Antigravity connects to your Anthropic API key to route Claude model requests directly through your account. This article walks through the exact setup process, model selection, billing, rate limits, and how to debug common connection errors.
If you've been running AI workflows manually, copying outputs between tools, or paying for multiple subscriptions just to access different language models, Antigravity was built for exactly that pain. The platform lets you wire up your own API credentials, route requests directly through Anthropic's infrastructure, and keep full control over billing, rate limits, and model selection. But getting that initial connection right, especially with Anthropic's credential management system, trips up a surprising number of users. This article walks through the exact process, what to watch for, and what becomes possible once everything clicks into place.
Antigravity is a platform built around the idea that you should own your AI stack. Rather than routing your requests through a middleman that marks up costs or obscures what model you're actually talking to, Antigravity passes your API credentials directly to the provider. With Anthropic, that means your requests go straight to Claude's servers, billed to your Anthropic account, at the exact pricing Anthropic publishes.
This matters for three reasons:
The tradeoff is that you bear the setup responsibility. If your credentials are wrong, expired, or scoped incorrectly, nothing works. Getting this right the first time saves hours of debugging.
Most AI platforms fall into one of two categories: they either own the API access and charge you a markup, or they ask you to bring your own credentials. Antigravity sits firmly in the second camp.
When you add your Anthropic API token to Antigravity, the platform stores it in encrypted form and injects it as the Authorization header on every request you initiate. You never see the raw token again after entry, Antigravity doesn't expose it in logs, and your requests are counted against your personal Anthropic account as if you had made them directly.
💡 Important: This also means that if your Anthropic account runs out of credits, Antigravity requests will fail with an authentication error, not a platform error. Always check your Anthropic billing dashboard first when debugging connection issues.
The bring-your-own-credentials model attracts a specific type of user: developers, technical product managers, and researchers who need reproducible AI workflows without giving up financial control. Common use cases include:
If you fit any of these descriptions, the setup process below is worth the 10 minutes it takes.
Two things need to be in place before you open Antigravity's credential panel: an active Anthropic account with a generated API token, and an understanding of which Claude model tier your account can access.
Anthropic issues API tokens through the Anthropic Console. The steps are straightforward:
antigravity-prod or antigravity-test💡 Name your tokens by purpose, not by date. A token called
antigravity-prodis easier to revoke or rotate thantoken-june-2024when you eventually need to audit access.
The token format begins with sk-ant-api03- followed by a long alphanumeric string. If yours doesn't start with sk-ant, you may be looking at a deprecated format. Generate a new one from the Console.
Anthropic operates a usage tier system that affects rate limits and model access:
| Tier | Monthly Spend Requirement | Rate Limit (RPM) | Access |
|---|---|---|---|
| Free | $0 (trial credits) | 5 | Limited models |
| Tier 1 | $5 charged | 50 | Standard models |
| Tier 2 | $100 spent | 1,000 | Full model access |
| Tier 3 | $500 spent | 5,000 | Priority access |
| Tier 4 | $5,000 spent | Custom | Enterprise |
For most Antigravity use cases, Tier 1 is the minimum you want. The free tier's 5 requests per minute cap will cause Antigravity to queue or fail requests in any real workflow.
💡 Check your tier before setup. Go to Anthropic Console, then Billing, then Usage Limits. Your current tier is displayed there. Upgrade before connecting if you're still on the free tier.
With your API token ready and your tier confirmed, the actual connection in Antigravity takes under five minutes. The exact UI varies by Antigravity version, but the underlying flow is consistent.
Antigravity will typically make a lightweight validation call to Anthropic's /v1/models endpoint to confirm the token is valid. If accepted, you'll see a green status indicator. If not, an error message will specify the problem.
💡 Don't add extra characters. Tokens copied from PDFs or emails sometimes pick up invisible formatting characters. Paste into a plain text editor first, verify the format, then copy again before pasting into Antigravity.
After the green light, run one test request before building anything on top of the connection:
That last step is important. A successful response in Antigravity's UI does not always mean the request was routed through your token. Confirming via the Anthropic Console that usage was logged rules out any caching or routing issues.
The first real request often surprises people who came from a shared-credential environment. With your own token:
Once connected, you have access to every Claude model your Anthropic tier permits. This is where the setup pays off: you're no longer limited to whatever model the platform licenses.
Anthropic organizes Claude into three capability tiers. Each serves a different speed-to-quality tradeoff:
| Model | Speed | Cost | Best For |
|---|---|---|---|
| Claude Opus 4.7 | Slow | High | Complex reasoning, long documents |
| Claude Opus 4.6 | Slow | High | Multi-step analysis |
| Claude 4.5 Sonnet | Medium | Medium | Coding, structured output |
| Claude 4 Sonnet | Medium | Medium | General writing, Q&A |
| Claude 3.7 Sonnet | Medium | Lower | Balanced everyday tasks |
| Claude 4.5 Haiku | Fast | Low | Classification, extraction |
| Claude 3.5 Haiku | Fast | Low | High-volume simple tasks |
| Claude 3.5 Sonnet | Medium | Lower | Content drafting |
Antigravity lets you switch between these in the model selector dropdown. The selection persists per workflow, so you can have one pipeline running Opus for document analysis and another running Haiku for batch classification, both under the same API token.
This question comes up constantly, and the answer is always: start with Sonnet.
Claude 4.5 Sonnet hits a quality-to-cost ratio that works for most production tasks. Reach for Claude Opus 4.7 when the task genuinely requires deep reasoning, not just when you want the "best" result. Reach for Claude 4.5 Haiku when you're running hundreds or thousands of requests and the individual outputs don't need to be polished.
💡 A practical rule: If you're spending more than $5 per day on a single Antigravity workflow using Opus, audit whether Sonnet would produce acceptable results at a fraction of the cost. Most of the time, it will.
One of the biggest surprises for developers new to BYOK platforms is discovering that billing and rate limits work differently than expected.
Anthropic charges per token: input tokens and output tokens are priced separately, and the rates differ by model. As of mid-2025, rough pricing looks like this (check Anthropic's pricing page for current figures):
These costs accumulate in your Anthropic account, not in Antigravity's billing system. Antigravity doesn't add a markup on top of Anthropic's pricing; you pay Anthropic directly.
Antigravity surfaces token usage in its workflow dashboard. Each completed request shows:
This makes it practical to audit expensive workflows before they run up your Anthropic bill. The estimated cost field uses Anthropic's published rates, so the number you see should match your Anthropic usage invoice at month end.
💡 Set a budget alert in Anthropic Console. Under Billing, you can configure alerts at specific dollar thresholds: $25, $100, whatever your comfort level. Antigravity won't stop you from spending past your limit automatically; Anthropic's alert system is the safeguard.
Even a correctly entered API token can cause problems. These are the most common failure modes and how to resolve them.
Error: 401 Unauthorized or Invalid API Key
This almost always means one of three things:
Fix: Go to Antigravity settings, delete the saved token, and re-enter it manually from the Anthropic Console. Do not copy from clipboard history if you've had the token stored there for more than a few days.
Error: 403 Forbidden or Rate limit exceeded
A 403 means your token doesn't have permission to access the model you selected. This typically happens when you're on the free tier and try to access a model that requires Tier 1 or higher.
A rate limit error (429) means you've hit your requests-per-minute cap. The solution is either to upgrade your Anthropic tier or add request spacing in your Antigravity workflow settings.
| Error Code | Meaning | Fix |
|---|---|---|
| 401 | Token invalid or missing | Re-enter token |
| 403 | Insufficient tier | Upgrade Anthropic account |
| 429 | Rate limit hit | Add delays or upgrade tier |
| 529 | Anthropic overloaded | Wait and retry |
| 500 | Server error | Retry; report if persistent |
If your Antigravity requests succeed but take 30 seconds or more, the issue is usually prompt length, not the connection. Anthropic processes tokens linearly, and a 10,000-token context window takes real time to process even on fast models.
Fixes for high latency:
Getting the token in place is just the starting point. What Antigravity actually enables is more interesting than the connection process itself.
With your API token active, Antigravity can run scheduled workflows that hit Claude at intervals. Common automated setups include:
Each of these workflows runs under your API token, with costs flowing to your Anthropic account. The per-workflow token cost is visible in Antigravity's dashboard.
Beyond automation, Antigravity supports building internal tools that use Claude as a reasoning layer. Product teams have used this to build:
The main advantage here is that these tools don't require you to build an API integration layer yourself. Antigravity handles the request routing; you just define the workflow and the model.
💡 Keep sensitive data out of prompts. Even with your own API token, prompts sent through Anthropic are subject to Anthropic's data handling policies. Review those policies before routing PII, financial data, or proprietary information through the API.
Once one workflow is running cleanly, most teams run into the question of how to scale without losing cost visibility.
The cleanest approach is a separate Anthropic API token per environment:
antigravity-development token with a spending limit set in Anthropic Consoleantigravity-production token with monitoring and alerts configuredantigravity-testing token rotated monthlyThis lets you track costs per environment, revoke a token without affecting others, and audit usage separately. In Antigravity, you add each token to the respective workspace or project.
Anthropic's Console shows usage broken down by token and by model. Cross-reference this weekly against Antigravity's workflow dashboard to catch cases where a workflow ran more than expected or selected a more expensive model than intended.
A simple weekly check:
Unexpected spikes usually mean a workflow loop ran more iterations than planned, or a context window grew larger than the average. Both are easy to fix once you spot them.
💡 Set
max_tokensin every Anthropic request. Antigravity surfaces this as a parameter you can configure per workflow. Capping output length prevents runaway costs from open-ended generation tasks that occasionally produce very long responses.
If you've got the API connection sorted and want to see what AI can produce beyond text, PicassoIA's image generation tools are worth a few minutes of experimentation. The platform gives you access to over 90 text-to-image models, from photorealistic generators to artistic styles, all without needing to wire up credentials or manage billing yourself. Start with a simple prompt and iterate from there. It's the fastest way to see what these models can actually do when you're not wrangling configuration panels.
