Skip to main content
Hive can answer with more than one AI model. Most of the time you do not need to choose: the workspace default is set by your admins. When you do want a particular model — a faster one for a quick lookup, a more capable one for a long analysis — pick it from the model picker in the Ask composer.
Hive does not publish a fixed list of models. The picker only offers models your workspace can actually use: those whose provider is configured for your deployment, that your admins allow, and that can serve a chat turn. The list changes as models are added and retired.

The model picker

Open the picker in the composer (the Hive default button — Hive auto when your workspace routes automatically, or Hive followed by the model’s name once you have chosen one). It shows:
  • Workspace default — uses your workspace and personal settings. The description names the model that will answer.
  • Hive Auto — shown instead when your workspace’s default is automatic routing: “Use automatic routing configured by your workspace.” Hive picks a suitable model for each turn. When your admins have set up Auto mode but the default is a named model, Hive Auto is offered as a choice of its own, below the default: pick it to have Hive route your next questions, short prompts to the Fast model and complex work to the Powerful one. It stays chosen for the chat until you pick another. Without Auto mode it is not offered.
  • Or pick a model — the specific models available to you, each shown with its provider’s icon. In Ask, a Balanced model shows about 2.7× credits and a Frontier model about 7× credits: how many times the credits of a typical answer on Hive Auto it typically uses (see Model tiers and credits). Open a model to see its description and its Credit use: how many times the credits of Hive Auto it typically uses, on a meter from Low to Very high. How is this calculated? explains that it is measured against a typical answer on Hive Auto: about £0.04 on Business, where credits are pounds, and about 27 credits on an earlier plan; the actual cost depends on the work. The multipliers are the same on every plan. Hive never shows a provider’s own prices.
“Workspace and billing rules apply to every turn” — your choice never bypasses your workspace’s credit limits.

The same picker everywhere

Every place you choose a chat or agent model in Hive uses this one picker, so it looks and works the same:
  • the Ask composer, and the Start a new chat in … box on a project page, which starts the new chat on the model you pick;
  • Agent Studio: the Agent model, including when you describe a new agent;
  • Workflows: a pinned model on an agent step on the tool-calling agent runtime;
  • Settings: your personal Default model, and the workspace Default chat model, Default agent model, Default workflow model, Default background model, and the Fast and Powerful models under Auto mode.
Image generation models have their own list. In Ask’s Generate image mode the composer’s picker shows only them, for that message; your default image model stays in Settings › Preferences. A workflow step on a workspace that is not yet on the tool-calling runtime offers only the runtime’s one model. Each place offers the choices that make sense there. Ask, your personal default and the workflow composers offer a default row, so you can leave the choice to Hive. Agent Studio always pins an explicit model. Auto appears as its own choice for your personal default and the workspace’s Default chat model, and, once Auto mode is set up, as Hive Auto in the Ask composer and a project’s Start a new chat in … box. In Agent Studio and Workflows the picker lists the models your workspace allows for agent work, which are priced models. Unavailable models are explained there as Cannot run agent work. rather than Cannot serve an Ask turn. With the keyboard, Tab moves through the list, Left arrow opens a model’s details and Right arrow returns from them, and Escape closes the picker. The composer’s picker opens upward, above the composer. When there is not enough room above it to show the list, as on a new chat where the composer sits in the middle of the page, it opens downward instead.

Models Hive cannot price

Every turn on Hive’s keys draws credits for the model that answers, so Hive runs only models it can price there. Custom models, models with no published price, and the most expensive -pro models run only on your organisation’s own model key (see Your own model key). Until your organisation adds its key, they are hidden from the Ask picker and from Agent Studio rather than offered and then refused; once it has, they appear. If your personal default model, or a model Auto routes to, is hidden this way, Ask uses the workspace default instead; choose a new personal default whenever you like.

Models that are retired

Model providers retire models over time. Hive checks every hour which models it can still use, and a model its provider has stopped serving is removed from every model list: the Ask picker, Agent Studio, the workspace model settings and allow-list, and the model lists AI clients read over MCP. If the provider serves the model again, it comes back within the hour. Models your organisation runs on its own model key with a provider Hive does not route are not affected. While a model of the same tier is still offered, nothing set up on a retired model stops working, and nothing you saved is changed:
  • A workspace default (chat, Fast, Powerful, agent, workflow or background), your personal default, and a saved agent pinned to a retired model run on the first model your workspace allows from the same provider and in the same tier — for example, an agent pinned to a retired Frontier OpenAI model runs on the next Frontier OpenAI model in the list. If that provider has none left in the tier, Hive uses the same tier from another provider. Hive never moves a choice to another tier: if your workspace allows no model of that tier, the choice is not available until an admin picks another model.
  • Hive’s own defaults follow the same rule: a chat where nobody chose a model, quick answers, chat titles and summaries, and background work run on a model that is still served.
  • A custom model your organisation added on its own key is never moved.
  • Each answer and agent run records, and is billed as, the model that actually ran.
  • The settings pages show the model that is used today. Saving them keeps that model.
  • If the retired model is served again, the original choice applies again automatically — unless an admin saved those settings while it was retired, which keeps the model shown when they saved.
A model a request names directly — for example from a browser tab left open since before the model was retired — is refused rather than swapped; choose another model in the picker.

Unavailable models

Admins also see models that exist but cannot be used, with the reason: If no model is available at all, Ask still answers with its default.

What your choice affects

An explicit model choice also governs the internal planning and retrieval steps for that turn, not just the final wording. Each answer records the model that actually produced it, and names it at the end of the line above the answer (“Hive · read HubSpot, Gmail · 6s · GPT-5.5”). The name shows as soon as Hive starts answering, including the model Hive picked when you left the choice to Hive, and changes if Hive moves the answer to another model before it finishes. To compare models on the same question, use Regenerate under the latest answer: the button asks again on the model that answered, and the arrow next to it opens the same model list as the composer so you can ask again on another. The new answer is added below; the first one stays in the chat, and in the conversation Hive reads when it answers again. Regenerate is not offered while an answer still has an approval waiting for you. That includes reading your connected apps. When Ask looks something up in a connected app, the model you chose decides which reads to make and pulls the relevant items out of what comes back, and those steps use credits like the answer does. A more capable model can make these reads slower and more expensive; a faster one makes them quicker. When you have not chosen a model, the workspace default is used. Agents read connected apps on the model selected for the agent, and background work such as Signals uses the workspace’s Default background model, which is always an Efficient model other than Claude Haiku. When that is not set, it uses Hive’s Efficient default. If a provider retires that model, Hive moves background work to its next Efficient model.

When a model fails

When the model working on an answer fails before you see any of it — the provider is down or overloaded, it stops responding, or it returns an error — Hive tries that step once more on another model, so you get an answer rather than an error:
  • After a Fast model, Hive uses your workspace’s Powerful model, or Hive’s own default Powerful model when none is set.
  • After a Powerful model, or a model you or an agent picked, Hive uses the next model in Hive’s default order that your workspace allows and you may use. When your organisation’s own model key is in use, it stays on that key’s provider, and a custom model of your workspace is never swapped.
  • A rate-limited or overloaded model is first tried again once on the same model, after a short wait, as before.
Hive switches at most once per answer, and only while nothing has been shown or done yet: once part of the answer has appeared, or a step has changed something (a file, an approval card), a failure ends the answer as before. A quick answer stays on an Efficient model. Hive also never switches to a model on which the answer would have needed your confirmation of its cost when the model you chose did not, or would cost more than the estimate you confirmed; when no such model is left, Ask tells you what went wrong. A request a provider’s safety filter blocked is not sent to another model. The model name above the answer shows which model answered, and the answer uses credits on that model like any other. If the second model fails too, Ask tells you what went wrong. Agents you ask in a chat follow the same rule. Apart from that, while you have credits, Hive never swaps the model you chose. On Business, when you have used your own credits and any extra usage the workspace allows, the composer says so, and answers stop until an Owner or Admin turns on extra usage or raises your top-up limit, or next month’s credits arrive. On an earlier plan, when your workspace is out of credits, or you reach your own monthly credit limit, the composer says so, and an Owner or Admin can add credits, or raise your limit, in Settings → Billing. Until then, as a member on a Pro or Max seat you keep 10 quick answers a day and 200 a month. Each is a short answer with one round of searches of your workspace’s knowledge, Business Brain and Signals, and no web search or files. They run on Hive’s Efficient model whatever model you chose, on an Efficient model your workspace’s allowlist admits: Hive’s own when it is allowed, otherwise another allowed Efficient model. A member on an Approver seat has no quick answers. If the allowlist admits no Efficient model, quick answers don’t run, and Ask says so (for a quick answer in Slack or Teams, Hive’s reply does) rather than using the model you picked or the workspace’s default, and the question isn’t counted. When your organisation’s own model key is in use, pick a model from that key’s provider. See Model tiers and credits.

Photos and images with any model

Every model can work with your photos. A model that sees images gets each picture you attach or choose, and each picture on a slide, as the image itself. A model that can’t see images gets Hive’s description of it instead: what it shows and any text in it, written by Hive’s vision model and labelled as such. Hive uses the description it made when the file was added; when there is none yet, it makes one for that answer on its Efficient vision model, which uses credits like the answer’s other steps, and reuses it after that. See How Hive reads the files you choose.

Models for agents

The Agent model picker in Agent Studio lists the same priced models as Ask. If an agent was set up with a model that is no longer listed, the picker shows it as unavailable and its runs stop with a message asking you to choose another model. Only a model that can’t take a step can change an agent’s model during a run. A step can’t be taken when the model is rate-limited, overloaded or unavailable, either before the step is sent or because the provider turned the step away without doing any of it (a 429, 503 or 529):
  • With When credits or limits run out set to Switch to a faster model, the run moves to Claude Haiku first, if your workspace allows it.
  • After that, or straight away when the agent is set to Stop, a run you or a teammate started, a scheduled run and a workflow’s Saved agent step move once to the model Hive picks the same way as for Ask (see When a model fails): a model your workspace allows and the person the run is for may use, on your organisation’s own key’s provider when that key is in use.
The rest of that run stays on the new model, and each step is charged at that model’s rates. The step the provider turned away is charged nothing, and the run’s own credit limit still applies to every step, so a move never lets a run spend more than its limit. A step the provider started working on and then failed is not sent again on another model. Shared agent links move only to Claude Haiku, as set above. Running out of credits, like every other workspace or agent limit, always blocks.

Model tiers and credits

Every answer uses AI credits: on Business your own, which are pounds spent at what the work costs Hive (1 credit is 1p); on an earlier plan, the workspace’s shared pool. How much depends on the model that answers and how much it reads and writes. Hive groups models into three tiers: Hive Auto picks a model for each turn: short, routine questions go to the workspace’s Fast model and longer or harder ones to its Powerful model. The Fast model answers most questions, so it must be an Efficient model, or a custom or unpriced model that only your organisation’s own key runs: see Hive Auto’s Fast model. Hive Auto decides by a fixed set of rules, read in this order, and the first that applies wins.
  1. Some turns always go to Powerful:
    • anything that makes or changes a file — a deck or slides, a spreadsheet, Excel workbook or sheet, a Word document, report or letter — whether you ask in your own words (“Make a 6-slide year-end review deck for Acme”, “Put these statements on one Excel sheet”) or pick the Document, Presentation or Spreadsheet mode, and any change to a file open beside the chat;
    • a chart, asked for in words or with the Visualize mode;
    • a request with several steps (“Pull my overdue invoices, then email each client, then log it”);
    • a turn while files are in play that is not a short lookup question: “Summarise these statements” or “Translate this” goes to Powerful, and “What is the total on this invoice?” does not. A short lookup is one line of up to 120 characters that opens with a question word such as “what”, “who”, “when”, “is” or “how many”, and has none of the words that make it a reading job, such as “summarise”, “compare”, “review”, “explain”, “check”, “every”, “all the”, “risk”, “issue”, “clause”, “change” or “difference”: “What is the total of all the invoices?” goes to Powerful. Files are in play when the answer reads them, not only when they are sent with the message: a follow-up in a chat with Library files selected, or after you attached files earlier in it, counts too.
    Saying “quick” does not move these to Fast, except a turn with files in play. Files you send with no message get a brief summary of each file on Fast, because Hive asks for that summary “briefly”, unless you picked a mode that makes a file or a chart.
  2. A speed you ask for in your own words near the start of the question: words such as “quick”, “brief”, “short answer” or “tl;dr” go to Fast, and words such as “thorough”, “detailed”, “comprehensive”, “in depth”, “deep dive” or “step by step” go to Powerful. A question that asks for both counts as neither.
  3. A question longer than 1,200 characters, one with a code block (between three backticks), or one that uses any of these words goes to Powerful: “analyse” (or “analyze”), “architect” or “architecture”, “compare”, “debug”, “implement”, “refactor”, “reason” or “reasoning”, “research”, “trade-off” (also “trade-offs”, “tradeoff” or “tradeoffs”) and “multi-step”. Only those exact words count: “analysis”, “compared” or “debugging” does not.
  4. Everything else goes to Fast, so “What is 2+2?” or “Who invented the light bulb?” is answered on the Fast model.
Hive Auto makes no AI call to decide, so the choice adds no noticeable wait before your answer starts and is not charged. The wording rules read English only. In another language the choice is mostly made by the mode you pick, the files in play, the question’s length and any code block: in a chat with files a short question in another language usually goes to Powerful, and without files a short but hard one usually goes to Fast. Pick a model yourself when a question needs a particular one: Hive Auto’s rules apply only when Hive Auto is chosen. The model name above the answer shows which model answered. On Business a typical answer on Hive Auto costs about 4.4p and a short one about 0.6p, so £100 buys about 2,200 typical answers. On an earlier plan a typical answer uses about 27 credits (27p) and a short one about 4, and £100 of credits buys about 370 typical answers on Hive Auto, or about 830 on an Efficient model. Typical cost per action on Business (actual cost depends on what the answer reads and writes): Typical credits per action on an earlier plan: On Hive Auto, Hive estimates deep research at Balanced cost (the Balanced column above). The research itself runs on the model Auto picks for your question, as with any Auto turn: your workspace’s Fast model for short, routine questions and its Powerful model for longer or harder ones. Any single action Hive estimates above £5 (500 credits on an earlier plan) waits for you to confirm it in Confirm a large action before it runs, which shows the estimate in pounds on Business and in credits on an earlier plan. On Business no single answer or research run reaches £5: even exhaustive research on a Frontier model is about £3.40. On an earlier plan, exhaustive research and standard research on a Frontier model ask first.

Where each tier can run

When credits run out

On Business, answers stop when you have used your own credits and any extra usage the workspace allows: see When credits run out. On an earlier plan, when the pool is empty, or you reach your own monthly limit, answers that use credits stop. As a member on a Pro or Max seat you keep 10 quick answers a day and 200 a month, short answers on an Efficient model with no web search or files, and the composer tells you how many are left. They start as soon as what is left can no longer pay for a full answer on the model you chose, so the last few credits never leave you without an answer. When the workspace is out of credits on an active paid plan, Owners and Admins buy a 5,000-credit top-up in one click from the composer (the button shows the price), or choose another pack from the Add credits link or under Settings → Billing → Credit top-ups. If you come back before its payment has gone through, the composer says that top-up is still on its way, with a link to its invoice, and buys another only if you choose Buy another anyway. A card payment your bank asked you to confirm and that you left unfinished is said to be waiting for confirmation instead, with a link to finish paying it; Buy another anyway then cancels it. A top-up does not lift a member’s own monthly limit: at that limit the composer’s link is Raise your limit in Settings → Usage, where an Owner or Admin changes the member’s cap under Member caps, or the cap for everyone in Spend controls. When the workspace is read-only, the composer says only “Sending is off while the workspace is read-only.”: the banner at the top of the page gives the reason and, for Owners and Admins, the fix (Choose a plan after a trial or subscription ends, or Pay with the amount when a payment is overdue). Other members are told to ask an Owner or Admin. A member without a seat cannot use Ask at all until an Owner or Admin gives them one, and a member on an Approver seat can ask only about items routed to them, in the Hive app, 10 questions a calendar month, paid from the pool: when it is empty, they cannot ask until there are credits again.

For admins: model settings

Owners and admins manage models under Settings → Workspace → Models.
1

Set the defaults

In the model picker, choose the Default chat model (or Auto for automatic routing), plus, under Auto mode, the Fast and Powerful models Hive uses for lighter and heavier work. Fast offers only Efficient models and custom or unpriced models that only your organisation’s own key runs (see Hive Auto’s Fast model). Leaving the default chat model on Deployment default leaves the choice to Hive. Then choose the defaults for other kinds of work (see Defaults for agents, workflows and background work). Choose Save model defaults.
2

Restrict the workspace (optional)

Under Select the models you allow in your workspace, pick the models your people may use and choose Save for workspace. Each model is shown with the same name and provider logo as in the model picker. An allowlist holds up to 50 models.
3

Add a custom model (optional, Enterprise)

On Enterprise, Add custom model registers a model by name and ID, and whether it Supports image input. On other plans the page says custom models are not included. A custom model runs only on your organisation’s own model key.

The Fast model must be Efficient

Hive Auto’s Fast model answers most questions, so it must be an Efficient model, or a custom or unpriced model that only your organisation’s own key runs. Claude Haiku counts as Efficient here. A -pro model also runs only on your own key, but it is priced as Frontier, so it can’t be Fast. The Fast list offers only the models allowed, and saving a Balanced or Frontier model as Fast is refused with “Auto’s Fast model answers most questions, so it must be an Efficient model, or a custom or unpriced model that only your organisation’s own key runs. Choose an Efficient model for Fast.” If your workspace saved a Balanced or Frontier model as Fast before this rule, Hive Auto does not use it. It answers those turns on your workspace’s Efficient model instead: the Default background model, else Hive’s Efficient default, else another Efficient model your allow-list admits, kept to your key’s provider when your organisation uses its own model key. Your saved setting is not changed. Under Fast, Settings → Workspace → Models shows “[model] isn’t an Efficient model, so Hive Auto answers on your workspace’s Efficient model instead. Choose an Efficient model for Fast.”, and Save model defaults is refused with the message above until you choose an Efficient model for Fast. Hive Auto is offered only when Fast and Powerful are both set, and both the model its Fast turns run on and the Powerful model are available to the member. A Balanced or Frontier Fast with no Efficient model to stand in for it (your allow-list admits none, or none from your own key’s provider) leaves Hive Auto unavailable until you choose an Efficient model for Fast or allow one. When your Powerful model can run, so Fast is the only reason, the note under Fast then says “[model] isn’t an Efficient model, so Hive Auto is unavailable until you choose an Efficient model for Fast.” When Hive Auto is off for another reason too, such as no Powerful model set, or while the list of models is loading or fails to load, the note states the rule only, with the same sentence as the refusal above.

Defaults for agents, workflows and background work

Beside the Default chat model, admins can set a default for each other kind of work. Each one applies only where nothing else chose a model: The agent, workflow and background defaults list the models your workspace allows for agent work, and Auto is not offered for them. A default never overrides a model someone chose: if a model an agent or a workflow step was set up with stops being available, its runs stop with a message asking for another model. Only work that relies on a default falls back to Hive’s own default model when that default can no longer run. A signal that is prepared again keeps the model that first prepared it while your workspace still allows it.

Your own model key

On Team and Enterprise, an Owner or Admin can store the organisation’s own OpenAI or Anthropic key in the Keys section of Settings → Workspace → Models. Members’ Ask, agent and workflow model work then runs on your provider account, and only on that provider’s models. Hive still uses 1 credit from the pool for each model call it orchestrates, and web search, knowledge search and the other services Hive runs on its own accounts use credits as usual. Signals and other background work stay on Hive’s Efficient model. Quick answers run on your key provider’s own Efficient model, and a quick answer uses no credit: Hive waives its fee for it. -pro models run only on your own key, and so does any provider model Hive has not priced yet. Custom models (one your workspace added with Add custom model, such as a fine-tune) come with Enterprise and also run only on your own key: on Team a call to one is refused with Custom models come with Enterprise, and the Models page says the custom models already added there do not run on your plan. If your provider refuses the key, for example because it is out of quota, the work stops and Owners and Admins are told. Hive never moves it to its own key. On a plan without your own key (the trial or Starter), a stored key is not used at all: model work runs on Hive’s models and uses credits, agents set to a custom model do not run, and the Keys section says so. See AI and models.

How the allowlist behaves

  • An empty allowlist means no restriction: everyone can use every model the deployment offers, and new models appear as they are added.
  • Once you restrict the workspace, new models are never added automatically — you opt each one in. This makes the allowlist a cost, data-residency and vendor-policy control.
  • A model you allowed that later stops being offered is simply ignored; it never widens access.
  • Quick answers run only on an Efficient model the allowlist admits. An allowlist with no Efficient model on it means none run.
Vetting AI vendors? Restrict the workspace to the providers you have approved. See AI and models for how Hive handles model traffic.

Models elsewhere in Hive

  • Your personal default — under Settings → Account → Preferences, Default model chooses the model a new regular chat starts on: Workspace default, Auto (when your workspace has Auto routing set up), or a specific model.
  • Workflows — Describe and the workflow assistant let you choose the model the workflow assistant drafts on. See Which model drafts. Agent and condition steps can use Workspace default or a pinned model. Web search and Generate image steps show Configured model and always run on the model the deployment’s web search or image service is configured with; a step saved earlier with another model ID shows a notice with Use the configured model, and deploy refuses any other model. An agent step on Workspace default runs on the workspace’s Default workflow model when one is set, and on Hive’s Efficient model otherwise. On the tool-calling agent runtime, an agent step can pin any model your workspace allows for agents. See Workflow steps.
  • Agent Studio — each agent pins its own model; Auto is not offered there. (Hive Auto in Agent Studio is an Inputs option, an alternative to Prompt for agents without a typed input contract, not a model.) See Build an agent.

AI and models

How model requests are routed and what is sent.

Workspace settings

Every admin setting in one place.