support flags in the API. Rental options from GET /v2/on-demand/rental-options and your rentals now include supportsMultipleSshKeys and supportsSshKeyUpdates, next to supportsCustomPorts. Check them before you pass more than one key in sshPublicKeyIds or change keys on a running instance. The dashboard uses the same flags to decide when to show Add another key, Add key, and Remove. See SSH keys.
  • Clearer expectedReservedUntil when extending a reservation. In POST /v2/on-demand/extend-reserved, send your reservation’s current end date as expectedReservedUntil. That’s the top-level reservedUntil from the extension quote, not the option’s newReservedUntil. If the end date changed since the quote, the extension is rejected. The field now only accepts UTC timestamps, such as 2025-01-31T12:00:00.000Z. See Extend a reserved rental.
  • Term lengths explained in the API reference. termPeriodDays now notes that any length from 1 to 90 days needs custom reservation lengths enabled for your account. Otherwise, use a standard term from GET /v2/on-demand/reserve-options. See Extending a reservation.

Bug fixes

  • SSH key changes on terminated rentals. Attaching or detaching a key on a terminated rental now returns RENTAL_SSH_KEY_UPDATE_NOT_SUPPORTED right away, and supportsSshKeyUpdates is false for that rental. See Change keys on a running instance.
  • Custom ports shown only where available. Bare metal rental options no longer report custom ports as supported, since new bare metal rentals can’t use them. Existing bare metal rentals that already have custom ports keep them. See Opening inbound ports.
  • Rent form keeps your instance name and custom ports. Changing another option on the rent page, such as toggling Monitoring, no longer clears the Instance Name or Custom ports you entered. Your instance name is now applied to the rental you create. See Opening inbound ports and GPU monitoring.
Week of October 19, 2026

New features

  • Upgrade to Pro from the Model Playground. Free Tier accounts now see a notice above the message box saying the model needs a Pro account ($5 one-time, no subscription). When you send a message, the Model Playground no longer sends a request that fails. Instead, an **Upgrade to Pro** dialog opens with your message, a Free vs Pro comparison, and a $5 payment form, so you never leave the page. Your message stays in the message box if you close the dialog. After your upgrade goes through, the Model Playground sends your message for you. See Plan.
  • Add funds from the Model Playground in an Organization. If your Organization has no funds left, the Model Playground shows an Add funds prompt when you send a message. Organizations with funds see no change. See Organizations.
  • Extend reservations by 1 to 90 days. POST /v2/on-demand/extend-reserved now accepts any whole number of days from 1 to 90, not just 7, 14, or 30. Extensions are still prepaid from your credits. They’re discounted the same way as new reservations: 1 to 6 days get no discount, and longer lengths get the rate of the longest standard term they fully cover. See Extending a reservation.
  • Pick a custom end date when extending. The Extend Reservation dialog now has a Custom end date option next to the 1 Week, 2 Weeks, and 1 Month choices. Select it to open a calendar at your reservation’s current end, then pick the day it should now end, up to 90 days later. The summary and Extend button show the price for that length. While a new date is being priced, Extend stays disabled. See Extending a reservation.

Updates

  • Quote any extension length. GET /v2/on-demand/get-reservation-extension-quote accepts an optional termPeriodDays parameter. The response still lists the standard terms in options, and adds the price for the length you asked for as requestedOption. Without the parameter, the response is unchanged.
  • Clearer messages when extending a reservation. The Extend Reservation dialog now explains what went wrong and what to do next. For example, if the price or dates changed since you opened it, the dialog asks you to check the updated details and click Extend again. If custom lengths aren’t available for your account yet, it asks you to choose 1 Week, 2 Weeks, or 1 Month.
  • Extend button always in view. In the Extend Reservation dialog, Cancel and Extend now stay visible at the bottom while the rest of the dialog scrolls.
  • Extensions over 90 days return BAD_REQUEST. Extension requests longer than 90 days now fail with BAD_REQUEST instead of INVALID_TERM_PERIOD_DAYS, matching new reservations.
  • Upgrade callout on the Plan page. The Free Tier card under Settings > Plan now offers an Upgrade to Pro for $5 option. The duplicate Upgrade Now button on the Pro card is gone. See Plan.
  • Recent Workloads opens on GPU Rentals. On Home, GPU Rentals is now the first tab under Recent Workloads. It opens by default if you’ve ever rented a GPU.
  • “New” tag on the Model Playground card. The Model Playground quick-start card on Home now has a New tag.
  • eu-north-9 is now contracted capacity. eu-north-9 is now set aside for customers with a capacity contract, and its pricing comes from that contract. It no longer appears in Hyperbolic’s public list of offered GPU configurations. If your account already has access, you still see eu-north-9 on the rent page. For public 8× H100 virtual machines in Europe, use eu-west-1 or eu-west-5. To ask about contracted capacity, contact sales. See Available regions.

Bug fixes

  • $5 deposit buttons open at $5. With a $0 balance, buttons that offer a specific amount, such as "Upgrade to Pro for $5” or “Top up $5+", used to open the payment form at $100. They now open at the amount shown. The header Deposit button and the Billing page are unchanged. See Billing.
  • Vision model pages load again. Opening a vision model’s page from Models could show an error instead of the Model Playground. Vision models now open in the vision playground, with image attachments and vision code samples.
Week of October 19, 2026

New features

  • Launch Managed Kubernetes from Home. Home now has a Managed Kubernetes quick-start card next to On-Demand GPU Clusters, Model Playground, and Dedicated Model Hosting. Click Launch Cluster to open the cluster launch page. Before, you could only get there from the sidebar. The card shows wherever Managed Kubernetes (beta) appears in your sidebar. See Managed Kubernetes.
  • New eu-north-9 region in Iceland. On-Demand now includes eu-north-9, which offers 8× H100 SXM5 virtual machines, the same configuration as eu-north-4. Expect instances to take 10–15 minutes to start. Port forwarding and bare metal aren’t available in this region. The region is rolling out gradually, so it appears on the rent page once it’s live for your account. See Available regions.

Updates

  • Health telemetry from virtual machine instances. Virtual machine instances created from today send Hyperbolic a small set of health metrics, such as GPU temperature, error counters, and memory use. We use them to spot failing hardware sooner and to troubleshoot faster when you contact support. Hyperbolic never collects logs, process names, file contents, or the contents of your network traffic. Collection is on by default. To opt out, create one file inside your VM and stop two services. See Telemetry Hyperbolic collects and Opting out.
  • Prorated refunds for failed reserved instances. When Hyperbolic terminates a reserved instance because it failed to deploy or start, your refund now covers only the unused time left in the current term. Before, you got the full term back no matter how much of it had already run. If the term never started, you still get a full refund. Prepaid extensions that haven’t started yet are also still refunded in full. If the term has already ended, no refund is issued. See How it works.
  • Four regions retired. us-central-1, me-central-1, ca-west-1, and uk-central-3 no longer have GPU capacity, so they’re no longer offered on the rent page and don’t accept new rentals. Existing rentals in these regions aren’t affected. See Available regions for the regions you can rent in.

Bug fixes

  • Reasoning models show their thinking in the Model Playground. Replies from reasoning models, such as GLM, no longer show up as empty messages. The model’s thinking now appears in a Thinking section above the answer. The section starts collapsed and reads “Thought for Ns” once the reply finishes. If the model uses up its Max Tokens on thinking before it writes an answer, the Model Playground now tells you so and suggests raising Max Tokens.
  • Clearer inference errors in the Model Playground. Errors that used to appear blank, or as a generic “Too many requests”, now show a readable message. When there’s no message to show, the Model Playground names the HTTP status code instead.
  • Out-of-credit and other 401 errors show their full message. When Hyperbolic rejects a Model Playground request, such as when your credits run out, you now see the message that explains what to do, followed by its error code. For example: “Your credits are used up. Add credits at https://app.hyperbolic.ai/settings/billing to continue. (insufficient_balance)”. Before, these showed a generic “The inference service rejected the request (HTTP 401).” See Billing.
Week of October 13, 2026

Updates

  • Clear messages for every rejected inference request. When Hyperbolic rejects an inference request with 401, the response now always includes an error object with a code and a readable message. The message tells you what to do next, such as checking your API key, signing in again, adding credits, or contacting support. Before, some rejections, such as running out of balance, showed nothing in the Model Playground. You can branch on these codes in your client:
    • api_key_missing, api_key_malformed, api_key_invalid, api_key_revoked: check your key under Personal API Keys. See API keys.
    • session_expired, session_invalid: sign in again.
    • account_not_found: sign in again, or contact support if it persists.
    • account_plan_unknown: contact support.
    • insufficient_balance: add credits under Billing. See Billing.
    The existing legacy_key_unsupported and inference_requires_pro codes are unchanged.
Week of October 13, 2026

New features

  • Use Models inside an Organization. When you’re working in an Organization, Models now appears in the sidebar. The Model Playground card and the inference tab under recent workloads also show on Home. Links to models, such as model cards, the model picker, and View all models, keep you in the Organization. Prices shown in an Organization are standard platform prices, which is what the Organization is billed. Favorites and recently used models stay with your personal account. See Organizations.
  • Pick a custom end date for reservations. The rent page now has a Custom end date option. Pick the day your reservation ends on a calendar, up to 90 days out. Below the calendar, you’ll see the length, discount, prepaid total, and an approximate end time in your time zone. Hover over a day to preview its price. If a longer term unlocks a bigger discount, a hint offers to move your date to it. For more than 90 days, use the Contact sales link. See How it works.

Updates

  • Inference requires a Pro plan. Inference is no longer included in the Free Tier. Requests made with a Free Tier account’s personal API key or session now return 401 with the code inference_requires_pro. To keep using inference, upgrade your plan under Settings > Plan. After you upgrade, access is restored within 30 seconds. Organization API keys and sessions in an Organization aren’t affected. See Plan.
  • New inference address for the Model Playground. The Model Playground now sends requests to https://api.hyperbolic.ai/inference, replacing https://inference.hyperbolic.ai. Chat, image, audio, and base model completions all use the new address under /v1. You don’t need to do anything.
  • Clearer error for legacy API keys. API keys issued in an older format aren’t accepted for inference. Requests with one now return 401 with an error that names it as a legacy key, instead of a generic invalid token error. To keep making requests, create a new key under Personal API Keys and use the sk_live_ key it gives you. See API keys.
  • Simpler Models page. The category toggle next to the search bar on the Models page is gone. It only offered one option, Text, so it didn’t narrow your results. Search works the same as before.

Bug fixes

  • Inference requests with an API key that doesn’t exist are now rejected right away. Before, a burst of requests with made-up keys could time out and cause valid requests to fail with 503 errors. Valid keys now keep working normally during such bursts. See API keys.
  • When a failed Auto Top-Up is retried after you add a new card or set a default card, your credits now appear right away. Before, the payment went through but your balance could take a while to update. Billing history now also shows failed Auto Top-Up charges and successful retries right away. Each retried payment is credited only once. See Auto Top-Up.
  • Hyperbolic can now check about four times as many uncached API keys at once. Under heavy load, inference requests using a newly created or recently idle key are less likely to slow down or fail with 503 errors. Requests with keys that don’t exist still return 401 quickly at higher rates. See API keys.
Week of October 13, 2026

New features

  • Use Organization API keys for inference. API keys created in an Organization now work for inference requests. Usage is billed to the Organization’s shared balance, not your personal balance. If the Organization’s balance runs out, requests are denied until an admin adds funds. See Organizations.
  • Reserve GPUs for 1 to 90 days. Reserved terms are no longer limited to 7, 14, or 30 days. You can now choose any whole number of days from 1 to 90, prepaid in full from your credits. A term gets the discount of the longest standard term it fully covers: 1 to 6 days get no discount, 10 days gets the 7-day rate, and 90 days gets the 30-day rate. See How it works.

Updates

  • Failed Auto Top-Ups retry when you update your card. If an Auto Top-Up charge fails, Hyperbolic now retries it right away when you add a new card or set a saved card as your default. Before, the charge waited for the next scheduled retry, which could take days. If the retry is declined, the charge stays on the normal retry schedule and you’re still notified. See Auto Top-Up.
  • Look up the discount for any reservation length. GET /v2/on-demand/reserve-options accepts an optional termPeriodDays parameter. The response still lists the standard terms in options, and adds the length you asked for and its discount as requestedOption. Without the parameter, the response is unchanged.

Bug fixes

  • Saving your Auto Top-Up settings, such as changing the top-up amount, no longer turns Auto Top-Up off. If Auto Top-Up was turned off this way, re-enable it from Billing. Auto Top-Up still turns off after repeated failed charges. See Auto Top-Up.
  • Bursts of concurrent inference requests that use a newly created or recently idle API key no longer fail with 503 errors. Concurrent requests for the same key now share a single authentication check. See API keys.
Week of October 13, 2026

Updates

  • More consistent SSH key errors when creating rentals. POST /v2/on-demand/rentals now checks SSH keys in a fixed order, so some requests may return a different error than before:
    • Passing more than one key in a single-key region returns MULTIPLE_SSH_PUBLIC_KEYS_NOT_SUPPORTED_FOR_REGION before Hyperbolic looks up the keys.
    • If you pass no key and have no saved keys, the request returns NO_PUBLIC_KEY.
    • If any key ID doesn’t exist, was deleted, or belongs to another account, the request returns SSH_PUBLIC_KEY_NOT_FOUND.
    See SSH keys.
Week of October 13, 2026

New features

  • See what happens when a reservation ends. The countdown on a reserved instance now says what comes next. With Reservation Protection on, it reads “Ends in 3 d 2 h, then converts to on-demand” in grey. With Reservation Protection off, it shows in red, and a tooltip names the setting. The countdown shows seconds in the last 24 hours. After the end, it reads “Converted to on-demand” or “Reservation ended · terminating”. See What happens at the end of the term.
  • Extend Reservation in the header. In the last 3 days of a reservation, an Extend Reservation button appears at the top of the instance details page. It hides once an extension is queued. See Extending a reservation.
  • Extend or upgrade from the Actions menu. The instance details page now has an Actions menu. It includes Extend Reservation whenever your reservation can be extended, and Upgrade to Reserved for on-demand instances that can be upgraded. Both open the same dialogs as the Billing card. See Extending a reservation.

Updates

  • Clearer billing actions. On the instance details page, Upgrade to Reserved and Extend Reservation on the Billing card are now buttons instead of text links. The Upgrade to Reserved button shows how much you can save.
  • Terminate moved to the Actions menu. The Terminate button on the instance details page now lives in the labeled Actions menu, which replaces the bare ⋮ icon. Instance cards keep their ⋮ menu. See Terminating instances.
  • Numeric rental IDs are deprecated in the API. Rental endpoints such as update, terminate, reboot, extend, convert to reserved, and storage volume attach now prefer the string rentalId. Pass "rentalId": "<rental-id>" instead of the legacy numeric id. Numeric ids still work for now, so existing integrations keep running.
  • Use rentalId to identify rentals in the API. Endpoints that act on a rental now document the string rentalId as the preferred identifier. This covers updating, terminating, and rebooting rentals, and attaching or detaching SSH keys. The legacy numeric id is marked deprecated in the API spec but still works, so existing integrations don’t break. We recommend switching to rentalId. See SSH keys.
  • List all active rentals in one call. rentalType is now optional on GET /v2/on-demand/rentals/active. Omit it to get your active virtual machine and bare metal rentals together in one paginated list, most recently started first. Pass rentalType=virtual-machine or rentalType=bare-metal to list one machine type, as before. With includeClusterPools=true, managed Kubernetes node pools appear when rentalType is omitted or set to virtual-machine. Bare metal listings never include them.

Bug fixes

  • After a reservation ends, the instance details page, instance list, and Home no longer show the instance as reserved with a past end date. The term badge now reads on-demand or hides, and the Billing card shows the post-reservation status.
  • You can no longer start an upgrade to Reserved while an instance is terminating.
Week of October 12, 2026

New features

  • Get help sourcing GPUs for Managed Kubernetes. The Launch Cluster page now shows a banner that connects you with our team. If Kubernetes capacity is fully booked, tell us what you need and we’ll help you find it. If GPUs are available but not the ones you want, we can source them for your cluster. See Create a cluster.

Bug fixes

  • Security updates. We patched third-party dependencies across the Hyperbolic dashboard and platform services to resolve known vulnerabilities. No action is needed on your part.
Week of October 10, 2026

New features

  • Change SSH keys on a live instance. You can now attach or detach a saved SSH key on a running or pending virtual machine rental without relaunching it. Use this to fix a launch with the wrong key, replace a lost key, or give a teammate access. Changes apply to every node without a reboot. This is available through the API today: POST /v2/on-demand/rentals/ssh-keys/attach and POST /v2/on-demand/rentals/ssh-keys/detach, each with a rentalId and sshPublicKeyId. See SSH keys.
    • Personal rentals use your personal keys. Organization rentals use the organization’s keys.
    • Any organization member can attach a key. Detaching is limited to admins and the member who created the rental.
    • A rental must keep at least one key. Detaching the last key returns CANNOT_DETACH_LAST_RENTAL_SSH_KEY.
    • Rentals that can’t change keys, including all bare metal rentals, return RENTAL_SSH_KEY_UPDATE_NOT_SUPPORTED.
    • If a change fails with FAILED_TO_UPDATE_RENTAL_SSH_KEYS, you can safely retry it.

Updates

  • Manage SSH keys from the dashboard. The SSH Access section on the instance details page now lets you change keys on eligible active rentals. Click Add key to attach a saved key, or paste a new key to save and attach it in one step. Click Remove to detach a key. Changes reach the running instance in about 20 seconds, with no reboot. See SSH keys.
    • You only see the actions your permissions allow.
    • An instance must keep at least one key. Remove is disabled on the last key, and a tooltip explains why.
    • You see a confirmation or error message after each change.

Bug fixes

  • The “SSH only available for creator” notice on the instance details page now appears only when you didn’t create the rental and can’t change its SSH keys.
Week of October 5, 2026

New features

  • Availability alerts. You can get an email when a GPU configuration you want becomes available. Manage alerts from the alerts page. Pause, resume, or delete them, and launch a rental straight from a matched alert. Alerts are grouped into active and done tabs, and you can have up to 10 active alerts. Learn more in Availability alerts.
  • Multiple SSH keys. You can now save more than one SSH key and choose one or more keys when you launch an instance. Each instance’s details page lists the keys attached to it. See SSH keys.
  • Organization SSH keys. Organizations can now share SSH keys. Members can add and view keys. Admins can manage any key in the organization. See Organizations.
  • Event codes. You can now scan an event QR code to claim free credit. Sign in or sign up, then choose your personal account or an organization to receive the credit. Some codes apply only to organizations, and you can create one during the flow. Each code can be redeemed once per account or organization, and once per person. After you redeem, you see the credit applied and your updated balance. See Organizations.
  • Instance management panel. SSH access and reboot controls now live together in one panel on the instance details page. Multi-node rentals show each node in a table, or as cards on mobile. See Managing instances.

Updates

  • Managed Kubernetes (beta).
    • Download a per-cluster SSO kubeconfig from Cluster info. See kubectl access.
    • New node group pages for compute, billing, and settings, plus a node detail view.
    • Cluster usage now appears in your billing history.
    • Hyperbolic now emails you about low balance, control-plane status, and cluster termination.
    • Refreshed layout with a Beta label in the navigation, an actions menu, and better mobile support.
    • The cluster API endpoint now appears in full in the Cluster info card, with a copy button. The Copy API endpoint item is no longer in the cluster actions menu.
    • The cluster creation wizard and order summary no longer show control-plane class or replica details. Hyperbolic manages the control plane for you.
    • Cluster bring-up progress now tracks worker provisioning, node joining, and add-on installation, with simpler status text.
    • Learn more in Managed Kubernetes.
  • Alert emails. Alert emails now include the connection type and a link to your alerts page. When you open an alert email, the dashboard tells you whether the GPUs are still available.
  • GPU count filter. GPU counts on alerts and the GPUs page now mean “N or more” GPUs. See GPU count filter.
  • Public API.
    • New paginated endpoints for active rentals and rental history, with filters for GPU type, region, and term type. The older unpaginated list endpoints are deprecated.
    • SSH key endpoints are now available, so you can list, create, rename, and delete keys.
    • New unified endpoints create, terminate, and reboot rentals of either machine type: POST /v2/on-demand/rentals, POST /v2/on-demand/rentals/terminate, and POST /v2/on-demand/rentals/reboot-instance. Set rentalType to virtual-machine or bare-metal when you create a rental. The per-type virtual-machine-rentals and bare-metal-rentals create, terminate, and reboot endpoints are deprecated but still work.
    • Endpoints to get a single rental, an active rental summary, and whether you have past rentals are now available.
    • Timestamps are now returned in RFC 3339 format.
    • GPU and storage counts in rental options are now rounded down to fixed steps.
    • New unified endpoints to create, terminate, and reboot rentals: POST /v2/on-demand/rentals, POST /v2/on-demand/rentals/terminate, and POST /v2/on-demand/rentals/reboot-instance. Set rentalType to choose a virtual machine or bare metal rental. Requests with fields that don’t apply to the chosen type now return a clear INVALID_INPUT_FOR_RENTAL_TYPE error.
    • The separate virtual machine and bare metal create, terminate, and reboot endpoints are deprecated. They still work, and each one points to its replacement.
    • The Virtual machine rentals and Bare metal rentals groups are now marked deprecated in the API reference and listed last. Their endpoints still work, but we recommend the unified rental endpoints for new integrations.
    • The current user response from GET /v2/users/me no longer includes the roles list. Use the role field instead.
  • Auto Top-Up reliability. If an automatic top-up is interrupted, Hyperbolic now retries it. New safeguards make sure a retry never charges you twice. See Auto Top-Up.

Bug fixes

  • Credits that Hyperbolic adds to your personal balance now unlock renting right away, the same as a card purchase. You no longer need to pay by card first. See Account management.
  • Instance status no longer flips away from Running while a node reboots.
  • Custom ports now show on the instance details page for all eligible instances.
  • Managed Kubernetes node details now show the node’s public IP. Before, the field was always empty. Use this IP to reach workloads on the ports you open. See Managed Kubernetes.
  • Retrying a Managed Kubernetes cluster launch after a network or server error no longer creates a second, separately billed cluster. The retry resumes the original launch. See Managed Kubernetes.
  • Managed Kubernetes node details now show the size of each NVMe disk. Before, every disk showed the node’s total disk capacity. See Managed Kubernetes.
  • Managed Kubernetes cluster and node group names are now validated on every request. Opening a cluster page with a malformed name now shows a not found page.
  • A temporary error from the cluster provider no longer marks a running Managed Kubernetes cluster as ended. When the provider removes a cluster, billing now stops at the time the deletion started.
  • The instance details page now shows each node’s internal IP for rentals where it was missing.
  • Availability alerts now match on offered configurations, not free GPU counts, and no longer count the same GPUs twice.
  • GPUs that just sold out no longer appear as available.
  • The login modal now closes automatically when you’re already signed in.
  • Signing in from a link now returns you to the exact page you opened, including its query parameters.
  • Fixed a 404 on the Support link in Private Cloud navigation.
  • Fixed the layout of past instance cards so the time sits under the price.
  • API error responses no longer include internal debugging details.
  • Terminating a bare metal rental that doesn’t exist or has already ended now returns RENTAL_NOT_FOUND instead of a generic lookup error.
  • If a new bare metal rental fails or ends while it’s provisioning, the launch screen now tells you. Before, it kept loading. See Managing instances.
  • An instance’s running time no longer shows a wrong value or flickers when the page first loads. It now shows a placeholder until the time is calculated, then counts up every second. See Managing instances.