Your Claude subscription stops paying for agents on June 15. The SDK and claude -p move to a separate credit billed at full API rates.

Interactive stays, programmatic moves

The whole change rests on one distinction Anthropic is now drawing hard: are you typing at Claude, or is your code calling it? Type at it, and nothing changes. Call it programmatically, and you are off the flat subscription and onto a meter. Here is exactly where every surface lands.

Anthropic's own wording, from the help article, is blunt: "Starting June 15, 2026, Claude Agent SDK and claude -p usage no longer counts toward your Claude plan's usage limits," while "your subscription usage limits stay the same and stay reserved for interactive use of Claude Code, Claude Cowork, and Claude." If your day is mostly chatting with Claude and running Claude Code by hand, you can stop reading. The rest of this is for anyone whose pipelines, cron jobs, or harnesses do the talking.

The credit, by plan

Every eligible plan gets a monthly programmatic credit equal to what the plan costs. It is denominated in dollars, not tokens, and it is spent at API list rates. Eligible users get a claim email before June 15 and opt in once.

Four details decide how this plays out. The credit is per-user, so a team cannot pool it. It does not roll over, so a quiet month is money left on the table. Overflow past the credit only happens if you have explicitly enabled usage credits, otherwise the SDK simply stops. And the dollar value is the headline number, not a bonus on top of it. Anthropic is handing you back exactly your subscription's price, priced at the same API rates anyone with a raw key pays.

What you pay once the credit is gone

Past the credit, the SDK bills at standard published API rates. These are the same numbers on the pricing page, and they have not moved. What changed is who pays them.

One asterisk worth keeping in mind on Opus. The 4.7 tokenizer can spend up to 35 percent more tokens than 4.6 on the same code or JSON, so even with the rate card flat, the effective cost per request runs higher than the sticker suggests. We dug into that in the Opus 4.7 pricing breakdown. Cache reads at a tenth of the input rate and batch jobs at half price are the two levers that stretch the credit furthest.

What the credit actually buys

Dollars are abstract, tokens are the thing you spend. Convert each tier's credit into tokens at a plain 50/50 input-output split and the runway gets concrete. Opus blends to $15 per million across that split, Sonnet to $9, Haiku to $3.

Now sit those numbers next to how a coding agent behaves. A single long Claude Code session on the Agent SDK, counting file reads, tool calls, and the retries that come with real work, runs into the millions of tokens fast. On Opus, a top-tier $200 credit is about 13M blended tokens, which a heavy agent operator can clear in days, not the month it is meant to last. Drop to Sonnet and the same credit stretches to 22M. Route the cheap work to Haiku and it is 67M.

Caching is the asterisk that rewrites this table. Agent loops reread the same system prompt and the same files on every turn, and at the cache-read rate the input side gets ten times cheaper. A workload with stable context can run several times longer on the same credit. Output tokens get no such discount, so the agents that burn through fastest are the verbose, reasoning-heavy ones that emit long completions every turn. Run the exact split for your own traffic on the calculator.

How big a change is this, honestly

Depends entirely on what you were doing. If you only ever ran Claude interactively, the answer is zero. If you were pointing an agent harness at a subscription and letting it grind, this is a genuine cost event, because that usage was being subsidized far below what the API charges. The hard part is that nobody can hand you a single multiplier, because the multiplier is a function of your own prior burn.

The numbers floating around the community are estimates, and they should be read that way. A widely shared analysis put the effective increase somewhere between about 12x for a light Pro user running OpenClaw and over 150x for heavy Sonnet automation on a Max 20x plan. Theo Browne framed it as up to roughly 25x less capacity and pushed back on the "free credits" language, arguing the credit just re-prices what used to be effectively unmetered. We would not repeat the 150x figure as fact, because the assumptions under it (hours per week, output tokens per hour, dollar value extracted) are not Anthropic's and not independently confirmed. The direction is certain. The magnitude is yours to measure.

This did not come out of nowhere. In April, Anthropic blocked third-party harnesses like OpenClaw from authenticating with subscriptions altogether, the change we covered in the April third-party pricing piece. It briefly tested pulling Claude Code from Pro entirely, then landed here. The June 15 credit model actually reinstates OpenClaw and its peers, just with a meter bolted on. Read generously, it is a compromise that keeps the harnesses alive. Read plainly, the free ride on programmatic agents is over.

Against OpenAI and Cursor

Compared to the obvious alternatives, Anthropic's deal is neither the most nor the least generous. ChatGPT Pro at $200 carries no programmatic envelope at all, the OpenAI API is a separate bill from day one, so a Max 20x plan plus its $200 SDK credit is actually more programmatic value than ChatGPT Pro on paper. OpenAI moved Codex onto API-token pricing back on April 2, and GPT-5.5 doubled the GPT-5 line to $5 input and $30 output, which puts its output rate above Opus 4.7's $25. Both labs raised the real cost of heavy programmatic use this spring, just through different doors.

Cursor is the one that undercuts everyone here. Cursor Ultra at $200 hands back something like $400 of credit at API rates, twice Anthropic's envelope, which makes it the cleaner pick for anyone running a pure agent fleet rather than splitting time between chat and code. The broader read across all three is the same: the window where you could run unlimited agents off a flat subscription is closing across the industry, and the credit is the polite version of that door shutting.

Your move before the 15th

If your Claude use is all interactive, you are done, the credit is irrelevant and your limits are untouched. For everyone running agents, the next three weeks are for measuring, not panicking.

Start by pricing your real burn. Pull the last month of token usage from your agent pipelines and run it against the API rate card above. That single number tells you whether the credit covers you with room to spare or whether you blow through it in a week. If it is the latter, decide deliberately whether to enable usage-credit overflow at API rates or let the SDK hard-stop when the credit runs out, because the default of overflow-off means your automation just halts mid-job otherwise.

For anyone running a serious fleet, this is the moment to compare three paths head to head: a Max 20x plan with overflow, a raw API key with caching and batch turned on, and Cursor Ultra's larger envelope. Caching alone can move the answer, since agent loops are cache-friendly by nature. There is also the Zed workaround making the rounds, running the claude CLI interactively inside a terminal stays on subscription limits while routing through the agent protocol counts as programmatic. It works today, but architecting around the interactive-versus-programmatic line is betting against Anthropic tightening it, and that is not a bet we would build a pipeline on.

Claim the credit when the email arrives, since the opt-in is one-time and the credit is the difference between your agents running and stopping. Then put the real numbers in front of yourself. The pricing page has every Claude rate next to its rivals, and the calculator turns your token mix into a monthly figure you can decide on.

Sources

• Anthropic Help Center: Use the Claude Agent SDK with your Claude plan - June 15 date, credit amounts per tier, no-rollover, affected tools
• Anthropic: Claude API pricing - Opus 4.7 $5/$25, Sonnet 4.6 $3/$15, Haiku 4.5 $1/$5, cache and batch rates
• The Decoder: Claude subscriptions get separate budgets for programmatic use - Announcement coverage and the full-API-rate framing
• Zed: Anthropic subscription changes - Third-party harness reaction and the interactive CLI workaround
• XDA: Claude subscriptions no longer include Agent SDK and claude -p usage - May 13 announcement timing and background
• VentureBeat: Anthropic reinstates OpenClaw and third-party agent usage - OpenClaw reinstatement with the credit attached
• OpenAI: Codex pricing - Codex moved to API-token pricing on April 2, 2026