AutoGPT pricing: hosted plans, self-hosting cost, and where the API key goes

AutoGPT pricing has three separate answers. The hosted AutoGPT Platform is paid with no free tier — Pro at $42.50 per month billed annually ($50.00 billed monthly) and Max at $272.00 per month billed annually ($320.00 billed monthly). Self-hosting the same Platform carries no licence fee. AutoGPT Classic is free, MIT and unsupported. All three live in one repository, which is why most price summaries for this product are wrong: they quote one and describe another.

The repository is healthy, whatever the old tutorials imply. The GitHub API reports Significant-Gravitas/AutoGPT as not archived, not disabled, not a fork, default branch master, pushed the same day this page was checked, with 187,475 stars. The most recent release is tagged autogpt-platform-beta-v0.8.0, published September 19, 2026, after v0.7.4 on September 4 and v0.7.3 on August 28 — roughly weekly. The tag string still carries "beta" as a naming legacy; the hosted platform itself is publicly sold, so do not read a waitlist or a closed release into that word.

Three products, one name

Two things follow from that table. First, Polyform Shield is a source-available licence, not an OSI-approved one — the README's own wording is "free for personal and internal business use; cannot be sold as a competing hosted service" — so the Platform is not open source in the sense MIT is, even though AutoGPT's marketing uses the phrase. Only classic/ and the rest of the repository are MIT. Second, if a guide tells you to put OPENAI_API_KEY in a .env next to SMART_LLM and FAST_LLM, it is describing Classic, whose current README documents exactly those variables and Python 3.12+ with Poetry. That is a migration-away target, not a recommendation. Sources: master README and classic/README.md, both read September 21, 2026.

Hosted AutoGPT Platform: what the subscription actually buys

Read from agpt.co/pricing on September 21, 2026. The annual figures are what the page renders by default; the monthly ones come from the price element's own data-monthly attributes in page source, and the toggle labelled "Save 15%" checks out both ways ($50.00 × 0.85 = $42.50, $320.00 × 0.85 = $272.00). Team is not purchasable today, so treat any figure you see quoted for it as invented.

Now the fact that breaks most AutoGPT budgets. The subscription is not a usage bundle. AutoGPT's pricing page states that automations — running agents and workflows — are pay-as-you-go via the credit wallet, at the same rate on every plan, and that the subscription gates AutoPilot Chat usage instead. The homepage FAQ puts it in plain words: the subscription covers chat, the credit wallet covers the agents you run. So Max's advertised 8.5x is 8.5x the chat allowance, and buying it changes nothing about what an agent run costs.

Two places where AutoGPT's own copy disagrees with itself, shown rather than resolved. The homepage says each plan "includes enough usage to put agents to work from day one … and a credit wallet for your agents' runs", while the comparison table says you "pay only for the automation runs you use"; the billing documentation punts to the pricing page for included-credit details, and the pricing page publishes no such number. Separately, none of the three public pages read here — the pricing page, the homepage FAQ and the Credits & Billing doc — states what a credit costs in dollars, and checkout sits behind a login. Confirm both in checkout before you budget. This page will not guess either one.

What the credit wallet charges for model tokens

AutoGPT does not publish a per-token rate on its pricing page, but the platform's open-source code does the arithmetic in the open. block_cost_config.py at master says rates are stored as "credits/1M tokens at the current credit-to-USD conversion, 1 credit ≈ $0.01, uniform 1.5x margin over the published provider price", with the constant _USD_PER_1M_DIVISOR = 150 annotated "150 = 100 cr/$ × 1.5x margin". Converting the shipped catalogue with those two constants gives the charged rate directly.

Those are AutoGPT's own catalogue entries, converted with AutoGPT's own constants on September 21, 2026 — 83 of its entries carry per-1M rates, spanning roughly $0.12 to $270.00 per million output tokens at the charged rate. The last column is filled only where the catalogue entry itself carries a provider_input_usd_per_1m / provider_output_usd_per_1m pair, which four entries do and the rest leave out; dividing the credits by 150 instead would reproduce the figure the 1.5x margin was set from, but that is the margin's own arithmetic rather than a vendor rate AutoGPT publishes. Three caveats that matter more than the numbers. This is what the shipped code does, not a ratio AutoGPT advertises anywhere on its pricing page, so treat it as a source-code reading. The live deployment layers per-account overrides over that catalogue, so your account can differ. And model slugs churn weekly, which is why the durable claim here is the 1.5x mechanism and not any one model name.

One consequence the code implies but no documentation states: the LLM block's credit charge is filtered on a credential id pinned to the platform's own system credentials, the ones literally titled "Use Credits for OpenAI" and "Use Credits for Anthropic". A credential you added yourself has a different id and matches no cost entry. Independently, AutoGPT's chat rate-limit module says outright that runs where the user supplied their own API key do not decrement the chat meter, "the user is paying the provider, not us". Bringing your own key therefore looks economically meaningful on the hosted platform even though bringing your own endpoint is impossible there — but that first half is a reading of the source, not a billing outcome anyone has tested on a paid account. Verify it on your own bill before relying on it.

AutoGPT self-hosted cost: no licence fee is not no cost

Installation itself is a git clone, a copied .env.default and docker compose up -d --build. Note the release installer status before you plan around it: the README says the Linux and macOS single-container release installer is "coming with the next appliance release" and that Windows users should continue with the manual guide for now. Re-check that line before you standardise on it, because it is written to change.

One more self-hosting cost that is easy to under-budget. AutoGPT's troubleshooting text says AutoPilot ships an approximately 8,000-token system prompt, and that on 4 CPU cores an 8B Q4 model prefills at roughly 3-4 tokens per second, "so a fresh turn takes ~35-45 min just to start generating". A consumer GPU brings that down to seconds. That is AutoGPT's own strongest argument for pointing the chat endpoint at a hosted API instead of CPU-only local hardware — and it is a published figure about local inference, not a measurement of anything here.

The AutoGPT API key question forks in two

AutoGPT has two independent LLM configuration surfaces, and its documentation states the problem itself: "the two paths read different env vars, so configuring one does not configure the other." Old tutorials and aggregator pages conflate them constantly.

The block layer's limit is structural, not a setting anyone forgot to expose. In blocks/llm.py the accepted providers are exactly AIML_API, ANTHROPIC, GROQ, OLLAMA, OPENAI, OPEN_ROUTER, LLAMA_API, V0. In util/llm/providers.py the OpenAI client is built as openai.AsyncOpenAI(api_key=api_key) with no base_url parameter on that path, and Anthropic and Groq are the same; the OpenAI-compatible helper that does take a base URL is only ever reached with four hard-coded URLs. So anyone telling you to "just change the base URL in AutoGPT" is describing a different product. The one editable host field, Ollama's, is wrapped in an SSRF guard checked against an operator-configured trust list, and AutoGPT's documentation states flatly that Ollama "is only available when self-hosting … It cannot be used with the cloud-hosted version."

The chat agent is the surface AutoGPT itself sanctions for an outside endpoint. Its local-LLM guide lists "a managed OpenAI-compatible API you don't pay AutoGPT for" as a supported target, and opens with "Self-hosting only — the cloud agpt.co deployment routes AutoPilot through Anthropic / OpenRouter and ignores the variables below." There is no plan — Pro, Max, or the forthcoming Team — that unlocks a custom base URL. The gate is deployment mode, not price tier.

# Self-hosted AutoPilot chat only. These variables do NOT configure the
# agent-graph LLM blocks, and the cloud deployment ignores them entirely.
CHAT_USE_LOCAL=true
CHAT_BASE_URL=https://api.kunavo.com/v1
CHAT_API_KEY=sk-...
CHAT_FAST_STANDARD_MODEL=claude-sonnet-4-6

Four documented limits before you commit to that route, all from the same guide. Extended thinking is unavailable under the local transport and auto-downgrades to the fast model, logged at WARNING rather than failing. From inside the containers the base URL must not be 127.0.0.1 or localhost, and CHAT_API_KEY deliberately does not fall back to OPENAI_API_KEY. Backends that do not report a context window — the guide names LiteLLM proxy, Jan and text-generation-webui — fall back to an assumed 32k regardless of the real model. And the setting is broader than chat: the same client serves the dry-run block simulator, onboarding extraction, prompt compression and the nightly dream pass, so budget for more than visible chat turns. Kunavo serves no embedding model, so the marketplace-search embedding step is not something Kunavo covers on this path either — and AutoGPT hard-locks that replacement anyway, since the pgvector column is declared vector(1536) and inserts of any other width fail.

Best model for AutoGPT: pick the mechanism, not the slug

The hosted platform does not ask you to guess. Its default is whatever the shipped catalogue flags as its recommended GA entry — llm_models.py states it as "the platform default IS the catalog's recommended model — one fact, one home", taking the first enabled recommended entry that is also GA. On September 21, 2026 that resolves to GPT-5.6 Terra. Because the catalogue changes weekly, the mechanism is the fact worth learning and the slug is not.

The lever that actually moves an agent bill is not model choice but the retry loop, and AutoGPT ships a control for it: the pricing page lists "Per-Task LLM Budget Caps — set max_budget_usd to bound LLM spend on any task". Set that before you tune models. Note also that the pricing page's marketing rows and the shipped catalogue are not the same list — the page advertises "frontier models from Anthropic, OpenAI, Google, xAI, Meta, Mistral, Cohere, and more" and names a Claude row of Opus 4.7/4.6/4.5, Sonnet 4.6/4.5 and Haiku 4.5, while the code catalogue at master carries different entries again. Check what your own account's picker offers rather than either list.

On a self-hosted AutoPilot pointed at your own endpoint, the constraints change shape. The local transport passes your model id through verbatim, so it has to be an id your own endpoint resolves; AutoGPT's guide asks for bare model names because its running example is Ollama, which cannot resolve a provider/model slug. The extended-thinking path is unavailable, so a reasoning-heavy model buys you less there than its rate suggests. A model whose real context window your endpoint does not advertise is treated as 32k. Pick for the fast path first, and confirm the window is reported.

A worked estimate for the self-hosted chat path

This is illustrative token arithmetic, not a measured task cost and not a bill ceiling. Assume one self-hosted AutoPilot day that sends 400,000 uncached input tokens — an approximately 8,000-token system prompt re-sent on each turn plus growing conversation — and receives 25,000 output tokens. Rates are live Kunavo catalog prices per million tokens.

Scale by your own days before treating that as a budget, and note what it does and does not compare. It is not a substitute for the hosted platform's subscription, which buys the managed service, the marketplace and the credentials AutoGPT supplies for its integrations — self-hosting gives all of that up. Kunavo's catalog amount is a billing floor rather than a cap: when the upstream reports its charge, the bill is the greater of catalog cost and upstream cost times the applicable markup. Cache charges and external tools sit outside this example. The minimum Kunavo top-up is $10 in prepaid credit, a funding minimum rather than a task fee or a subscription — see billing details.

Which route wins when

If the answer is a gateway on the chat path, the setup is an OpenAI-compatible base URL and a key — start at the Kunavo quickstart and create an account when you are ready to fund one. Keep a working route available while you try it, run one bounded task, then read what your account actually recorded for it. For the wider comparison, OpenAI-compatible API covers the protocol, OpenRouter alternatives compares the incumbent gateway AutoGPT compiles in, and the AI agent API directory records where each client's configuration boundary sits. If you are budgeting a workflow builder rather than an agent platform, n8n AI API cost and Dify API cost use the same separation between software fees and model tokens.

FAQ

Checked September 21, 2026: agpt.co pricing and homepage copy, the Credits & Billing doc, the GitHub API for repository status and release dates, the master sources of README.md, classic/README.md, blocks/llm.py, util/llm/providers.py, block_cost_config.py and the model catalogue, and AutoGPT's local-LLM and Ollama documentation. Not established here and deliberately absent: the dollar value of a credit, whether any plan includes starting credits, the absolute size of the chat allowance, and whether a trial exists. Kunavo has run no compatibility test against AutoGPT; the CHAT_BASE_URL route above is AutoGPT's documented scenario, and every dollar example is illustrative token arithmetic rather than a measured task cost.