Article by Paul Sawers published on The New Stack on June 16, 2026, about the suspension by Anthropic — "on the very day it was scheduled to go live" — of the billing split meant to separate Agent SDK usage from Claude subscription limits. Anthropic's cited message: *"We're pausing the changes to Claude Agent SDK usage described below.
Article by Paul Sawers for The New Stack, published on June 16, 2026: Anthropic suspends, "on the very day it was scheduled to go live," the billing split meant to move Agent SDK usage out of Claude subscription limits. The message to subscribers is brief — "We're pausing the changes to Claude Agent SDK usage described below. For now, nothing has changed."
The immediate context. The decision comes after a difficult week: on June 9, Anthropic released Fable 5 and Mythos 5, its first generally available Mythos-class models, equipped with hardened cybersecurity safeguards; a few days later, an export control directive from the US government forced it to withdraw both models for all its customers worldwide. The pricing suspension then appears as "a little good news" offered to a rattled developer base.
What was at stake. For third-party tools built on the Agent SDK, the change was not trivial. Zed's post, signed by Franciska Dethlefsen, noted that subscriptions were subsidizing this usage by roughly 15 to 30× the equivalent API cost — a figure the article explicitly attributes to an analysis by engineer Matthew Diakonov — and that the new credits would be billed at full API rate. Zed pointed to a workaround: launching the official Claude CLI in a terminal rather than going through the Agent SDK kept subscription limits intact. Hence the article's phrase, set off in its own paragraph: "The same tool, billed differently depending on how you invoked it."
The timing damage. Companies that had already passed the change on find themselves caught out. Conductor, a multi-agent coding tool built on the Agent SDK, has to issue a denial to its customers.
The underlying tension. It extends beyond Anthropic. As early as April, Boris Cherny, head of Claude Code, justified an earlier restriction by explaining that subscriptions "weren't built for the usage patterns of these third-party tools" — an admission that flat-rate plans and open-ended agentic usage don't mix. GitHub reached the same conclusion and acted on it, removing in June Copilot's flat-rate premium requests model in favor of token-based billing, despite protests. The same week, a proposed class action was filed in California, alleging that Max tiers fall well short of the multipliers advertised for intensive coding sessions.
The final take. Between government pressure, a planned IPO, and rumored price cuts at OpenAI, Anthropic is trying to keep its developer base on its side — and the suspension contributes to that, for now.
Key takeaways
The fact. Anthropic suspends, on the very day it took effect (June 15), the billing split meant to move Agent SDK usage out of subscription limits. "We're pausing the changes to Claude Agent SDK usage described below. For now, nothing has changed." No date announced for a revised version — only "working to update the plan to better support how users build with Claude subscriptions."
The sentence to remember from the whole story."The same tool, billed differently depending on how you invoked it." It names what the announced regime actually did — pricing the invocation channel rather than usage. A pricing policy that distinguishes technically equivalent paths mechanically produces workaround arbitrage, not usage moderation. Cf. [[dethlefsen-zed-anthropic-subscription-changes-2026-05-14]], whose very first recommendation was precisely this workaround.
The structural admission, and the most durable one.Boris Cherny (head of Claude Code), in April, during an earlier restriction — subscriptions "weren't built for the usage patterns of these third-party tools." The article rightly reads this as "an acknowledgement that flat-rate pricing and open-ended agentic usage don't mix well."This observation outlives the suspension. The regime was cancelled; the incompatibility between flat-rate plans and open-ended agentic usage was not. → Any return of this issue was predictable as early as April.
The industry precedent that sets the trajectory.GitHub removed in June Copilot's flat-rate premium requests model in favor of token-based billing — "a change that drew its share of complaints but went ahead regardless." Two major players, same diagnosis, opposite outcomes: GitHub held firm, Anthropic backed down. The comparison is the article's main analytical contribution: it turns an incident into a sector-wide trend and suggests that Anthropic's retreat owes more to context than to substance.
The context that explains the retreat (three simultaneous pressures). 1. Fable 5 and Mythos 5, released on June 9 with hardened cybersecurity safeguards, withdrawn for all customers worldwide a few days later under a US government export control directive. Cf. [[anthropic-claude-fable-5-mythos-5-2026-06-09]]. 2. A planned IPO. 3. Rumored price cuts at OpenAI — marked "rumored" by the author, not to be overstated. Cf. [[sfeir-gpt56-sol-terra-luna-coding-agentique-pricing-2026-07-13]]. → Reading: the suspension is as much a developer-base retention move in a difficult week as it is a pricing-policy revision. Anthropic could not afford a second source of discontent.
The timing damage, rarely noted. companies that had already passed on the change to their customers find themselves having to issue denials. Conductor (a multi-agent coding tool built on the Agent SDK) publishes "Anthropic has delayed the subscription updates to Claude plans. You can continue to use your Claude plan with Conductor as normal." → The real cost of a platform U-turn: it is not measured at the provider but at the third parties who adjusted product, pricing, and messaging around an announcement. Worth keeping in mind for any dependency on a platform whose pricing policy is not contractual.
Traceability of the 15-30× figure. the article is the source that traces the chain back — Zed states it, but attributes it to an analysis by engineer and entrepreneur Matthew Diakonov. It is neither Anthropic data nor a Zed measurement. Always cite it as a third-party estimate. Zed's post, which does not source it, is corrected here by the press.
The class action, to be followed separately.the same week, a proposed class action was filed in a California federal court, alleging that Max tiers "fall well short of their advertised usage multipliers during heavy coding sessions."Proposed — a contemplated action, not certified; draw no conclusions on the merits. But the signal matters: the dispute over the gap between advertised multiplier and actual throughput moves from public forum to the courtroom.
Consolidated timeline of the story. (useful for any public statement): April — first access restriction for third-party tools, Cherny's statement → May 13 — announcement of the split and Agent SDK credits → May 14 — Zed's post ([[dethlefsen-zed-anthropic-subscription-changes-2026-05-14]]) → June 15 — planned effective date, suspension the same day → June 16 — this article + addendum to Zed's post. The broader timeline appears in girard-acp-deux-protocoles-un-sigle-2026-08-02.
What to take away for decision-making. the question is not "does my subscription cover my agent today?" (an unstable answer) but "does my architecture survive a change in pricing regime?" This is exactly Zed's optionality argument — and it is validated here by a full round trip in one month.
Meta / to link. a direct follow-up to dethlefsen-zed-anthropic-subscription-changes-2026-05-14, which it corrects (source of the 15-30×) and complements (context, Copilot precedent, lawsuit); model context in anthropic-claude-fable-5-mythos-5-2026-06-09; Boris Cherny in cherny-wu-reflecting-year-claude-code-2026-07-17; competitive pricing pressure in sfeir-gpt56-sol-terra-luna-coding-agentique-pricing-2026-07-13; token FinOps stakes in tokenomics-foundation-linux-finops-token-economics-about-2026-06-03 and gupta-token-budget-wars-marginal-token-utility-2026-05-28; the protocol involved in agentclientprotocol-introduction-2026-08-02.
Key figures
une subvention de l'usage agentique par l'abonnement de l'ordre de 15 à 30× le tarif API
"We're pausing the changes to Claude Agent SDK usage described below. For now, nothing has changed."
— Anthropic
subscriptions were not designed for third-party tool usage patterns
— Boris Cherny
les paliers Max restent très en deçà des multiplicateurs d'usage annoncés en session de codage intensive
— recours collectif envisagé en Californie
entre contrôle à l'export, projet d'introduction en bourse et baisses de prix supposées chez OpenAI, Anthropic cherche à retenir sa base de développeurs
— Paul Sawers
The knowledge graph extracted from this fiche — 8 entities, 17 relations.