AI News — June 14, 2026

Model releases

• [2026-06-12] Anthropic — suspended all access to Claude Fable 5 and Claude Mythos 5 to comply with a US government export-control directive, received at 5:21pm ET, ordering it to cut off access by any foreign national — inside or outside the US, including foreign-national Anthropic employees. Unable to enforce that selectively, Anthropic disabled both models for every customer; all other Claude models are unaffected. The government cites a potential “jailbreak” — by Anthropic’s account a narrow technique that amounts to asking the model to read a codebase and fix its software flaws. Anthropic is complying but publicly disagrees, calling the technique narrow, already known, and present in rival models, and warning that recalling a commercial model on this basis would “halt all new model deployments for all frontier model providers”; it says it is “working to restore access as soon as possible,” with no date. Three days after Fable 5’s splashy GA, the model that tops SWE-bench Pro (80.3%) vanished from the API and every plan — the first time a frontier model has been pulled from the market by government order. (official, source, source)

  For anyone who treats an API model string as a stable contract, this is a different failure mode than the dated, announced deprecations the API normally guarantees: an export directive can take a model dark with hours of notice and no migration path, so “pin it on the API” — the usual hedge against silent in-place swaps — does nothing against a regulatory recall.

Coding agents

• [2026-06-12] Anthropic — the Fable 5 suspension pulls the model out of the coding tools it shipped into day-one on June 9: it was generally available in GitHub Copilot’s picker and Claude Code, and on Amazon Bedrock and Microsoft Foundry — and because the directive covers all access to the model, none of those surfaces can serve it while the cutoff holds. Teams that spent the included-access window (June 9–22) standardizing a workflow on the SWE-bench leader now have it gone with no restoration date and should fall back to Opus 4.8 or Sonnet. It’s a sharp reminder that betting a workflow on a single just-released frontier model carries availability risk well beyond pricing. (official, official, source)

  The fallback isn’t automatic, either: anything that hard-coded the claude-fable-5 model string — a CI job, an agent definition, a Bedrock or Foundry deployment config — now errors rather than gracefully dropping to another model, so getting work moving again means editing those references over to Opus 4.8 or Sonnet by hand, not waiting for the directive to lift.

Watch list

• Fable 5 / Mythos 5 restoration: suspended since June 12 under the export directive; Anthropic says it is working to restore access “as soon as possible” but has given no date. The June 22 plan-window deadline is moot while the model is dark — watch for (a) a restoration date and (b) whether the export-control framing reshapes the original June 23 move to usage-credit gating. (official)

  What matters as much as the date is the shape of the restoration: if access returns wrapped in nationality-based gating to satisfy the directive, that compliance layer likely outlives this incident and becomes a standing condition of frontier-model access rather than a one-off.

• Anthropic Agent SDK billing split — Monday, June 15 (tomorrow): Claude Agent SDK, claude -p, and third-party SDK apps move off plan limits onto a separate per-user Agent SDK credit that must be claimed once and can’t be pooled or shared. (official)

  This is the watch-list item that resolves first — tomorrow morning — and the open unknown is non-human logins: a per-user credit a service account can’t claim could hard-stop SDK automation the instant the meter flips, so anyone with Claude in CI wants to confirm the claim path over the weekend rather than discover it Monday.

• Gemini 3.5 Pro (late June): still in limited Vertex AI enterprise preview — no model card, public benchmarks, or confirmed pricing for the promised 2M-token context and Deep Think mode. (source) (unconfirmed)

  Under three weeks to the late-June target and still no model card; a published benchmark and rate are what would make this plannable, since a 2M-token context and Deep Think stay specs no one outside the allowlist can budget against until then.

• xAI Grok V9-Medium (mid-June): the ~1.5-trillion-parameter, Cursor-trained coding model finished training May 25 and is still expected this month, but there is no model card, API string, or release yet. (source) (unconfirmed)

  Mid-June has arrived with nothing shippable to show; an independent coding benchmark is the tell to wait for, since “Cursor-trained” and a 1.5-trillion-parameter count are positioning, not evidence the model actually challenges the coding leaders.

• MiniMax M3 weights — still overdue: the “within 10 days” window closed June 11 and the MiniMaxAI Hugging Face org still tops out at M2.7; the 59.0% SWE-Bench Pro claim stays self-reported until the weights and technical report land. (source) (unconfirmed)

  Three days past the deadline with the org still at M2.7, the silence is itself the signal — for an open-weights pitch the downloadable weights are the product, so the 59.0% remains a number with nothing to reproduce it against until a repo and report appear.