Axiom's open-source repo says AI agents turned a 15,000-page proof into about 43 million Lean lines in 24 days. Lean checks each step, letting builders verify the proof as code. Unconfirmed via testingcatalog.

CIIL released Anuvani-Diarization, an open-source diarization model for 22 Indian languages plus English. It processes six hours of audio in 55 seconds on one B200, about 700x real-time. Weights are CC BY-NC 4.0; inference code is MIT or Apache-2.0. huggingface.co/CIIL/Anuvani-…

One author's Review-it run flagged 1 ghost citation among 70 references: 68 verified, zero self-citations, LOW RISK badge in sampled frames. One demo only, not proof of general accuracy, but useful as a pre-submission citation check.

Claude Duet is rolling out to US Max users first. US Pro users come next. International access has no timeline yet. Based on Karan Ahuja's post, not an official announcement.

A single prompt to Gemini 4 Argon High used 2% of one PRO user's 5-hour usage limit. That screenshot is one user's observation, not a confirmed Google-wide rate. Builders on PRO should expect even trivial Argon High requests to dent daily quotas.

Artificial Analysis Index v5: Claude Opus 5.5 leads at 58; released open-weight models score 32, a 26-point gap. Kimi K3.1 Pro leads open weights at 46 but its weights aren't out yet.

Morph is offering students 12 months of free API access to DeepSeek V4.1 Flash, GLM-5.3 and Kimi K3, verified via SheerID with no credit card. A community post notes the signup page was returning 404 errors for some users, so try Morph docs if it fails.

Grok Bot shows per-app Allow buttons for Calendar, Reminders and Contacts on iOS and Android in a permissions mockup, not proof the access is live yet. Users can decide which phone data to grant if the rollout reaches them.

Judgment Labs opens general access for agent trace triage. Agent swarms can review production traces, spot failures, and test fixes. The Judgment MCP works with Claude, Codex, Cursor, or any MCP client. GA status comes from an X post screenshot, not independent confirmation.

An unconfirmed report flags CVE-2026-116692, a critical MLflow scorer flaw letting attackers run server code, affecting 3.12.0rc0 up to but not including 3.17.0. Upgrade to 3.17.0.

Claude Pro's Opus 5.5 reportedly reaches 700–900 million weekly tokens at ~90 tok/s, versus ~220 million for GPT-6.1 Sol on GPT Plus. A single X post cites an anonymous source; providers haven't confirmed these figures.

ChatGPT may connect to iCloud Mail and Calendar on iOS, web and desktop. Plugins spotted in the store let it read and manage email and events. Unconfirmed third-party sighting, not an OpenAI announcement.

A Gemini Tester screenshot shows new sidebar menus: Images, Scheduled, Students, plus Notebooks and '+ New notebook'. It also links to 'Official Discord API Documentation'. This is an observed image, not a Google announcement or confirmed rollout.

GPT-6.1 Sol fell to 3.9 on Agent Arena's shell-command reliability, down from GPT-6 Sol's 11.7 and GPT-5.6 Sol's 12.0; Claude Opus 5.5 is a reference line at 15.8. One observed chart shows this, not independently verified results.

Sentient OS hands an open screen task—like a form or email—to its own cursor, storing what it learns in local plain-text files. A single launch post offers the first 1,000 users lifetime free access with code FINAL_1000; it has not been independently tested.

Genie 4 is a steerable world model: change weather, camera and soundscape with natural language, powered by Gemini as the simulation brain. It arrives on the API for the first time soon. No date or pricing yet.

MiniCPM-OCR 1.0 is a 1.3B open-source OCR model claiming 93.68 on Real5-OmniDocBench and 1.49x PaddleOCR-VL-1.6 throughput. These are one X post's promotional-card claims, not verified benchmarks.

Packet capture tests flag a Gemini 4 Argon cache anomaly that spikes single-request costs near 10x at 0% cache hit. Quotas stay $15/week on Pro and $200/week on Ultra, but silent spikes can drain them faster. Third-party packet capture report, not confirmed by Google.

A user screenshot shows Claude Max at $124.99/month for 5x usage and $249.99/month for 20x usage, up from $100 and $200. Anthropic hasn't confirmed the change. Regional pricing may vary.

A single X post claims Genie 4 is here for interactive worlds — with excitement but no official details, specs, or release date to back it yet.