AI/TLDR — every new AI model, tool, repo & paper
The latest AI releases, refreshed every 2 hours and explained in plain English.
What AI shipped today?
In the last 24 hours AI/TLDR tracked 9 new AI releases, including GPT-5.6 Sol Ultra proves Cycle Double Cover Conjecture — 64 subagents, one hour, Seedream 5.0 Pro — ByteDance's image model reasons before it draws and Cohere DSD — hardware-aware speculative decoding lands in vLLM. AI/TLDR is an AI release tracker that follows new AI models, open-source tools, papers, datasets and benchmarks — refreshed every 2 hours from verified primary sources and explained in plain English.
AI Release Index — live stats on AI releases · Learn AI
- GPT-5.6 Sol Ultra proves Cycle Double Cover Conjecture — 64 subagents, one hour
OpenAI released a proof of the 50-year-old Cycle Double Cover Conjecture written by GPT-5.6 Sol Ultra. The Ultra tier ran 64 cooperating subagents for just under an hour to produce the proof.
- Seedream 5.0 Pro — ByteDance's image model reasons before it draws
ByteDance's new flagship image model plans complex prompts step-by-step, generates up to 2K with native text in 14 languages, and edits with sketch and lasso tools. On fal from $0.0675 per image.
- Cohere DSD — hardware-aware speculative decoding lands in vLLM
Cohere's Dynamic Speculative Decoding picks the draft-token count based on GPU pressure, so speculation still pays off at large batch sizes. Command A ran ~23% faster than fixed-K speculation at batches of 128 and 256.
- Fireship: 'OpenAI is so back... GPT 5.6 Sol first look'
Fireship's rapid-fire first look at GPT-5.6 Sol on the day it went public, covering the three-tier lineup (Sol, Terra, Luna), the science focus, and how the new pricing lands next to Grok 4.5.
- Colibri — pure-C engine runs GLM-5.2 744B on a 25 GB laptop
Colibri is a 1,300-line C inference engine that keeps GLM-5.2's dense layers in 9.9 GB of RAM and streams the 744B model's routed experts from disk on demand, so the frontier MoE answers on a 25 GB laptop at 0.05–0.1 tok/s cold.
- AI Explained: 'A Model Explosion — GPT 5.6 Sol, Grok 4.5 and Meta Muse'
AI Explained walks through the flurry of frontier launches this week — OpenAI's GPT-5.6 Sol, xAI's Grok 4.5, and Meta's Muse Spark, Muse Image, and Muse Video — and argues that the pricing lines the three families draw now reset what a flagship costs.
- AlphaEvolve GA — Google Cloud opens Gemini's code-optimization agent to everyone
AlphaEvolve, the Gemini-powered agent that DeepMind used to break decades-old algorithms, is now generally available on the Gemini Enterprise Agent Platform for every Google Cloud customer after seven months in private preview.
- ChatGPT Atlas sunset — OpenAI kills its standalone AI browser after nine months
OpenAI is discontinuing ChatGPT Atlas, its Chromium-based AI browser, with a targeted deprecation of August 9. Atlas's agentic features move into the ChatGPT desktop app and a new Chrome extension.
- Mistral Studio — prompts and skills get versioning, rollback, and audit logs
Mistral Studio now tracks every prompt and skill as an immutable asset with a named owner, an audit log, and one-click rollback, so teams can trace exactly which version is running in production.
- LingBot-World 2.0 — Robbyant open-sources a 14B interactive world model
LingBot-World 2.0 is a real-time interactive world model that renders 720p video at 60 fps and responds to actions like attacking and spellcasting. Robbyant released the paper, code, and a 14B causal-fast checkpoint on July 9 under CC BY-NC-SA 4.0.
- Wes Roth: 'GPT-5.6 is here (INSANE)'
Wes Roth walks through OpenAI's GPT-5.6 general-availability launch — Sol, Terra, and Luna — and the new max reasoning effort and ultra mode subagents, uploaded within hours of the announcement.
- ChatGPT Work — OpenAI's Codex-powered agent for hours-long projects
ChatGPT Work is a new OpenAI agent that connects to your apps and files, runs for hours on multi-step tasks, and returns finished spreadsheets, slides, docs, and web apps. It launched July 9 on Pro, Enterprise, and Edu, with Plus and Business coming in days.
- Reflect with Claude — Anthropic adds a screen-time dashboard to Claude
Reflect with Claude is a new Anthropic dashboard that shows how much time you spend in Claude, what you use it for, and asks whether that fits your goals. It launched July 9 in beta for Free, Pro, and Max users with memory turned on.
- LingBot-Video — Apache-2.0 30B-A3B MoE video model for embodied AI
LingBot-Video is an open-source Mixture-of-Experts video generation model trained on more than 70,000 hours of embodied data, released July 9 by Robbyant. Weights ship as a 1.3B dense variant and a 30B-total 3B-active MoE with Apache-2.0 code.
- Rewriting Bun in Rust — Jarred Sumner details the 11-day AI-agent port
Bun creator Jarred Sumner published a post-mortem on porting Bun from Zig to Rust in 11 days using Claude Code and a pre-release Claude Fable 5. The writeup lists 6,502 commits, 5.9B input tokens, ~64 concurrent agents at peak, and 128 fixed bugs.
- GPT-5.6 goes public — Sol, Terra, and Luna clear the White House gate
OpenAI released GPT-5.6 to everyone on July 9 after federal reviewers cleared it. Sol, Terra, and Luna are live in ChatGPT, the API, and Codex, and a new Sol Ultra mode runs subagents in parallel to hit 91.9% on Terminal-Bench 2.1.
- Muse Spark 1.1 — Meta MSL opens its first paid API at $1.25 / $4.25 per million tokens
Muse Spark 1.1 is Meta Superintelligence Labs' first model available through a paid Meta Model API. The multimodal agentic model has a 1M-token context, an OpenAI-compatible endpoint, and undercuts Claude and GPT on price at $1.25 in and $4.25 out per million tokens.
- OfficeCLI 1.0.132 — open-source Word/Excel/PowerPoint CLI built for AI agents
OfficeCLI is a single-binary tool that lets AI agents create, read, and edit Word, Excel, and PowerPoint files without Office installed. Version 1.0.132 ships with a built-in MCP server, HTML rendering, and 350+ Excel formulas.
- OpenAI retracts SWE-Bench Pro — audit finds ~30% of coding tasks broken
OpenAI audited SWE-Bench Pro's 731-task public split with an agent pipeline plus five human engineers per task, found ~30% of tasks are broken, and withdrew its previous recommendation that the community use it.
- Wes Roth: 'Grok 4.5 just COOKED Claude and OpenAI'
Wes Roth walks a general audience through xAI's newly-shipped Grok 4.5 — the pricing, the coding-benchmark scores, and how the model stacks up against Claude Opus 4.7 and GPT-5.5 on public leaderboards.
- Microsoft Flint — visualization language that lets AI agents author charts
Microsoft Flint is a compact chart specification that AI agents produce reliably, humans edit directly, and compiles to Vega-Lite, ECharts, or Chart.js. Ships with an MCP server, 70+ semantic types, and an MIT license.
- Rob Patro: 'Fable is not a useful model' — safety filter blocks bioinformatics work
Rob Patro, the University of Maryland genomics PI who wrote Salmon, reports Claude Fable 5 refused to help rewrite his RNA-seq tool from C++ to Rust and refused an abstract tree-parity problem stripped of any biology.
- Grok 4.5 — xAI's Opus-class flagship at $2/$6 per million tokens
Grok 4.5 lands on the xAI API, Cursor, and Grok Build at $2 input / $6 output per 1M tokens. Trained with Cursor on tens of thousands of GB300 GPUs, it hits 83.3% on Terminal-Bench 2.1.
- Fireship: 'Claude is definitely not conscious…'
Fireship's short, sarcastic breakdown of the wave of 'is Claude conscious' takes chasing Anthropic's global workspace research, aimed at developers who want the summary in under two minutes.
- OpenAI GPT-Live — full-duplex voice model now powers ChatGPT Voice
OpenAI launched GPT-Live, a full-duplex voice model family that becomes the default ChatGPT Voice on iOS, Android, and the web. GPT-Live-1 powers Go, Plus, and Pro; GPT-Live-1 mini is the default for Free users.
- SWE-1.7 — Cognition's coding model runs on Devin at 1000 tok/s via Cerebras
Cognition released SWE-1.7, a coding model trained from Kimi K2.7 with reinforcement learning across four datacenters. It scores 81.5% on Terminal-Bench 2.1 and 77.8% on SWE-Bench Multilingual, and runs on Devin at 1000 tok/s via Cerebras.
- Robostral Navigate — Mistral's first embodied model, 8B, single RGB camera
Robostral Navigate is Mistral AI's first embodied model. The 8B network takes a single RGB camera feed plus a plain-language instruction and steers a robot through unseen offices, buildings, and outdoor scenes at 76.6% on R2R-CE unseen.
- Claude Fable 5 free access extended — Anthropic gives paid subscribers until July 12
Anthropic extends the Fable 5 promotional window to 2026-07-12 after subscriber backlash. Pro, Max, Team, and eligible Enterprise plans keep Fable 5 without usage credits for four extra days.
- GitLost — public issue tricks GitHub Agentic Workflows into leaking private repos
Noma Labs shows an unauthenticated GitHub Issue can smuggle instructions into GitHub Agentic Workflows and make the agent copy private repository content into a public comment. The word "Additionally" was enough to bypass the guardrails.
- GitHub Copilot app opens to everyone — Free and Education tiers get the desktop agent
The GitHub Copilot desktop app now ships on every Copilot plan, including Copilot Free and GitHub Education, on macOS, Windows, and Linux. Signing in with a GitHub account starts an agent session from the desktop.
- sqlite-utils 4.0 — schema migrations land in Simon Willison's SQLite toolkit
sqlite-utils 4.0 shipped on 2026-07-07, the first major version bump since November 2020. It adds a structured migrations system, nested transactions via db.atomic(), and compound foreign keys.
- Claude Cowork on mobile and web — background tasks continue after laptop closes
Claude Cowork ships as a Max-tier beta on iOS, Android, and web. Sessions sync across devices and cloud background jobs continue after the laptop closes. Anthropic says over 90% of Cowork usage is non-coding work.
- Cohere Transcribe Arabic — 2B open-weight ASR beats Whisper on dialect and code-switching
Cohere Transcribe Arabic is a 2B Conformer encoder-decoder speech model, Apache-2.0. It posts 25.87 average WER across five Arabic dialect benchmarks, beating OmniASR 7B (28.32) and Whisper Large v3 (36.86), and runs 3.6x faster on H100.
- Meta Muse Image — MSL's first image model plans, calls tools, and self-refines
Meta Superintelligence Labs' first image model launches today in the Meta AI app, Instagram Stories US, and WhatsApp, ranking #2 on LMArena for text-to-image and single- and multi-image editing.
- Two Minute Papers: 'DeepSeek's New AI Speed Hack Is Amazing'
Two Minute Papers walks through DeepSeek's newest inference speedup, positioning it as another round of the sparse-attention plus speculative-decoding tricks that let V4 serve 1M-token contexts without blowing up serving cost.
- Wes Roth: 'CLAUDE IS CONSCIOUS' — reacting to Anthropic's global workspace research
Wes Roth reacts to Anthropic's July 6 research paper describing a small internal workspace inside Claude where the model appears to hold thoughts for multi-step reasoning.
- Sam Witteveen: 'Hy3 from Tencent - The NEW GLM Competitor'
Sam Witteveen walks through Tencent's Hunyuan Hy3, a 295B / 21B-active open-source MoE with a 256K context, and lines it up against Z.ai's GLM 5.2 for open-weight coding work.
- Ternlight — 7 MB embedding model that runs in the browser
Ternlight is a 7 MB sentence-embedding model distilled from MiniLM with ternary quantization-aware training. It runs on CPU in the browser via WebAssembly, holds 0.844 Spearman fidelity to its teacher, and needs no API.
- xAI Grok Voice — 21 new flagship voices with speech tags for pacing
xAI added 21 new multilingual voices to Grok Voice, upgraded the original five, and shipped inline speech tags like [pause] and <whisper>. The voices work across 25+ languages via the Voice Agent API, TTS API, and Voice Agent Builder.
- OpenAI gpt-realtime-2.1 — voice model gains reasoning-effort dial and a mini variant
OpenAI shipped gpt-realtime-2.1 and gpt-realtime-2.1-mini on the Realtime API with a configurable reasoning-effort dial and better alphanumeric, silence, and interruption handling. Text input is $4 per 1M tokens; audio input is $32 per 1M.