AI News — June 25, 2026

Model releases

• [2026-06-24] Anthropic — in a letter to US senators and White House officials (dated June 10, surfaced publicly June 24 via Bloomberg, CNBC, and Reuters), Anthropic accused operators affiliated with Alibaba and its Qwen AI lab of carrying out the largest known distillation attack on its models to date: 28.8 million exchanges with Claude through roughly 25,000 fraudulent accounts between April 22 and June 5, deliberately targeting Claude’s most commercially valuable capabilities — software engineering and agentic reasoning. Distillation here means training a cheaper model on a stronger one’s outputs; Anthropic frames this as systematic, unauthorized harvesting of US frontier-model capability to build a rival at a fraction of the cost, and says Alibaba “ignored the Trump Administration’s warnings.” An Anthropic spokesperson said combating illicit distillation “requires coordinated action between government and industry.” Alibaba did not immediately comment. It matters because it shifts the US–China AI fight onto a new axis: alongside the June 12 export-control suspension of Anthropic’s own Fable 5 / Mythos 5, the company is now arguing the bigger leakage risk is a competitor pulling capability out through ordinary API access — a framing aimed directly at the policymakers writing frontier-model rules. (Bloomberg, CNBC)

  If this framing lands with the policymakers Anthropic is lobbying, the fallout reaches ordinary API consumers rather than just Alibaba: the defense against harvesting-at-scale is tighter account verification, per-account rate ceilings, and output-pattern monitoring, so the friction a few thousand fraudulent accounts created gets paid down by every legitimate integration that now has to prove it isn’t one.

Watch list

• Alibaba / Qwen distillation fallout — Anthropic’s complaint is now in front of the Senate Banking Committee (Chair Tim Scott, Ranking Member Elizabeth Warren) and the White House. Watch for Alibaba’s formal response, whether the administration acts on the “coordinated government–industry” ask Anthropic is pushing, and whether other US labs surface similar distillation claims — the accusation reframes “model security” from jailbreaks and weight theft toward abuse of legitimate API access at scale. (Bloomberg)

  What would turn this from one company’s complaint into a sector-wide shift is corroboration: if a second major lab publishes its own distillation telemetry, “who is allowed to call a frontier API, and how do you prove it” moves from Anthropic’s policy ask toward an industry default, and KYC-style verification on model endpoints stops being optional.

• Fable 5 / Mythos 5 restoration — both still suspended under the June 12 export-control directive, now 13 days on with no restoration date; usage-credit metering has been live since June 23, and Anthropic’s updated consumer terms — reportedly effective July 8 — enumerate identity-verification data that analysts read as the likely mechanism for a US-users-first return. Prediction markets still price roughly even odds of restoration before July 1. Watch for a date and whether access returns gated by identity or nationality. (official notice, terms update, prior coverage)

  The identity-verification path back and today’s lead story point at the same machinery from opposite ends — attributable, verified access is both how Anthropic would satisfy the export directive and how it would stop the next 25,000-account harvesting run — so the thing to watch on July 8 is whether the terms read as a one-off unlock for Fable 5 or the start of identity becoming a standing condition of frontier-API access.

• Gemini 3.5 Pro (late June) — still a limited Vertex AI enterprise preview with GA targeted for June; with month-end days away there’s still no model card, public benchmarks, or pricing for the promised 2M-token context and Deep Think mode. The open question is now whether Google holds the June date or lets it slide into Q3. (prior coverage)

  The figure that decides whether the 2M-token window is usable or just a spec-sheet number is per-token price at that length: a context that large only changes how you build — whole-repo or whole-corpus prompts instead of retrieval — if filling it is something a budget can absorb, and that’s the one number the preview still withholds.

• xAI Grok V9-Medium — the ~1.5-trillion-parameter, Cursor-trained coding model is reportedly live in the consumer Grok product on X and SuperGrok but still has no API string or model card on x.ai, so nothing testable has shipped for anyone building on it. (prior coverage) (unconfirmed)

  Should the API string and model card finally appear, the test worth running is whether being trained on Cursor’s editing traces actually buys measurably better real-PR performance than a general model — a coding-specific training signal is the whole bet here, and until it’s callable there’s no way to check it.

• MCP spec finalization (July 28) — the 2026-07-28 release candidate (stateless core, Extensions framework, Tasks, MCP Apps, authorization hardening, formal deprecation policy) is in its validation window; watch for the rest of the SDK support landing before the cutover. (official, prior coverage)

  For anyone maintaining a server today, the piece to track is SDK timing rather than the spec text: if your language’s SDK lands Tasks and the Extensions framework only after the July 28 cutover, you’re stuck straddling the old and new wire formats during the gap — so the risk to plan around is an SDK that trails the release, not the release itself.