AI News — June 15, 2026

Model releases

• [2026-06-15] Anthropic — retired the original Claude Sonnet 4 (claude-sonnet-4-20250514) and Claude Opus 4 (claude-opus-4-20250514) across the Claude API, Claude on AWS, and Microsoft Foundry. Deprecated since April 14, the dated 4.0 strings now fail rather than fall back silently; the recommended replacements are Claude Sonnet 4.6 and Opus 4.8. The same docs note Claude Mythos Preview (claude-mythos-preview) retires June 30. The retirement lands the same day Anthropic’s Agent SDK billing split takes effect, so anything still pinning the old strings in CI or an agent config breaks today rather than gradually. (official)

  Because the request now errors instead of degrading quietly, the failure is loud — which is the better failure mode, but it means the fix is to grep CI configs, SDK initializers, and agent settings for the -20250514 suffixes today rather than waiting to notice a broken pipeline.

• [2026-06-14] MiniMax — its M3 open weights finally landed on Hugging Face, closing the “within 10 days” gap that had slipped past June 11: the MiniMaxAI/MiniMax-M3 repo is now live, with third-party quantizations already appearing, so the 1M-context sparse-attention model that posted a self-reported 59.0% SWE-Bench Pro is at last downloadable rather than API-only. A genuine open-weights drop, though the headline benchmark still awaits independent confirmation now that anyone can run it. (official, source)

  Open weights here means the model can run on your own hardware and be fine-tuned or inspected rather than only called over an API, and the third-party quantizations already appearing signal it will fit on smaller GPUs than the full checkpoint demands.

  For Solution Architects: A team whose data cannot leave its own network can now evaluate M3’s 1M-token context for whole-repository or long-document work on self-hosted infrastructure — an option that was off the table while it was API-only — with the caveat that the 59.0% SWE-Bench Pro figure is still self-reported and worth re-running on your own eval set before you commit to it.

Coding agents

• [2026-06-13] GitHub Copilot — Copilot CLI v1.0.62 shipped with per-subagent configuration: a subagent’s model, reasoning effort, and context tier can now be set via user settings or the /subagents picker. The release also lets plugins ship installable extensions through the marketplace and adds content search with match navigation to the diff view. The per-subagent model and effort controls are the substantive part — they let a CLI workflow route cheaper subtasks to a smaller model instead of running every step at one tier. (official)

  Until now the CLI ran every subagent at one tier, so the only knob was a global default; the /subagents picker makes the cost-versus-quality tradeoff a per-task, visible choice instead of an edit to a config file you have to remember exists.

  For Software Developers: In a multi-step task you can drop a file-search or formatting subagent to a smaller model and low reasoning effort through the /subagents picker while keeping high effort on the planning subagent — trimming token spend on the mechanical steps without downgrading the one step that actually needs the larger model.

Watch list

• Fable 5 / Mythos 5 restoration — both remain suspended under the June 12 US export-control directive; Anthropic says it is working to restore access “as soon as possible” but has given no date. Watch for a restoration date and whether access returns with nationality-based gating attached, which would likely outlast this incident. (official)

  Worth watching because the lasting question is not the outage itself but whether a temporary export-control response hardens into a permanent access-control regime; if gating returns attached, it sets a precedent other vendors may follow.

• xAI Grok V9-Medium (mid-June) — the ~1.5-trillion-parameter, Cursor-trained coding model finished training in late May and is still expected this month, but mid-June has arrived with no model card, API string, or release. (prior coverage)

  This matters once it actually ships: a model trained specifically on Cursor coding data is a direct bid at the agentic-coding tier, and it only becomes significant when there is an API string to call and an independent SWE-bench number to compare.

• Gemini 3.5 Pro (late June) — reportedly in limited Vertex AI enterprise preview with GA targeted late June; still no model card, public benchmarks, or pricing for the promised 2M-token context and Deep Think mode. (prior coverage)

  The signal to watch for is pricing and published benchmarks alongside GA, since the 2M-token context and Deep Think mode only become real planning inputs once teams can weigh their cost against the current frontier options.