Granular Hosted: models and credits, no API keys

What Granular Hosted is

Granular needs an AI model behind it to do any work. Normally you supply that yourself — you sign in with your own Claude or Codex account, or paste your own provider API key. That works well, but it assumes you already have one.

Granular Hosted removes that assumption. It is a paid add-on where Granular runs the model connection for you:

• No separate AI subscription. You do not need a Claude Pro or ChatGPT plan.
• No API keys. Nothing to create, paste, rotate or leak.
• One bill. Model usage arrives on your Granular invoice, not on four provider invoices.
• Many models, one switch. Change model mid-conversation without re-authenticating anything.

It stacks on any plan, including Free. It is a separate purchase from Granular Pro — Pro lifts the limits on projects, sessions and workers, while Hosted supplies the models.

The models, and what each is good for

The hosted lineup is 19 models from 9 providers — the frontier labs plus the strong open-weight challengers. Every one of them is agent-capable: it can reliably drive tool calls, which is what Granular actually needs. Models that only chat well are deliberately left out.

What follows is the same information the app shows when you hover a model in the picker: strengths scored out of 5 for coding, reasoning, writing and everyday chores, and the cost of a typical turn in credits. The turn used for that estimate is a deliberately conservative, cache-free floor — real turns use prompt caching and usually go further than the numbers suggest.

Anthropic

• Claude Opus 5 — the most capable model: elite coding and reasoning, at flagship price.
  Coding 5 · Reasoning 5 · Writing 5 · Everyday 2  ·  ≈13 credits a turn, about 8 turns per $1
• Claude Fable 5 — as capable as Opus 5, with an extra edge for long-form writing.
  Coding 5 · Reasoning 5 · Writing 5 · Everyday 2  ·  ≈13 credits a turn, about 8 turns per $1
• Claude Opus 4.8 — last-generation flagship, still top-tier for coding and writing.
  Coding 5 · Reasoning 4 · Writing 5 · Everyday 2  ·  ≈13 credits a turn, about 8 turns per $1
• Claude Sonnet 4.6 (the default) — the balanced default: much of Opus 5's quality, far cheaper.
  Coding 4 · Reasoning 4 · Writing 4 · Everyday 4  ·  ≈8 credits a turn, about 13 turns per $1
• Claude Sonnet 5 — Sonnet-class reasoning at a fraction of Opus 5's cost.
  Coding 4 · Reasoning 4 · Writing 4 · Everyday 4  ·  ≈5 credits a turn, about 20 turns per $1
• Claude Haiku 4.5 — fast everyday model: quick chores over hard reasoning.
  Coding 3 · Reasoning 2 · Writing 3 · Everyday 5  ·  ≈3 credits a turn, about 33 turns per $1

OpenAI

• GPT-5.1 — OpenAI's flagship: reasoning on par with Opus 5.
  Coding 4 · Reasoning 5 · Writing 4 · Everyday 3  ·  ≈4 credits a turn, about 25 turns per $1
• GPT-5.1 Codex — OpenAI's best coder: matches Opus 5 on code.
  Coding 5 · Reasoning 4 · Writing 3 · Everyday 2  ·  ≈4 credits a turn, about 25 turns per $1
• GPT-5.1 Codex Mini — a cheap, capable coder: much of Codex's skill, far less cost.
  Coding 4 · Reasoning 3 · Writing 2 · Everyday 4  ·  ≈1 credit a turn, about 100 turns per $1
• GPT-5 Mini — cheap, fast general model: everyday chores over deep reasoning.
  Coding 3 · Reasoning 3 · Writing 3 · Everyday 5  ·  ≈1 credit a turn, about 100 turns per $1

Google

• Gemini Pro — Google's flagship: top-tier reasoning and huge context.
  Coding 4 · Reasoning 5 · Writing 4 · Everyday 3  ·  ≈6 credits a turn, about 17 turns per $1
• Gemini Flash — fast, cheap, and capable: a great everyday workhorse.
  Coding 3 · Reasoning 3 · Writing 3 · Everyday 5  ·  ≈1 credit a turn, about 100 turns per $1
• Gemini Flash-Lite — the lightest, fastest option: quick chores, not hard problems.
  Coding 2 · Reasoning 2 · Writing 3 · Everyday 5  ·  ≈1 credit a turn, about 100 turns per $1

The challengers

• Grok 4.6 (xAI) — xAI's flagship: a well-rounded, up-to-date generalist.
  Coding 4 · Reasoning 4 · Writing 4 · Everyday 3  ·  ≈4 credits a turn, about 25 turns per $1
• DeepSeek v3.1 — strong open coder and reasoner, a fraction of Opus 5's cost.
  Coding 4 · Reasoning 4 · Writing 3 · Everyday 4  ·  ≈1 credit a turn, about 100 turns per $1
• Kimi K2 Thinking (Moonshot) — capable long-context reasoning at a very low price.
  Coding 4 · Reasoning 4 · Writing 3 · Everyday 3  ·  ≈1 credit a turn, about 100 turns per $1
• GLM-4.6 (Z.ai) — strong coding agent, a fraction of Opus 5's cost.
  Coding 4 · Reasoning 3 · Writing 3 · Everyday 4  ·  ≈1 credit a turn, about 100 turns per $1
• Qwen3 Max (Alibaba) — Alibaba's largest: strong multilingual coding, low cost.
  Coding 4 · Reasoning 4 · Writing 3 · Everyday 3  ·  ≈2 credits a turn, about 50 turns per $1
• Mistral Medium 3.1 — a solid all-rounder: dependable writing and everyday work.
  Coding 3 · Reasoning 3 · Writing 4 · Everyday 4  ·  ≈1 credit a turn, about 100 turns per $1

If you are unsure, start on Claude Sonnet 4.6 — it is the default because it balances quality and cost well. Reach for a flagship when a task is genuinely hard, and drop to a 1-credit model for chores. The picker is per-message, so switching costs you nothing but a click, and hovering any model shows this same card before you commit.

How credits work

Hosted usage is metered in credits, and the rule is deliberately simple:

100 credits = $1 of model usage, at the provider's own price.

Granular takes its margin once, at the moment you buy. After that the model runs through at cost. That has a consequence worth understanding: a credit buys the same amount of real inference no matter which model you spend it on. An expensive model does not have a hidden penalty attached — it just burns credits faster, in exact proportion to what it genuinely costs.

Your balance has two parts, and they behave differently:

• Monthly credits — included with the subscription, and they reset at the start of each billing period. Use them or lose them.
• Top-up credits — bought separately, and they roll over indefinitely.

Granular always spends the monthly credits first, so your rolled-over balance is the last thing to be touched. That is the right order: it protects the credits that do not expire.

Plans and credit packs

Granular Hosted is a monthly add-on, and the only difference between the tiers is how many credits arrive each month:

• AI Base — $20/month · 1,600 credits a month
• AI Plus — $50/month · 4,000 credits a month
• AI Max — $100/month · 8,000 credits a month

Need more in a given month? Credit packs are one-time purchases whose credits roll over until spent: $10 for 800 or $25 for 2,000.

Two honest notes for choosing. First, every tier and pack buys credits at exactly the same rate — there is no bulk discount and no small-buyer penalty, so pick on how much you will actually use, not on deal-hunting. Second, remember the split from the section above: monthly included credits reset each billing period, while pack credits keep. If your usage is uneven, a smaller tier topped up with packs in busy months often fits better than a big tier you only sometimes fill.

Both live in Settings → Billing & plan inside the app. Checkout opens in your browser, and your unlock code is shown on the success page and emailed to you.

Watching your balance

The credits gauge sits beside the model picker, under the prompt, and updates as you work. Two places give you the fuller picture:

• The gauge itself — remaining credits at a glance, without leaving the conversation.
• Settings → Billing & plan — what you are subscribed to, what renews when, and the email each subscription bills to.

Metering happens on Granular's server, in the background, after your turn has already streamed back to you. Accounting never blocks or breaks a conversation.

When credits run out

Nothing breaks and nothing is lost — hosted turns simply stop until there is balance again. Your sessions, files, terminals and history are all untouched, because none of them depend on the hosted connection.

You have three honest ways forward, and none of them is wrong:

• Top up or move up a tier if hosted models are how you want to work.
• Switch to your own Claude or Codex login if you already pay for one — the work continues immediately.
• Switch to your own API key stored in the Vault, and pay the provider directly.

Granular is deliberately not a trap. Hosted is a convenience, not a lock-in — the app is fully usable on a model connection you own.

Using your own model instead

Switching away from Hosted takes one change in the model picker. Your options are covered in detail elsewhere in these docs:

• Connect an LLM — sign in with an existing Claude or Codex account.
• Integrate Codex — drive your logged-in Codex CLI.
• Connect Gemini — Google Gemini with a Vault-injected key.
• The Vault — where any provider key belongs: encrypted, injected into your terminal, never printed in chat.

You can keep Hosted subscribed and still use your own login day to day. The picker is per-conversation, so nothing forces you into one or the other.

Related