Claude Max API credits: what $100 or $200 a month covers, and what Thariq's AI new-tab page sends to the cloud

Since Anthropic announced it on October 7, a Claude Max subscription comes with monthly credit for the Claude API: $100 on Max 5x, $200 on Max 20x, the same as the plan's price. The next day Thariq, who works at Anthropic, showed what he does with his: a Chrome extension that rewrites his new-tab page every day from the sites he actually reads.

Video: Thariq's demo on X. "It starts with what you actually read. Then opens the pages you follow. And writes you a fresh homepage." The newspaper on screen is a designed mock-up; his real output is linked below.

Both are worth understanding separately. The credits are a real benefit with a handful of rules that will catch people out, especially anyone who already has an API account. The demo is more interesting than it looks. It is a small, working example of a pattern Thariq says Anthropic is moving toward, where the agent loop runs in Anthropic's cloud and borrows your browser for the parts only your browser can see, including pages you are signed into.

What do Claude Max and Team plans get in API credits?

Per Anthropic's docs page and Help Center article:

The launch post says the credits are "designed to allow our users to experiment with building tools, apps, and agents that call our API," and that "they can be used on any of our models." Simon Willison called it "really generous," and on the face of it, it is: a Max 20x subscriber now gets the subscription price back as API spend.

The catch is in the timing. Credits arrive each billing cycle shortly after the plan payment, and "Unused credits expire at the end of each billing cycle (or monthly for annual plans) and don't roll over." New subscribers wait seven days before they can claim, and some subscribers report that a first grant claimed late in their cycle expired within days. "Use them or lose them," as Willison put it. For Team plans, each grant uses the seats present at the start of the billing month, so added seats count from the next one.

What do the Claude API credits cover, and what don't they?

They cover the Claude API (the Messages and Message Batches APIs), Claude Managed Agents, the Claude Agent SDK and the Console Playground. They don't cover interactive Claude Code, extra usage in the Claude apps, or Claude on Bedrock, Vertex AI, Microsoft Foundry and Claude Platform on AWS.

The claude -p case is where most of the confusion sits, and the Help Center answers it directly:

API credits cover claude -p and the Claude Agent SDK when you run them
yourself with an API key from your linked Claude Console organization,
because that usage is billed as Agent SDK usage. When you're signed in with
your Claude plan instead, claude -p and Agent SDK usage still draw from your
plan's usage limits and don't use your API credits.

Runs started by the Claude Code GitHub Action, an IDE extension or the desktop app count as Claude Code, so the credits never cover them, "even with -p." Open an interactive Claude Code session with a key from an organization that holds only these credits and you get "Credit balance too low."

The line is drawn by the type of usage, not by the key. That matters because the credits are additive. Logged in with your plan, the Agent SDK and claude -p keep running on your subscription limits, as the Agent SDK article now says: "You can still use the Claude Agent SDK, claude -p, and third-party apps with your subscription limits." That wasn't always the plan. Earlier this year Anthropic announced a change to how Agent SDK and claude -p usage would be billed, then paused it on June 15, saying "nothing has changed." The credits FAQ now says that earlier credit "isn't available." This one is additive: it comes on top of what your subscription already covers.

Two gaps remain. No page says whether server-side tools such as web search, or the runtime of Managed Agents sessions, are paid from the credit, and none says whether you may use it to power something you sell. The claim flow asks you to accept Anthropic's Supplemental Credit Terms, and the published version, effective March 2024, says credits can't be transferred or resold.

How do you claim Claude Max API credits without a billing surprise?

On claude.ai, go to Settings > Billing (Organization settings > Billing on Team, as an Owner or Primary Owner), find the API credits section, choose "Link organization," and pick or create a Claude Console organization where you hold the Owner, Admin or Billing role. No card is needed, and plans bought in the iOS or Android app are eligible but claim on the web. The Help Center suggests using the same email in the Console as on claude.ai. The rollout runs over several days, so a missing button may just mean you haven't been reached yet.

Three rules decide which organization you should pick.

The link is effectively permanent. "Each plan links to one Console organization, and each Console organization can receive credits from one plan. You can't change the linked organization yourself," per the docs. To change it, you contact support.

Every key in that organization spends the credit. "Every API key and workspace in the linked organization draws from the same balance." If you link the organization your production keys live in, the free credit and your production traffic share a pool.

Auto-reload ignores the free credit. This one will cost people money. The docs say included credits are spent first, but "Auto-reload is based only on your purchased balance. It reloads when that balance reaches your threshold, even if included credits remain." Link an existing organization with auto-reload switched on, let something the credit doesn't cover (an interactive Claude Code session on the same key, say) drain the purchased balance to the threshold, and your card is charged while $200 of free credit sits unused.

For most people the answer is a new organization, used only for personal projects, with auto-reload off and no purchased balance. Then, when the credit runs out, requests stop: "Usage is never charged to your Claude plan." If you want a hard limit per project inside it, give each project its own workspace with a spend limit. One more mismatch to know about: credit usage counts toward the organization's monthly spend cap, which resets on the first of the calendar month, while the credit refreshes on your billing date.

Linking also moves the organization to at least the Start rate-limit tier, above the Evaluation tier new organizations may start in. The docs guarantee at least the Build tier only when monthly credits are "over $200," so on my reading a Max 20x credit of exactly $200 guarantees Start, not Build. That is worth checking in the Console after claiming.

How does Thariq's AI new-tab page work?

Thariq's follow-up: "it's built on top of Claude Managed Agents, doing code generation in the cloud sandbo[x], using Sonnet 5.5 each run costs ~$0.75." The code is on GitHub as ThariqS/ai-newtab, with a README and an ARCHITECTURE.md that are more candid than most launch demos. From those:

• It is a Chrome extension with no server. It replaces the new-tab page and talks to Anthropic directly with your API key, which it keeps in the extension's local storage. It isn't in the Chrome Web Store; you build it and load it unpacked.
• The agent loop runs in Managed Agents. The extension starts a session, streams its events, and answers tool calls. Sonnet 5.5 is the default model, with Haiku 5.5 and Opus 5.5 in a picker.
• It learns what you read from your browser history. A getHistory tool returns a ranked summary of your most-visited domains. There's no RSS and no list of sites to maintain.
• It reads pages in your own browser. A getPageHtml tool lets the agent choose up to eight pages. The extension opens them in a background window of your signed-in Chrome, cleans the HTML, and uploads it into the agent's sandbox. It doesn't use web search or web fetch.
• The output is code, not a template. The agent writes one React component, which the extension renders in a sandboxed iframe, because "it is model-written code." Free-text instructions set both content and style: a broadsheet, a single calm column, a green-on-black terminal.
• The schedule is your browser. A one-minute alarm in the extension rebuilds the page once it is older than 6, 12, 24 or 48 hours. There is no cron. If Chrome isn't running, nothing happens.

This is exactly what Thariq described a day earlier: "we're increasingly going to be moving towards Claude's 'brains' in the cloud and giving Claude 'local hands' to operate on your computer." Here the brain is a Managed Agents session and the hands are two tools that only a browser can run, reading history and opening logged-in pages. The extension answers those tool calls over the stream it opened itself, so, in the architecture document's words, "There is no inbound connection and no public endpoint."

What does Thariq's ai-newtab extension send to Anthropic?

The extension reads pages with your browser's signed-in context. Thariq's own sample output has a section built from his X timeline, items from Hacker News and Reddit, and a college football preview. That is the point of reading pages inside your browser: an ordinary server fetch without your login gets the logged-out page, not your timeline. It also means the cleaned HTML of up to eight pages goes into an Anthropic sandbox on every build.

The project is upfront about it. Per the README, the request carries "top domains, visit counts, a few page titles" plus up to eight cleaned pages, which "can include private content." It skips search engines, webmail, banking and logged-out marketing pages, and deletes the session and every uploaded page when the build ends. The architecture document tells anyone redistributing it to confirm retention and zero-data-retention eligibility "against current docs before distributing this," and the README warns that "anyone with access to the browser profile can read the stored API key." The agent's shell is switched off because "scraped pages are untrusted input," and the agent is told never to invent a link.

So the trade is plain. You lend an agent the session you are signed into, for up to eight pages per build, in exchange for a front page made from what you actually read. Whether that is acceptable depends on what is in your history. And because you install it unpacked from source, it is also a tool with your reach: it holds permission to read your history and open tabs as you, so read what you build before you load it.

Is the AI new-tab page a good use of Claude Max API credits?

Using Thariq's figure of about $0.75 a build on Sonnet 5.5, one page a day comes to roughly $23 a month, a little over a tenth of the Max 20x credit. Mind the interval setting. Automatic rebuilds are on by default once you have built the page by hand, and at the shortest interval, six hours, the same estimate is closer to $90 a month. A 24-hour interval matches "generated everyday," and Haiku 5.5 is the cheaper model in the picker. Actual cost depends on how much the agent reads.

The bigger point is what the credit is for. Anthropic's developer account on X said the credits work "in your code or third-party harnesses," and the docs say to "build and run your own applications and agents." A daily page is a good first project because it is small and runs every day, which makes a monthly use-it-or-lose-it allowance easy to plan around. It also teaches the parts of Managed Agents that are hard to learn from docs. The architecture file lists several, from how uploaded files are mounted to why a session waiting on an unanswered tool call idles forever.

If a browser-bound schedule doesn't suit you, the other route is server-side: a Managed Agents deployment on a schedule that runs whether Chrome is open or not. You lose the signed-in pages, which is exactly why Thariq put the reading in the browser.

How do Claude's API credits compare with Google and OpenAI?

Google has done a version of this since January. Google AI Pro and Ultra subscribers get $10 and $100 a month in Google Cloud credits, usable on Vertex AI, Cloud Run and the Gemini API, and per Google's developer program FAQ they last a year rather than a month. OpenAI sells credits for ChatGPT plans that fund extra usage of its own features, and its help article is explicit that "These credits are not API credits." Willison notes that OpenAI does let you use a Codex subscription for personal API use, which he calls "a better deal for heavy API users," and says Anthropic's credit "goes at least some way to overcoming that difference."

What should a Claude Max subscriber do with the API credits?

Claim them into a new Console organization with auto-reload off, then check the tier and the expiry date in the Console. Use an API key from that organization for the scripts and agents you would otherwise run with claude -p, and scope it to those commands. If ANTHROPIC_API_KEY is exported in your shell, Claude Code uses the key instead of your subscription, so an interactive session in that shell would hit the credits-only organization and stop with "Credit balance too low." And pick one small, daily thing to build, because an allowance that resets every month rewards a habit more than a big project.

If that thing is Thariq's new-tab page, decide first which of your logged-in sites you are happy to see read by an agent every morning.