Claude Code limits Pro plans to Sonnet 4.5 and moves Opus behind extra usage

• om_patel5's screenshot post says Anthropic's support docs now limit Claude Code on the $20 Pro plan to Sonnet 4.5 by default, and require "extra usage" purchases to access Opus models.
• The same support-doc screenshot lists Opus 4.7, Opus 4.6, and Opus 4.5 as supported, but explicitly gated behind separate paid usage for Pro subscribers.
• In parallel, Ben Tossell's billing thread and his follow-up reply describe invoice emails that did not line up cleanly with the usage dashboard, even with extra usage turned off.
• A separate 7_eito_7 thread claims Claude Code can switch to metered billing when it detects HERMES.md in Git history, but that claim is still circulating as user-reported behavior rather than an official Anthropic explanation.

Anthropic appears to have buried the most consequential change in a support note rather than a launch post. You can see the supported-models warning in the screenshot evidence, compare it with user reports of receipt and dashboard mismatches, and then hit the stranger edge case, where one thread claims a filename in Git history can flip billing routes.

The clearest evidence in the pool is a screenshot of Anthropic's Claude Code model list. In om_patel5's post, the warning under "Supported models" says: "When using a Pro plan with Claude Code, you will only be able to use Opus models after enabling and purchasing extra usage."

That matters because the same screenshot shows Sonnet 4.5 alongside three Opus variants. Pro users still see Opus 4.7, 4.6, and 4.5 listed as compatible models, but not as part of the flat subscription.

For Claude Code users who treated Pro as the bundled path into Anthropic's best coding model, this is a product-line change disguised as a billing footnote.

The second thread is less about list pricing than legibility. In Ben Tossell's billing thread, Tossell said invoice emails kept arriving even though he had not used agents over the weekend, and said the receipts did not seem to match what appeared in his dashboard.

His later follow-up reply narrows the complaint: he was mainly confused by date mismatches between invoices and dashboard views, while acknowledging that some background "heartbeats" were expected.

That does not prove overbilling. It does show that once Claude Code mixes subscription access, extra usage, and background activity, users can lose a clean mental model of what triggered a charge.

A separate cluster of posts adds an unverified, but very specific, claim about how billing switches happen. In 7_eito_7's thread, the author says overseas reports described a case where HERMES.md appearing in a GitHub repository led Claude Code onto a separate metered billing path, with one reported loss of $200 in a day.

The follow-up posts break the claim into parts:

• 7_eito_7's cause summary says Claude Code includes Git commit history in the prompt.
• The same summary post says detection of HERMES.md can activate a rule that changes the billing route.
• 7_eito_7's mitigation post says the filename and commit history are the specific things to watch in the reported cases.

There is no official Anthropic source in the evidence pool confirming that mechanism. What the posts do add is a new concrete fear around Claude Code billing: not just higher charges, but hidden triggers inside repo context.