Skip to content

Models and providers

Player One works with a wide range of AI providers, and which of them your agents run on is yours to decide. You connect accounts, where an account is one connection to a provider, either a subscription you sign in to or an API key you paste, and each account brings a set of models with it. A model is the brain that actually answers. Player One picks among the models your accounts offer, and you can overrule that choice at any level, from a single conversation to the whole installation.

Connecting an account

A consumer subscription is a plan you already hold, such as a ChatGPT plan: you connect it by signing in, pay a flat monthly fee, and live with the provider's own limits, so when you reach one you wait rather than pay more. An API key is billed per token instead: you paste the key, pay for what you use, and have no ceiling, which suits work that has to run whatever the hour.

Most people end up with both, and an account can be added or removed whenever you like. The connect list shows the providers this installation can reach, plus a custom option for a service of your own, connected by address and key. Each connect screen carries a tick box confirming that the provider handles your data under its own terms, and the connect button stays grey until you tick it. Tokens, cost and caching weighs the two ways of paying, and Connect your AI accounts has the steps.

Amazon Bedrock is on the list too, billed per token at Bedrock's own prices. You connect it with a Bedrock API key and choose the region the key was made in: US East (N. Virginia), US East (Ohio) or US West (Oregon). The models it lists depend on that region, and the three US regions list the most. If Player One runs on a machine of your own that already carries AWS credentials, Bedrock appears by itself and needs no key.

The Providers page

Admin > Providers (/admin/providers) shows one card per connected account: how it is signed in, which plan it is on, how much of it you have used, and a Test button that checks that one account rather than a healthy neighbour. Give each connection a short name and it is told apart everywhere it appears. Cards and model pickers label an account by how it is paid for: API for a key billed per token, Subscription for a plan you sign in to, and Included for what comes with Player One. Update routing opens your assistant in the side panel to go through what your agents use now and change it once you confirm.

Accounts on the same provider are grouped together, and the cards are dragged into the order you want. That order does two jobs. Across providers, the first one's best model is what answers when nothing else has been chosen. Within a provider, it is the failover order: the account tried first, and the one that takes over when it cannot.

You can also park a single account, for an hour or until you turn it back on. It stays signed in, nothing routes to it, the others carry the work, and the card reads a red Off until you lift it.

Choosing a model

For one conversation, the chip in the composer's control row names the model the chat is running on. Click it to pick another, or type /model. That choice sticks to the chat and beats everything else. Beside it, on models that allow one, an effort level sets how hard the model thinks before it answers. When the chat has not set a level of its own, the chip shows the level the route gives that model, tagged route, and the picker's rows show the level each model would land on, so a swap is made knowing what it brings. Switching the chat to a different model clears a level you had set, so the new model starts on its own route's level; pick a level together with the model to keep one.

For one agent, open it from Manage > Agents (/agents), go to its Advanced tab and set its models there. Every chat with that agent starts on those unless you move it. See Agents.

For everything else, Admin > Agent Models (/admin/models) holds the default chain: the first model answers and the rest are its fallbacks, in order. Edits save themselves.

Click a model in the chain and its information panel opens. It grades the model for power, cost and speed from 0 to 10, where 10 is the most capable, the dearest and the fastest, so a high cost grade means expensive. It also has an Ignore control, which makes automatic routing skip that model everywhere on this installation while leaving it yours to pick by hand. A row says why it will not serve: a red Ignored or Off until for something you chose, an amber Cooling until or Unavailable for something that clears itself.

Effort on every model entry

Every model entry in a chain carries its own effort level, wherever that chain is: the default chain, a routing rule, an agent's own models, or a route set for one chat. Each row has a level select beside it offering Default, which leaves it to the installation's effort settings, the model's own levels, and Off where the model can genuinely switch reasoning off. The level belongs to that entry and follows it: pair your strongest model at high with a cheaper fallback at medium, and when the fallback answers it runs at medium. A level the model cannot express stays selected with an amber runs at chip naming what it will actually run at.

Say "put the coordinator on our strongest model at high, with the cheaper one at medium behind it" and your assistant writes the rule with a level on each entry and simulates it, reading which levels each model accepts before naming one.

Effort, rules and simulating

The Effort tab (/admin/models/effort) sets how hard models think by default, for all of them or one at a time. The order a turn's level is decided in: the chat's own level wins, then the level on the routing entry that names the model, then the agent's own default, then the per-model and global settings on this tab, then the model's catalogue default. Last of all the model's own list clamps anything it cannot express.

The Advanced Routing tab (/admin/models/routing) sends particular kinds of work to particular models, picked from named scenarios rather than written as code: coordinators, job workers, coding, research, voice calls. Beneath your rules, Default rules lists the built-in ones for background work, such as chat titles, file summaries and ranking your agents' search results, each on a cheap or fast model. A rule of your own for the same job wins, and Override starts one.

The Simulate tab (/admin/models/simulate) answers "which model would take this?" without sending anything, which is the honest way to check a rule you have just written. You describe the situation, such as the agent, the channel or the kind of task, and it names the chain that would answer and what decided it: one of your rules, the default chain, the built-in default, or a pick by need. Beneath the chain it shows an Effort block: one line per entry with the level it would run at, where that level came from (the entry's own setting, or which default), and the clamp when an entry asks for more than its model has.

When an agent searches the web it uses a search service, and the Web Search tab (/admin/models/web-search) shows which.

  • Automatic is the default. It uses the first search service that has a search key of its own, from Brave Search, Gemini with Google Search, Kimi, Perplexity, Firecrawl, Exa and Tavily, and with none of them, DuckDuckGo, which needs no key and costs nothing. The card names the service in use now.
  • Below it are the search services that run a model of their own, such as Gemini with Google Search. Pick one to use it instead of Automatic.
  • A search service needs its own search key, even when the same company is connected for chat on Providers; such a row reads "Connected, but web search needs its own key". Ask your assistant to add a search key: it asks for it in a secure prompt and keeps it in the vault.

A connected AI account never chooses a search service for you. Kimi can search on your connected Kimi API account with no separate key, but only once you ask your assistant to switch web search to Kimi: the tab does not list it, and Automatic never picks it because of an account. A Kimi Code account does not work for search.

Each paid search service charges in its own way, per search or per token, on its own account.

Image models

Pictures are handled by different models from the one that writes your text. Admin > Image Models has two tabs. Generation (/admin/images) picks the model that draws new images, from the image-capable models your connected accounts offer, or leaves it on Automatic and lets Player One choose; images are billed per picture rather than by the word. Vision (/admin/images/vision) picks the model that describes an image an agent is sent when its own chat model cannot see images. A chat model that can see reads the image itself.

Other model pages

Three more pages under Admin > Models each choose the model for one kind of job.

  • Embedding Models (/admin/embedding-models) picks the model that turns text into embeddings, so text can be compared by meaning. Changing it means everything already embedded is embedded again, which costs tokens, so change it deliberately. It does not change what your agents find when they search.
  • Voice Models (/admin/voice) picks the models that speak, narrate and transcribe: see Voice.
  • Judgement Models (/admin/judgement-models) picks the model your agents use for quick typed judgements, such as which option fits a message, answered with a probability rather than prose. It needs a judgement provider connected on Providers.

When something runs out

Nothing stops when one account reaches its limit. Player One tries every other usable account on that provider first, each under its own name, and only then moves to the next model in the chain. Where that happens mid reply, a card in the thread names the model that ended up answering, and opening it shows what was tried and why each attempt failed. A model that no longer exists on its provider is skipped without pausing the account, and when nothing in the chain could answer the card names each real cause rather than a vague one.

A sign-in that has died for good is different, because it will not come back on its own. The account is marked Signed out on its card, its meter in the status corner at the top right is crossed out, an alert appears on your Dashboard (/dashboard), and a note arrives in your main chat. Signing in again from the Providers page is the only fix.

A sign-in that keeps failing without the provider saying it is dead is told to you once, after a day. A note arrives in your main chat naming the provider and the account, saying how long the renewal has failed, and saying that Player One keeps retrying every hour. If it has not recovered, sign in again from the Providers page. A Claude account whose organisation does not allow subscription sign-in says so on its card, and signing in again will not help: add a Claude API key, or connect an account your organisation allows.

New models arrive by themselves

Player One's list of models is kept current for you. Your installation checks for a newer list about every half hour, so a new model, a new price or a retired one reaches you without waiting for an update. When there is news, a Model updates window shows it once, on the next page you open, and once read it stays read on every device. Got it closes it, and Update routing opens your assistant in the side panel to review your model routing with you. The same button is on the Providers page. Every model Player One works with, with each provider's list prices, is on 1p.dev/models.

When a model is replaced by a newer one, the choices you saved for it move to the newer model on their own: your default chain, your routing rules, an agent's own models, and a chat's model on its next message. A newer choice you make yourself always wins. A model withdrawn with no successor is dropped from what you saved and shows as Unavailable in a chain. A paused model is hidden and skipped, and a choice saved for it falls back to that provider's recommended model.

Recent changes

Recent additions to the catalogue, as of early November 2026, include:

  • Claude Haiku 5.5, which takes over the Haiku tier from Haiku 4.5, alongside Claude Sonnet 5.5 and Claude Opus 5.5, on Anthropic and several other providers. Claude Fable 5.1 is Anthropic's most capable model, at about two and a half times Opus 5.5's price, so point it at the work that earns it.
  • GPT-6.1 Sol, which OpenAI's own API now recommends in place of GPT-6 Sol at the same price, and GPT-6.1 Terra, a cheaper tier, on OpenAI and several other providers. GPT-6 Sol stays for coding and everyday work, GPT-6 Luna is the efficient one for high-volume work, and GPT-6 Astra is the dearer top tier.
  • Amazon Bedrock as a provider, with Claude and a few models from other makers: see Connecting an account.
  • Grok 4.7 on xAI, and new models from Xiaomi MiMo, Alibaba Model Studio, the Volcengine Coding Plan, NVIDIA, Mistral, Venice and MiniMax.
  • Eleven v4 and Eleven v4 Turbo on ElevenLabs, for spoken replies and narration: see Voice.
  • GPT Transcribe on OpenAI, now the model that turns voice messages and live-call speech into text when OpenAI is the one chosen. Grok Voice Transcribe 2.0 on xAI and Voxtral Mini Transcribe 2 on Mistral join it, GPT Realtime 2.1 arrives for live calls, and Grok TTS for spoken replies.
  • Gemini Embedding 2 on Google and Voyage 4 on Voyage, for embeddings.
  • Player One Provider, included with Player One, which transcribes your voice messages and does nothing else.

Saved choices that moved on their own:

  • Claude Sonnet 5 to Claude Sonnet 5.5, except on GitHub Copilot and OpenCode Zen, where Sonnet 5 stays.
  • Claude Haiku 4.5 to Claude Haiku 5.5.
  • Claude Opus 5 and Claude Opus 4.8 to Claude Opus 5.5, and Claude Fable 5 to Claude Fable 5.1.
  • On Amazon Bedrock, the Global versions of older Claude models to Opus 5.5, Sonnet 5.5 and Fable 5.1. Any other Bedrock choice the catalogue does not carry stays saved but is refused, with the model to pick instead, until you choose one.
  • GPT-5.5 and the GPT-5.6 models to GPT-6 Sol and GPT-6 Luna, except GPT-5.6 Terra on OpenAI and Vercel AI Gateway, which moves to GPT-6.1 Terra. GPT-5.5 Pro moves to GPT-6 Astra.
  • Grok 4.6 and Grok 4.5 to Grok 4.7.
  • Gemini 3.7, 3.6 and 3.5 Flash to Gemini 3.8 Flash.
  • GLM 5.2 and GLM 5.1 to GLM 5.3, and DeepSeek V4 Flash to DeepSeek V4.1 Flash.
  • Eleven v3 to Eleven v4. Voices cloned before Eleven v4 may need retraining in ElevenLabs to sound their best.
  • GPT Image 1 to GPT Image 2.
  • GPT-4o mini Transcribe, GPT-4o Transcribe and Whisper to GPT Transcribe.

Where they live

All of it sits in the Models group of Admin. Providers (/admin/providers) is where accounts are connected and ordered. Agent Models (/admin/models) is where the models your agents think with are chosen, with its Default, Effort, Advanced Routing, Simulate and Web Search tabs. Image Models (/admin/images), Embedding Models, Voice Models and Judgement Models each choose the model for one kind of job. Usage (/admin/usage), under Installation, shows what has been spent, broken down by model and by agent. Stay in control puts that in the wider picture of what is yours to change, and Your installation covers the rest of the Admin area.