New AI model releases, papers, and open source projects, June 14 to 15 2026

Why "what released on June 14 to 15" has no clean answer

People search this expecting a dated changelog for the whole field. There is not one. Hugging Face ranks models, datasets, and spaces by a trending score that decays over hours; GitHub Trending does the same for repositories. Both answer "what is popular right now," which drifts every time you reload. The records that genuinely pin to a calendar day are narrower: a tagged release, a commit timestamp, a model-card revision.

For this two-day window the ecosystem itself was quiet. Release trackers like llm-stats list Kimi K2.7 Code (Moonshot AI) as the nearest frontier release, and it landed on June 12, two days before the window opened. No major open-weight model dropped a downloadable file on June 14 or 15. So the honest way to read the window is to pick a record that carries a real date and read it, rather than screenshotting a trending list that will look different tomorrow.

The window in numbers, from one repo's git log

This running series tracks one open-source macOS agent, Fazm, day by day, because a single repo gives you something the ecosystem-wide trending page cannot: exact timestamps you can re-run yourself.

An empty day is not a non-event; it is a maintainer resting, and it tells you the unit of a "new project" is a release, not a calendar square. June 15 made up for the weekend with 18 commits and three tagged builds in a single day.

The break a model release caused, before anyone announced it

Here is the part the trending pages cannot show you. The load-bearing change of June 15 was not a feature. It was a fix for a failure that a model-provider catalog change introduced silently. As Fazm's adapter observed it in June 2026, Anthropic folded Opus into the default pseudo-model (now carrying a 1-million-token-context label) and removed the standalone Opus entry from the catalog.

Fazm had long dropped default outright, because it used to be a pure 1M-context credit trap. Once Opus moved inside default, that blanket drop quietly took Opus (shown in the UI as "Smart") out of the model picker. No error, no notice, just a missing option. The served model under the hood was already claude-opus-4-8, even though the advertised version text still read "Opus 4.7". That mismatch is exactly why the fix matches on the word opus, not on a version number.

Reading the fix at the line level

No roundup of this window reads a dated change line by line, because most of what they cover has no public diff to read. This one does. Commit ac94f9ed changed acp-bridge/src/index.ts by 18 insertions and 6 deletions. The constant LATEST_OPUS_MODEL_ID is defined one line above the changed block as "claude-opus-4-8".

// acp-bridge/src/index.ts  (commit ac94f9ed, +18 -6, v2.9.67, 2026-06-15 13:18 PT)
const LATEST_OPUS_MODEL_ID = "claude-opus-4-8";   // one line above the fix

const transformed = availableModels
-  .filter(m => m.modelId !== "default")            // old: blanket-drop "default"
   .map(m => {
     const stripped = m.modelId.replace(/\s*\[1m\]/gi, "");
+    const haystack = `${m.modelId} ${m.name} ${m.description ?? ""}`.toLowerCase();
+    const isOpus = stripped.includes("opus") || haystack.includes("opus");
+
+    // 2026-06: Anthropic folded Opus INTO `default` (1M-context label) and
+    // REMOVED the standalone opus entry, so blanket-dropping "default"
+    // silently strips Opus ("Smart") from the picker. Remap instead.
+    if (stripped === "default") {
+      if (!isOpus) return null;                     // non-Opus default: still dropped
+      return { modelId: LATEST_OPUS_MODEL_ID, name: "Opus", description: m.description };
+    }
     // Legacy: a standalone claude-opus-4-7 gets bumped to the latest id.
     const modelId = stripped === "claude-opus-4-7" ? LATEST_OPUS_MODEL_ID : stripped;
     return { ...m, modelId, name: m.name.replace(/\s*\[1m\]/gi, "").trim() };
   })
+  .filter((m): m is { modelId: string; name: string } => m !== null);

The shape of the fix is the lesson. The old code trusted a string id (default) to mean one fixed thing forever. The new code asks a question instead: is this entry Opus-bearing? It checks the id, the name, and the description, remaps the Opus-bearing default to the canonical Opus id, and sends it explicitly via session/set_model so it sidesteps the 1M-context credit gate. A non-Opus default is still dropped. That is the whole change: 24 lines touched, shipped the same afternoon the break appeared.

The rest of the June 15 burst

The Opus fix was the headline, but three releases went out that day. Each is a small, reviewable diff in the same public log.

The history-limit bump is the one most aligned with what wrappers around Claude Code exist for: keeping full conversation state reachable instead of letting older sessions silently disappear. A 20-row cap is a small auto-hide; raising it to 500 is the same instinct as refusing to auto-compact a live chat. See the read on controlling context compaction for why that instinct matters.

Why a 24-line picker fix is the better answer than a trending screenshot

If you came here to find the model everyone launched on June 14 to 15, the honest answer is that nobody did; the nearest frontier release was two days earlier and the weekend was quiet. If you came to find something new you could actually run and read inside the window, the answer is a desktop agent release and its diff. That is not a consolation prize. "New model releases" rarely break in a press release; they break in the tools that wrap them, when a catalog reshuffle quietly removes an option. Watching an open-source project catch that and fix it in 24 reviewable lines tells you more about whether the tool will survive the next model release than any benchmark a vendor self-reports.

Fazm is that kind of wrapper: it runs the real Claude Code agent loop (and Codex) over ACP inside a native macOS app, with sessions that survive a restart, one-click chat forking, and no auto-compacting of context. The Opus-catalog fix is one small entry in a public log you can read the same way for any other day. See the May 31 to June 1 read for the previous window, or the note on Anthropic's latest model for why claude-opus-4-8 is the served flagship.

Questions people ask about this window