Two compliance events converge this week with a Chinese lab development that puts the open-weight coding market in sharper focus. An operator who takes no action today risks an automation outage tomorrow.
Anthropic ends flat-rate agent access — credit pool launches 15 June
From tomorrow, any usage of the Claude Agent SDK, the claude -p command, Claude Code GitHub Actions, and third-party applications authenticating through the Agent SDK will no longer draw from a subscriber's standard usage pool. Anthropic moves these to a separate monthly credit (USD 20 for Pro, USD 100 for Max 5x, USD 200 for Max 20x) billed at full API rates, with no pooling across users and no rollover.
The credit is opt-in. Users who have not yet claimed it will find automations stop when their current plan limits are reached, unless overflow billing has been enabled. Interactive use of Claude (chat, Claude Code in a terminal, Claude Cowork) is unaffected.
The practical impact depends on how deeply Agent SDK calls are embedded in a team's pipelines. A CI pipeline that makes a few API calls per commit will likely stay within the credit ceiling; a product that routes customer requests through an agent loop will not. Operators should audit which workflows authenticate through the Agent SDK, estimate their monthly token volume, and either migrate to direct API billing or accept the credit cap before today ends. The New Stack details the billing mechanics and the opt-in claim process; TechTimes has the timeline.
Moonshot AI opens Kimi K2.7-Code — with benchmark caveats
On 12 June, Moonshot AI released Kimi K2.7-Code on Hugging Face under a Modified MIT licence. The model is a Mixture-of-Experts architecture with 1 trillion total parameters, 32 billion active per forward pass, 384 experts, and a 256,000-token context window, designed for long-horizon, agentic software engineering. API access is priced at USD 0.95 per million input tokens and USD 4.00 per million output tokens. It is the fifth major release from Moonshot in under a year.
Moonshot reports gains of 21.8% on Kimi Code Bench v2, 11.0% on Program Bench, and 31.5% on MLS Bench Lite versus K2.6, alongside a roughly 30% reduction in reasoning-token usage. VentureBeat notes that all three are proprietary first-party suites and that the model has not yet been submitted to DeepSWE, the independent benchmark now widely used as a signal for model routing decisions. The token-efficiency claim is verifiable at the API level; the headline performance deltas should be treated as directional until independent results arrive.
The release coincides with Bloomberg reporting that Moonshot AI is seeking up to USD 2 billion in a new funding round at a USD 30 billion valuation, a sevenfold increase since December 2025 when the company was valued at approximately USD 4 billion. Kimi's annual recurring revenue reached USD 200 million in April. The combination of a permissively licensed frontier-class coding model and venture-scale backing positions Moonshot as a credible open-weight challenger to US and European labs in the agentic coding segment.
EU AI Act Article 50: 49 days, no extension
The Digital Omnibus agreement reached in May moved the high-risk AI compliance deadline from 2 August 2026 to 2 December 2027. It did not move the Article 50 transparency obligations, which remain effective 2 August 2026, now 49 days away.
From that date, providers of chatbots and virtual assistants must design them so that users are informed they are interacting with an AI system. Providers of systems generating synthetic audio, image, video, or text must mark outputs in a machine-readable format and make them detectable as AI-generated. Systems already on market before 2 August receive a four-month grace period on machine-readable watermarking only; the user-disclosure requirement itself takes effect immediately.
The European Commission published draft implementation guidelines in May. For any operator with a product reaching EU users that includes a conversational interface or generative content output, compliance work should be in final testing now. The Digital Omnibus gave high-risk AI builders 16 extra months; it gave the rest no additional time.
The day's items share a single logic: the cost and accountability of deploying AI is hardening. Anthropic is formalising the price of autonomous agents; the EU is removing the last grace period on conversational AI disclosure; and Moonshot is opening frontier-class coding capability at per-token prices US labs do not match. Operators who designed their AI infrastructure around flat-rate access or voluntary disclosure will be rebuilding parts of that architecture in the weeks ahead.