AI News — June 13, 2026

Model releases

• [2026-06-12] OpenAI — retired GPT-5.2 from ChatGPT: GPT-5.2 Instant, Thinking, and Pro are no longer selectable, and any existing conversation on a GPT-5.2 model automatically continues on the matching GPT-5.5 model. The cut follows OpenAI’s stated policy of keeping a ChatGPT model available for ~90 days after its successor ships (GPT-5.3 here). This is a lifecycle, not a launch — but it removes a model some workflows had pinned, and the silent in-place migration to GPT-5.5 means anyone who tuned prompts or evals against GPT-5.2’s behavior is now running on a different model without changing anything. (official, source)

  The durable lesson is that a ChatGPT model selection was never a stable contract: anything that genuinely needs a fixed model version belongs on the API, where deprecations are dated and announced rather than swapped underneath an open conversation.

Coding agents

• [2026-06-12] OpenAI — Codex gained a browser Developer mode with an opt-in “Enable full CDP access” toggle (Settings → Browser): with the Chrome DevTools Protocol open, Codex can inspect network requests, monitor console logs, profile performance, and diagnose runtime errors in a running web app, plus Windows support for Computer Use so it can drive desktop apps on Windows. The same update bundle added access controls and usage-limit guidance. Giving the agent CDP access turns web debugging from “paste the stack trace” into the agent reading the live console and network panel itself — the data a developer would otherwise screenshot or copy in by hand. (official, source)

  That same CDP channel is why the access is opt-in and toggled per session rather than on by default: full DevTools visibility means the agent can read whatever crosses the network panel — auth headers, cookies, response bodies — so it’s reach you grant deliberately, which is also why the update shipped alongside access controls.

  For Software Developers: A frontend bug that only surfaces at runtime — a fetch that comes back 500 with nothing thrown client-side, say — used to mean reproducing it yourself and pasting the failing request and console output back to the agent. With CDP access on, Codex opens the page, walks the path, and reads that 500’s response body and the console trace itself, so the loop from “it’s broken” to a proposed fix no longer waits on you to relay what the panels show.

• [2026-06-12] Anthropic — Claude Code shipped point releases (v2.1.174/176) that fix the /model picker to show the family that “Default” resolves to (Opus as its own row on Max/Team Premium/Enterprise, Sonnet on Pro/Team) and stop the “Fable 5 is now consuming usage credits” banner from showing on enterprise usage-based accounts; session titles are now generated in the conversation’s language. Minor on their own, but the model-picker and Fable 5 banner fixes are housekeeping around this month’s two Anthropic metering changes — the Agent SDK split on the 15th and Fable 5 leaving flat-rate on the 22nd. (official)

  Of the two fixes, the /model picker correction is the one with teeth once metering arrives: “Default” silently resolving to Opus versus Sonnet is a real cost difference, so seeing which family it lands on stops being cosmetic the moment those tiers start drawing on credits.

Watch list

• Anthropic Agent SDK billing split — Monday, June 15 (2 days out): now officially documented — 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. Anyone running Claude in CI or under a service account has the weekend to claim and decide on overflow credits. (official)

  Two days out, the open question is what actually happens Monday morning: whether a job already running under a service account keeps drawing on plan limits until someone claims the credit, or hard-stops the instant the meter flips — the difference between a quiet migration and a Monday outage for anyone with Claude in CI.

• 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 as of today there’s no model card, API string, or release; with xAI now inside a public SpaceX, the launch lands under earnings-cycle scrutiny. (source) (unconfirmed)

  Mid-June is now, and a model that finished training May 25 with still no model card or API string is the gap to watch close — what would make it matter is an independent coding benchmark, since “Cursor-trained” and a 1.5-trillion-parameter count are positioning, not evidence it actually beats 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; until the weights and technical report land, the 59.0% SWE-Bench Pro claim stays self-reported and the model stays one you can’t actually run. (source) (unconfirmed)

  The concrete tell is the MiniMaxAI org still topping out at M2.7 — until a new repo and the technical report land there, the 59.0% SWE-Bench Pro is a self-report with nothing to reproduce it against, and an open-weights model nobody can download is just a press release.

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

  With GA now pushed to late June, pricing is the gap that matters more than the date: a 2M-token context and a Deep Think mode are expensive features to serve, and until there’s a published rate they stay specs in a preview the allowlist sees and no one else can budget against.