AI CompassThe day’s major AI releases, written up and sourced.
Weekly · 2 to 6 November 2026 · 63 launches across 5 days · the whole week in 10 minutes
Weekly · 2 to 6 NovCreativeOpenAI
Moonshot AI released open weights for its 2.8T-parameter Kimi K3.1 Pro, Thinking Machines Lab debuted its 17B-active Inkling 2 Small MoE, Google Antigravity brought cloud VM Remote Agents to consumer plans and OpenAI unveiled GPT-6 Cyber across 63 confirmed launches from 2 to 6 November 2026.
Four product and engineering shifts defined the 63 launches published between 2 and 6 November 2026: image and video generation moved from single-shot prompts to multi-turn editing, consumer assistants gained subsecond avatars and persistent mini-apps, open-weight labs scaled sparse models to 2.8 trillion parameters and coding agents expanded into cloud virtual machines and autonomous security pipelines.
In multimodal media and everyday apps, OpenAI’s ChatGPT Images 2.6 cut ChatGPT rendering times by 50% while adding vector SVG export, looping GIFs and markup-guided edits across two new API endpoints, and Google’s Gemini Omni 1.2 Flash lowered 720p video pricing by 30% to $0.07 per second alongside multi-turn 15-second shot refinements and 4K upscaling. Across 5 and 6 November, Meta paired 870 ms, 25 fps Muse Realtime Voice and Avatars with chat-built Muse Spaces and Active Memory models.
For builders running open weights and background agents, Moonshot AI’s Kimi K3.1 Pro shipped native MXFP4 weights across 2.8 trillion parameters and a 1M-token window, Thinking Machines Lab’s Inkling 2 Small reached 86.3% on SWEBench Pro v2 Hard with 17B active parameters and Ricursive Intelligence’s Ripple-65B Preview fit frontier coding accuracy onto a single workstation under Apache 2.0. Meanwhile, Google Antigravity Remote Agents moved consumer repository runs into cloud VMs, OpenRouter Auto Code cut coding-agent token bills by 74% and OpenAI’s GPT-6 Cyber automated root-exploit validation across 948 Linux kernel flaws.
01CreativeOpenAIConfirmed
OpenAI rolled out ChatGPT Images 2.6 globally on 4 November across ChatGPT and Codex alongside developer endpoints gpt-image-2.6-flare and gpt-image-2.6-sunburst ($8 input and $30 output per 1M image tokens), adding looping GIF generation, reusable character references and denser multi-column typography.
Paired across 2 to 6 November with Google’s Gemini Omni 1.2 Flash on multi-turn video and Pruna AI’s 8-step P-H3-Image for 24 GB local workstations, Images 2.6 gives product teams 6-second low-quality Flare drafts, tunable layout-planning budgets and clean vector SVGs.
Read full OpenAI breakdownAll OpenAI launchesCreativeIntroducing Images 2.6 (openai.com)
02ModelGoogleConfirmed
Google made Gemini Omni 1.2 Flash generally available across the Gemini API, AI Studio, Vertex AI and Gemini Ultra on 5 November, extending native single-pass clips from 10 to 15 seconds with 360p or 720p drafting and post-generation 1080p or 4K upscaling.
Conditioned on character images, 3-second motion clips and 1-second voice cues for multi-speaker lip-sync, its conversational shot editing complements OpenAI’s ChatGPT Images 2.6 while raising the bar for detectors such as Pangram’s Image Flash 1.0.
Read full Google breakdownAll Google launchesModelGemini Omni 1.2 Flash: Fast and cost-effective with longer clips (blog.google)
03ProductMetaConfirmed
Meta shipped a two-day overhaul of its Muse consumer app on 5 and 6 November, starting with Realtime Voice and a two-step causal Diffusion Transformer streaming 448x768 avatars at 25 fps and 870 ms latency that beat Runway (78% to 22%) and HeyGen (88% to 12%) in blind live trials.
A day later Meta added Muse Spaces for pinning chat-built trackers and mini-apps to navigation alongside specialised Active Memory models distilled from Muse Wonder that cut critical context forgetting by 2.4 times.
Read full Meta breakdownAll Meta launchesProductGive Your Muse a Voice (about.fb.com)
04Open sourceMoonshot AIConfirmed
Moonshot AI published Kimi K3.1 Pro open weights on Hugging Face on 4 November alongside a 2 to 3 times faster Highspeed API tier, activating 104B of its 2.8 trillion parameters per token across a 1,048,576-token window using Kimi Delta Attention and Gated DeepSeek Sparse Attention.
Leading Claude Opus 5.5 and GPT-6 Astra on SWE Marathon v1.1 (51.7 versus 46.3 and 42.5) and scoring 92.7 on BenchCAD, it anchors early November’s open-weight surge alongside single-workstation Ripple-65B Preview and Meituan’s 1.6T LongCat-2.5.
Read full Moonshot AI breakdownAll Moonshot AI launchesOpen sourceKimi K3.1 Pro technical blog (kimi.ai)
05ModelThinking Machines LabConfirmed
Introduced on 5 November on Tinker Playground and the Tinker fine-tuning API ahead of its planned mid-November Apache 2.0 weight release, Inkling 2 Small pairs 594B total parameters with 17B active per token and configurable visual reasoning.
Following Cognition’s SWE-2 Flash post-trained on first-generation Inkling-Small, Inkling 2 Small reached 31 on the AA Intelligence Index and delivered a 6.02 times speedup on DeepSeek-V3.2 sparse attention kernels.
Read full Thinking Machines Lab breakdownAll Thinking Machines Lab launchesModelIntroducing Inkling 2 Small (thinkingmachines.ai)
06Coding agentGoogleConfirmed
Google expanded Remote Agents across the Antigravity desktop app, CLI and web interface on 6 November for Google AI Pro and Ultra personal accounts, running GitHub repository workflows and pull requests inside cloud virtual machines while local laptops sleep.
Together with DeepSeek Harness Creator Mode for live local UI plugins, Mistral Studio Managed Agents and OpenRouter Auto Code, the 2 to 6 November launches split coding harnesses between extensible local shells and unattended cloud sandboxes.
Read full Google breakdownAll Google launchesCoding agentRemote Agents documentation in Google Antigravity (antigravity.google)
07ModelOpenAIConfirmed
Unveiled on 5 November for vetted Daybreak Red defenders via the API and Codex, GPT-6 Cyber is the first OpenAI model rated Critical in cybersecurity, scoring 48.5% on ExploitGym and completing 95% of RemoteExploitBench tasks that mainline GPT-6 Astra refuses.
It directly powered OpenAI’s Linux kernel audit that built 948 working root exploits at $1,279 per distinct flaw and uncovered 81 high-severity Firefox bugs, arriving alongside the Frontier Model Forum’s cyber jailbreak rubric.
Read full OpenAI breakdownAll OpenAI launchesModelGPT-6 Cyber: Advancing the frontier of aligned cybersecurity agents (openai.com)
| Release | Company | Category | Key spec / benchmark / price | Story |
|---|---|---|---|---|
| ChatGPT Images 2.6 | OpenAI | Creative | 50% faster in ChatGPT, SVG and GIF output, $8 in / $30 out per 1M image tokens | Breakdown |
| Gemini Omni 1.2 Flash | Model | 3 to 15s clips, multi-turn edits, 4K upscale, $0.07/s at 720p (-30%) | Breakdown | |
| Muse Realtime Avatar | Meta | Product | 2-step causal DiT, 448x768 at 25 fps, 870 ms latency, 88% win rate vs HeyGen | Breakdown |
| Kimi K3.1 Pro | Moonshot AI | Open source | 2.8T / 104B active MoE, 1M context, MXFP4 open weights, 51.7 SWE Marathon v1.1 | Breakdown |
| Inkling 2 Small | Thinking Machines Lab | Model | 594B / 17B active MoE, 86.3% SWEBench Pro v2 Hard, 31 AA Intelligence Index | Breakdown |
| Ripple-65B Preview | Ricursive Intelligence | Open source | 65B dense Apache 2.0, 1M context, 80.52 TerminalBench 2.1 ($32M R&D) | Breakdown |
| LongCat-2.5 | Meituan | Model | 1.6T / 48B active MoE, discrete visual tokens, 88.4 Terminal Bench v2.1 | Breakdown |
| Tencent Hy4 | Tencent | Model | Multimodal CLI, GUI and 3D engine control, roughly $0.90 in / $2.70 out per 1M | Breakdown |
| OpenRouter Auto Code | OpenRouter | Infra | 74.3% Deep SWE at $1.98/task (74% lower cost), 0.65s p90 routing latency | Breakdown |
| GPT-6 Cyber | OpenAI | Model | 48.5% ExploitGym, 95% RemoteExploitBench, 948 Linux kernel LPEs (Daybreak Red) | Breakdown |
63 launches from 2 to 6 November 2026, distilled and edited by .
Spotted an error? Email aiforsuccess.in+compass@gmail.com or tag @ai_for_success on X
Monday 2 November
Tuesday 3 November
Wednesday 4 November
Thursday 5 November
Friday 6 November
