AI Developer Digest
0 main-section items passed quality gate | ~32 scanned | ~29 excluded | Sources checked: 26 Scan window: June 15 (post-prior-digest) – June 16, 2026 (24h). This is a genuinely light day: nothing new from tier 1 labs, GitHub release repos, arXiv, leaderboards, or infra blogs cleared the quality bar within the window. The most actionable item in this digest is a deadline, not a release: the Gemini CLI hard stop is now 2 days away.
This Week's Signal
After yesterday's double Anthropic breaking change (Sonnet 4 / Opus 4 retirement + Agent SDK billing split), today is quiet by comparison — no lab shipped a new model, no GitHub release repo had a feature-level change, and no leaderboard moved. The only repo activity was routine maintenance: llama.cpp pushed twelve backend-maintenance builds (SYCL/Vulkan op coverage, an NVFP4 edge-case fix, Eagle3 speculative-decoding sampling support), LiteLLM shipped a stable-branch backport (v1.89.1), and Transformers shipped a two-line patch (v5.12.1). None of these clear the bar for a full entry. The thing that actually demands developer attention today is a previously-announced deadline closing in: the
geminiCLI hard stop is now 2 days out (June 18), with no grace period for Pro/Ultra/free-tier users.
Must-reads this digest:
- Gemini CLI hard stop — 2 days (June 18) — if any CI/CD pipeline, GitHub Action, or script still calls the
geminicommand on a non-org account, it stops working Thursday with no grace period
[BREAKING] Breaking Changes
No new breaking changes in the June 15–16 window. (Yesterday's Claude Sonnet 4 / Opus 4 retirement and Agent SDK billing split are now in effect — see the June 15 digest — but nothing new triggered today.)
Model Releases
No new model releases in the June 15–16 scan window. Gemini 3.5 Pro remains in limited Vertex enterprise preview; Kimi K2.7 Code (June 12) and Grok V9-Medium (June 10) remain outside this window with no new developments today. See Worth Watching for status.
API & SDK Changes
No API or SDK changes met the quality bar within the window. Transformers v5.12.1 (June 15) is a two-line dependency/tokenizer patch — see Quick Hits.
Research
Nothing cleared the quality bar this period. Direct arXiv cs.AI/cs.CL search did not surface a June 15–16 submission from a recognized lab with benchmark numbers or a linked repo; results returned were either off-window or lacked confirmable code/benchmarks. See Horizon for a note on this gap.
Tooling
No tooling release this period reached feature-level significance — see Quick Hits for the three patch/maintenance releases that did ship (llama.cpp builds, LiteLLM v1.89.1, Transformers v5.12.1).
Benchmarks & Leaderboards
No new leaderboard movements or SOTA changes in the June 15–16 window. Claude Mythos 5 continues to lead SWE-bench Verified at 95.5% (unchanged since June 13), followed by Claude Fable 5 (95%) and Claude Opus 4.8 (88.6%). Third-party Kimi K2.7 Code evaluations remain pending.
Trends & Emerging Tech
Nothing cleared even the lower Trends bar (score ≥2) this period. Simon Willison published "The Fable 5 Export Controls Harm US Cyber Defense" on June 16, which would otherwise be a candidate, but simonwillison.net returned a 403 on fetch — per the non-negotiable rule against citing anything not actually read, it's excluded. See near-misses.
Technical Discussions
Nothing cleared the quality bar this period.
Quick Hits
- llama.cpp b9660–b9672 (June 15 22:05 UTC – June 16 18:54 UTC) — 12 builds: NVFP4 edge-case fix in llama-graph, Eagle3 speculative-decoding backend sampling support, Vulkan
col2im_1dop and gated-delta-net support, SYCL EXPM1/floor/trunc/round op support, LFM2 tool-call double-escaping fix, BoringSSL vendor bump. Backend maintenance only, no new model support or published benchmarks. [github.com/ggml-org/llama.cpp/releases] - LiteLLM v1.89.1 (June 16, 03:31 UTC) — Stable-branch backport: "1.84.8 patch set + MCP/model-info/DB fixes to stable/1.89.x." No new features. [github.com/BerriAI/litellm/releases/tag/v1.89.1]
- Transformers v5.12.1 (June 15, 2026) — Two-line patch: corrected PEFT minimum version bound (#46605) and fixed auto-tokenizer resolution for the Mistral tokenizer when
mistral-commonis installed (#46667). [github.com/huggingface/transformers/releases/tag/v5.12.1]
Worth Watching (Announced, Not Yet Shipped)
⚠️⚠️⚠️ Gemini CLI Hard Stop — June 18 (2 DAYS — URGENT)
(Countdown updated)
Source: Google Developers Blog | Link: https://developers.googleblog.com/an-important-update-transitioning-gemini-cli-to-antigravity-cli/
The gemini CLI and Gemini Code Assist IDE extensions stop serving requests June 18 for Google AI Pro, Google AI Ultra, and free-tier (Code Assist for individuals) users. Hard stop — no grace period. Replacement is agy (Antigravity CLI). No 1:1 feature parity at launch; weekly compute-based cap replaces the 1,000 req/day limit, with multi-day cooldowns reported when exhausted. Google Cloud org accounts on Standard/Enterprise license are not affected. Audit CI/CD pipelines, GitHub Actions workflows, and scripts calling gemini before Thursday.
⚠️⚠️⚠️ Gemini API Unrestricted Key Deadline — June 19 (3 days)
(Countdown updated) Source: Google AI for Developers | Link: https://ai.google.dev/gemini-api/docs/api-key All unrestricted Gemini API keys blocked June 19. Restrict via AI Studio → API Keys → "Restrict to Gemini API." ~2 minutes; no code changes required.
⚠️⚠️ Gemini Image Models Shutdown — June 25 (9 days)
(Countdown updated)
Source: Google AI for Developers | Link: https://ai.google.dev/gemini-api/docs/deprecations
gemini-3.1-flash-image-preview and gemini-3-pro-image-preview shutting down June 25. Migrate to stable image model equivalents.
⚠️⚠️ GPT-4.5 Retirement from ChatGPT — June 27 (11 days)
(Countdown updated)
Source: OpenAI Platform Changelog | Link: https://platform.openai.com/docs/changelog
GPT-4.5 removed from the ChatGPT product surface June 27. API route retirement unconfirmed. Audit any gpt-4.5 model identifiers.
⚠️⚠️ Kimi K2.7 Code Third-Party Benchmarks — Expected ~June 22 (6 days)
(Carried) Kimi K2.7 Code weights landed June 12. Third-party SWE-bench Verified and LiveCodeBench evaluations typically appear 7–14 days post-weight release. Watch paperswithcode.com and swebench.com around June 20–25.
⚠️⚠️ Grok V9-Medium — API Release Still Pending
(Status unchanged) xAI deployed Grok V9-Medium to Tesla fleet and X users as of June 10 (1.5T parameters, 32B active). No API model ID, no pricing, no confirmed public benchmark numbers as of June 16.
⚠️ Claude Fable 5 / Mythos 5 Reinstatement — No Timeline Announced
(Carried)
Source: Anthropic | Link: https://www.anthropic.com/news/fable-mythos-access
Both models remain suspended under the US export-control directive issued June 12. No return date. Migrate to claude-opus-4-8 for agentic workloads.
⚠️ Gemini 3.5 Pro — GA Still Pending (Limited Vertex Enterprise Preview)
(Carried — status unchanged) Source: Google AI for Developers | Link: https://ai.google.dev/gemini-api/docs/models Expected: 2M token context, Deep Think reasoning mode. No general availability date.
⚠️ Aion 1.0 Open Weights — July 2026 (~2 weeks)
(Countdown updated) Source: Windows Developer Blog | Link: https://blogs.windows.com/windowsdeveloper/2026/06/02/build-2026-furthering-windows-as-the-trusted-platform-for-development/ Microsoft Aion 1.0 Instruct open weights on Hugging Face in July 2026. No confirmed specific date.
⚠️ Claude Opus 4.1 Retirement — August 5 (50 days)
(Countdown updated)
Source: Anthropic | Link: https://platform.claude.com/docs/en/about-claude/model-deprecations
claude-opus-4-1-20250805 retires August 5. Migrate to claude-opus-4-8.
Apple iOS 27 / macOS Golden Gate / Core AI GA — Fall 2026 (September)
(Carried — status unchanged) Source: Apple Developer / WWDC 2026 | Link: https://developer.apple.com/ios/ Includes Siri Extensions API, Core AI (replaces Core ML), Foundation Models multi-provider support.
Claude Mythos 5 General Availability — No Timeline
(Carried — suspended under same export-control order) Source: Anthropic | Link: https://www.anthropic.com/news/expanding-project-glasswing
⚠️ OpenAI Reusable Prompts / Evals Platform / Agent Builder Shutdown — November 30 (167 days)
(Carried)
Source: OpenAI | Link: https://platform.openai.com/docs/deprecations
Export eval configs before October 31 (read-only from that date). Migrate Agent Builder to Agents SDK. Move prompt content from v1/prompts to application code.
<details> <summary>🔭 Horizon — Open Questions, Emerging Patterns & Grounded Speculation</summary>
This section operates under different rules than the digest above. Evidence-grounded speculation is allowed. Pure prediction is not. Every claim here must cite a source from this digest or a real paper/benchmark. Label each entry by type so the reader knows what kind of thinking they're engaging with.
[PATTERN] Disruptive-event days are followed by maintenance-only days, not by a steady drip Yesterday's digest carried two simultaneous Anthropic breaking changes plus a notable LiteLLM feature release. Today, the same set of tier-1 sources and GitHub repos produced nothing above patch-level maintenance (llama.cpp backend ops, a LiteLLM stable-branch backport, a two-line Transformers fix). Across the last several digests, the news cadence looks bursty rather than continuous — concentrated around specific announcement days (model retirements, billing changes, export-control actions) with quiet stretches between. The practical implication for anyone consuming this digest: a quiet day is not a signal that nothing is coming, just that nothing landed in this specific 24h window. Grounded in: June 15 digest (two [BREAKING] entries + LiteLLM v1.89.0); this digest's Quick Hits (zero feature-level changes across llama.cpp, LiteLLM, Transformers)
[OPEN QUESTION] How many automated pipelines will silently break on June 18 because the audit never happened?
The Gemini CLI hard stop has been carried in this digest series for several days with shrinking countdowns (3 days → 2 days), and the source is explicit that there's no grace period for the affected account tiers. Unlike a deprecation with a soft warning period baked into the API (e.g., Anthropic's retired-model error responses, which at least fail loudly), a CLI binary that stops authenticating mid-pipeline can fail in ways that are harder to immediately attribute to "the Gemini CLI deadline" — a CI job just starts failing. The open question: of the teams running gemini in scripts today, how many have actually completed the audit this digest has recommended for days, versus how many will discover the cutoff when a pipeline breaks on Thursday morning?
Grounded in: Gemini CLI Hard Stop entry (this digest, Worth Watching); Google Developers Blog hard-stop announcement (carried from prior digests)
[IF THIS CONTINUES] llama.cpp's Eagle3 backend sampling support is a small signal for where speculative decoding is heading
Build b9669 (June 16) adds "backend sampling support for eagle3" — Eagle3 is a speculative-decoding technique that predicts multiple draft tokens per step. Backend-level sampling support (as opposed to a reference Python implementation) is what's needed to make a draft technique practical to actually deploy on a llama.cpp-based inference stack. This is a small, unbenchmarked change today, but if llama.cpp continues adding backend support for newer speculative-decoding methods at this pace, draft-model speedups that are currently confined to research repos and big-lab serving stacks (vLLM, TensorRT-LLM) become available to anyone running local inference. Worth tracking whether a follow-up build adds a --draft-model flag or benchmark numbers for Eagle3 specifically.
Grounded in: llama.cpp b9669 "spec: add backend sampling support for eagle3" (this digest, Quick Hits)
[RESEARCH THREAD] Nothing new surfaced from arXiv this cycle — but the gap itself is the data point For the second consecutive digest, direct enumeration of arXiv cs.AI/cs.CL new submissions failed to surface a single paper meeting the quality bar (recognized lab or GitHub repo, plus concrete benchmark numbers) within the scan window. Today's search returned only off-window or tangential results (a "Benchmark of Benchmarks" survey, a curiosity-driven test-generation paper with no June 2026 date). This may reflect genuinely thin output, or it may reflect that arXiv listing pages are hard to enumerate via web search alone (the June 15 digest noted direct 403s on the listing pages themselves). Worth revisiting with a more targeted search strategy (specific subfields, specific lab names) in the next digest rather than concluding research output has actually slowed. Grounded in: this digest's Research section (nothing cleared the bar); June 15 digest's arXiv 403 note
</details>Excluded: ~29 items below quality gate threshold, outside scan window, or already covered in prior digests. Near-misses: Simon Willison "The Fable 5 Export Controls Harm US Cyber Defense" (June 16 — fetch returned 403, content unverifiable, otherwise a plausible Trends candidate); Hugging Face "JFrog Artifactory: An Enterprise Guide (and What Changes in June 2026)" by Jeff Boudier (fetch returned 403, date and content unverifiable); Cohere "The future of work debate has an evidence problem" (June 15 — research/policy framing, no technical/developer substance); LangChain 1.0 alpha (announced early June, outside 24h window, already a known trajectory item); Ollama v0.30.8 (June 12 — outside window, broadens GGUF hardware support via llama.cpp, carried forward consideration only); OpenAI Python SDK v2.41.1 (June 10 — outside window); Unsloth unsloth-zoo release (June 12 — outside window); Hacker News (no Show HN or technical thread with score >200 identified for June 16); LMArena/SWE-bench (no leaderboard movement within window — Claude Mythos 5 still leads SWE-bench Verified at 95.5%, unchanged since June 13); AWS ML blog, Azure AI blog, Groq blog, Together AI blog (no posts confirmed within window).