Dated release log, verified June 22, 2026

New AI model releases, papers, and open source projects: June 12-13, 2026

I went looking for the new models everyone assumes shipped in this window. The honest finding: the biggest AI story of June 12-13 was a model being pulled, not launched. Below is the verified, dated list, then the part no recap covers: the exact file in Fazm that lets the same agent keep working when a model disappears.

M
Matthew Diakonov
7 min read
Direct answerverified June 22, 2026

No new frontier or open-weights model launched inside the June 12-13, 2026 window. The largest AI item was the June 13 US government directive that prompted Anthropic to suspend access to its Fable 5 and Mythos 5 models for foreign nationals. The genuine in-window releases were developer tooling and research: luau-wasm 0.1a0 and a SQLite column-provenance writeup (June 13), and an OpenAI GPT-Realtime-2 audio tool (June 12). GLM-5.2, the week's open-weights headline, is dated June 16-17, just outside this window.

0
new models in the 48-hour window
0
verified items dated Jun 12-13

Primary sources: simonwillison.net Jun 13, Jun 12, and anthropic.com.

What actually shipped, dated honestly

Each row is tied to a primary source with a real timestamp. Rows marked in window have a source dated June 12 or June 13. The one model row is labelled context because its release date sits outside the window, exactly the row most recaps quietly backdate.

Policy/Anthropic statementJun 13in window

US directive to suspend access to Fable 5 and Mythos 5

The single biggest AI item dated inside the window, and it is the opposite of a launch. Anthropic published a statement responding to a US government directive to suspend access to two of its models for foreign nationals on national-security grounds. If you are searching this window for a shiny new model, the real headline is a model being pulled.

anthropic.com, Jun 13, 2026
Open source/Simon WillisonJun 13in window

luau-wasm 0.1a0

A genuine June 13 open-source release: an early alpha compiling the Luau language to WebAssembly. Small, specific, real. This is what actually shipped in-window, not a model.

github.com/simonw/luau-wasm, Jun 13, 2026
Research/Simon Willison researchJun 13in window

Mapping SQLite result columns back to source table.column

A dated June 13 research writeup on recovering the originating table and column for each cell in an arbitrary SQLite query result. The closest thing to a 'paper' inside the window, and it is engineering research, not an LLM benchmark.

github.com/simonw/research, Jun 13, 2026
Tool/Simon WillisonJun 12in window

OpenAI WebRTC audio session with document context

A June 12 browser tool for talking to OpenAI's GPT-Realtime-2 over WebRTC with optional pasted-document context. A wrapper around an existing model, not a new model.

tools.simonwillison.net, Jun 12, 2026
Model/Z.aioutside windowcontext

GLM-5.2 (open weights, Z.ai)

The week's actual open-weights headline, but the MIT-licensed weights and the detailed writeup landed June 16, three days after this window. A 753B-parameter mixture-of-experts model with a 1M-token context window. If a 'June 12-13' page leads with GLM-5.2 as fresh same-day news, it has the date wrong.

simonwillison.net, Jun 17, 2026

The story of this window is a model being pulled

Most dated roundups for any given week are built on an assumption: something new launched, so list it. June 12-13, 2026 breaks that assumption. The most consequential AI event with a timestamp in this window was Anthropic's June 13 statement on a US government directive to suspend access to Fable 5 and Mythos 5 for foreign nationals. A model that engineers were building on became, for a group of users, unavailable, overnight, by external decree.

That is the uncomfortable lesson behind the whole "what released today" genre. The model is not the durable part of your setup. Access can be revoked, models get deprecated, providers change terms, a government issues a directive. The durable part is the layer above the model: your agent loop, your sessions, your tools. If that layer is welded to one model endpoint, a window like this one is a problem. If it is model-agnostic, this window is a footnote.

The part no recap covers: repointing the agent in one file

Fazm is a native macOS app that wraps Claude Code and Codex through the Agent Client Protocol (ACP). Because the agent runs as a spawned process the app controls, the model endpoint is just an environment variable Fazm sets on launch. When a model gets restricted, you point the same loop somewhere else. Here is the actual code path, not a marketing diagram.

What happens when a model you used becomes unavailable

Model restricted

Provider deprecates, revokes, or a directive pulls access

2

Set custom endpoint

Paste any Anthropic-compatible gateway URL in Settings

⚙️

Bridge rewrites env

ACPBridge sets ANTHROPIC_BASE_URL, drops the bundled key

Same loop continues

Chat, history, and tools stay intact

The mechanism lives in Desktop/Sources/Chat/ACPBridge.swift. Fazm reads the customApiEndpoint setting, validates it through validCustomAPIEndpoint(), and only on a valid absolute http(s) URL forwards it to the agent process. A malformed value is ignored and chat falls back to the default endpoint, so a typo never silently bricks the app.

env Fazm sets on the spawned ACP agent (ACPBridge.swift)

Two details make this honest rather than hand-wavy. First, the bundled Anthropic key is replaced with the placeholder sk-fazm-custom-endpoint so it is never leaked to a proxy you route through. Second, Codex is a fully separate swappable backend, managed by CodexBackendManager, so you can move a chat from Claude to a GPT-5.x model without losing the window. The agent loop, not the model, is the asset Fazm protects.

Want a setup that survives the next model getting pulled?

Walk through routing Fazm through your own gateway and swapping backends per chat, live.

Questions about the June 12-13, 2026 window

Frequently asked questions

What new AI models were released on June 12-13, 2026?

No new frontier or open-weights model launched inside the exact June 12-13, 2026 window. The verified dated items were developer tooling and research: luau-wasm 0.1a0 (a Luau-to-WebAssembly alpha) and a SQLite column-provenance research writeup on June 13, and an OpenAI GPT-Realtime-2 WebRTC audio tool on June 12. The most significant AI news dated to June 13 was a policy event, not a release: Anthropic's statement on a US government directive to suspend access to its Fable 5 and Mythos 5 models for foreign nationals.

Was GLM-5.2 released on June 13, 2026?

GLM-5.2 is the week's open-weights headline, but its MIT-licensed weights and the widely-cited writeup are dated June 16-17, 2026, not June 12-13. Z.ai began rolling it toward coding-plan subscribers around mid-June, so some recaps stretch its date back into this window. To be precise: it is adjacent context for June 12-13, not an in-window launch.

What was the Fable 5 and Mythos 5 suspension about?

On June 13, 2026 Anthropic published a statement responding to a US government directive to suspend access to two of its models, Fable 5 and Mythos 5, for foreign nationals on national-security grounds. It is the clearest signal of why model access is not guaranteed: the specific endpoint you build on can be restricted, deprecated, or pulled with little notice.

Why does a tool's model-agnostic design matter for a window like this?

When the biggest news in a window is a model being pulled rather than launched, the durable asset is the layer above the model: the agent loop, your sessions, your tool wiring. Fazm wraps Claude Code and Codex through the Agent Client Protocol, lets you swap the backend per chat, and exposes a custom API endpoint that overrides ANTHROPIC_BASE_URL, so the same loop can point at any Anthropic-compatible gateway if one model becomes unavailable.

How does Fazm actually repoint the agent at a different model?

In Desktop/Sources/Chat/ACPBridge.swift, Fazm reads the customApiEndpoint setting, validates it through validCustomAPIEndpoint(), and on a valid http(s) URL sets env["ANTHROPIC_BASE_URL"] on the spawned ACP agent process while replacing the bundled key with a placeholder ("sk-fazm-custom-endpoint") so it never leaks to your proxy. Codex is a separate swappable backend managed by CodexBackendManager. The chat, its history, and its tools stay intact across the swap.

Where can I verify the June 12-13 dated items myself?

The per-day archive pages on simonwillison.net timestamp each item: simonwillison.net/2026/Jun/12/ and simonwillison.net/2026/Jun/13/. The Fable 5 / Mythos 5 statement is on anthropic.com. Every claim on this page links to its dated primary source so you can check the date rather than trusting a roundup that smeared a whole week together.

How did this page land for you?

React to reveal totals

Comments ()

Leave a comment to see what others are saying.

Public and anonymous. No signup.