# forck.live — 20–26 April 2026

> AI lab announcements, news reports and podcasts.

forck.live follows what companies announce about their own products. Press may
corroborate an official story; it never leads and never appears as an item of its
own. A missing date stays missing. A title is the publisher's words — where a
company does not publish in English, an English rendering is listed separately
as `Title in English` and named as this site's translation, never as theirs.

This file is one issue of the weekly archive: every confirmed
first-party announcement of ISO week 2026-W17, 20–26 April 2026, newest
first. The window never moves, so this URL always names the same seven days —
but a post published inside them and detected later is added when it arrives,
so a closed week can still gain entries.

Every entry is one first-party announcement, and carries the same eight fields
in this order before any others:

    Company · Announced · Category · Coverage · Announcement · Group · Source · Record

`Source` is the company's own page the announcement was traced to and `Record`
is this site's page for it. Both are safe to cite.

`Coverage` is a number of outlets, or `not counted` — which means this record
reached no count of its own, never that nobody wrote about it.

`Category` is this tracker's own analysis, and is authoritative wherever it is
stated. `not stated` means either that nothing has analysed the post yet or
that the analysis placed it outside this taxonomy; neither is a verdict about
the announcement.

`Announcement` is whether the post is news of something the company did, rather
than a how-to, a customer story, a changelog listing or a staffing note.

`Group` is which of this record's groups the entry falls in, carried on the
entry itself so it holds where the groups are headings and where they are not.
One of: `models`, `covered`, `announcements`, `routine`.

Anything after `Record` is held detail that only some announcements carry.

- This week's page: https://forck.live/week/2026-W17
- The current issue, and the catalogue of the other weeks: https://forck.live/briefing.md
- Catalogue: https://forck.live/llms.txt

## Announcements

### OpenAI — Our principles

- Company: OpenAI (openai.com)
- Announced: 2026-04-26T16:00:00+00:00
- Category: not stated
- Coverage: not counted
- Announcement: no
- Group: routine
- Source: https://openai.com/index/our-principles
- Record: https://forck.live/items/167-our-principles
- Subject: GPT / ChatGPT / API

OpenAI restates its mission to ensure AGI benefits all of humanity and announces that Sam Altman has shared five principles guiding the company's work.

### Sakana AI — Learning to Orchestrate Agents in Natural Language with the Conductor

- Company: Sakana AI (sakana.ai)
- Announced: 2026-04-26T15:00:00+00:00
- Category: research-paper
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://sakana.ai/learning-to-orchestrate
- Record: https://forck.live/items/4686-learning-to-orchestrate-agents-in-natural-language-with-the-conductor
- Subject: Sakana models

Sakana AI announces a research paper accepted at ICLR2026 on a Conductor model trained with reinforcement learning to orchestrate multiple AI agents via natural language, achieving state-of-the-art results on LiveCodeBench and GPQA-Diamond. The research also powers a new multi-agent system called Sakana Fugu.

### Sakana AI — Trinity: An Evolved LLM Coordinator

- Company: Sakana AI (sakana.ai)
- Announced: 2026-04-25T15:00:00+00:00
- Category: research-paper
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://sakana.ai/trinity
- Record: https://forck.live/items/4687-trinity-an-evolved-llm-coordinator
- Subject: Sakana models

Sakana AI announces a new paper, 'TRINITY: An Evolved LLM Coordinator', accepted at ICLR2026. The paper introduces a system that coordinates multiple LLMs by dynamically assigning roles (Thinker, Worker, Verifier) using a lightweight coordinator with fewer than 20K learnable parameters, trained via evolutionary algorithms. The system achieves state-of-the-art results on LiveCodeBench (86.2% pass@1) and generalizes zero-shot to unseen tasks, outperforming individual models including GPT-5, Gemini 2.5-Pro, and Claude-4-Sonnet.

### Amp — Opus 4.7

- Company: Amp (ampcode.com)
- Announced: 2026-04-25
- Category: model-update
- Coverage: not counted
- Announcement: yes
- Group: models
- Source: https://ampcode.com/news/opus-4.7
- Record: https://forck.live/items/7751-opus-4-7
- Subject: Amp

Amp released Opus 4.7, which now powers its smart mode. The model scored approximately 72% on internal evals, up from 65% for Opus 4.6. It follows prompts more closely and is better at complex coding tasks.

### OpenAI — Modeling an AI jobs transition

- Company: OpenAI (openai.com)
- Announced: 2026-04-25
- Category: research-paper
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://openai.com/index/modeling-ai-jobs-transition
- Record: https://forck.live/items/168-modeling-an-ai-jobs-transition
- Subject: GPT / ChatGPT / API

OpenAI introduces a framework analyzing 921 occupations and 148 million U.S. jobs to assess AI's impact on automation, reorganization, growth, or minimal disruption.

### Lovable — Testing GPT-5.5 in early access: what we are seeing so far

- Company: Lovable (lovable.dev)
- Announced: 2026-04-24T14:30:00+00:00
- Category: model-update
- Coverage: not counted
- Announcement: yes
- Group: models
- Source: https://lovable.dev/blog/gpt-5-5-now-in-lovable
- Record: https://forck.live/items/12087-testing-gpt-5-5-in-early-access-what-we-are-seeing-so-far
- Subject: Lovable / AI app builder

Lovable tested GPT-5.5 in early access and reports it outperforms GPT-5.4 on complex tasks, with a 12.5% improvement on the hardest benchmark at the same cost, 23.1% fewer tool calls per request, and 9.9% higher likelihood of resolving blocking tasks. The model is described as more cost-efficient and better at unblocking builders during stalled build sessions.

### Aleph Alpha — Sovereign AI for the World: Cohere and Aleph Alpha to Form Global AI Powerhouse as Nations and Enterprises Demand Control Over Their Technology

- Company: Aleph Alpha (aleph-alpha.com)
- Announced: 2026-04-24
- Category: partnership-acquisition
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://aleph-alpha.com/en/news/cohere-aleph-alpha-global-ai-powerhouse
- Record: https://forck.live/items/16857-sovereign-ai-for-the-world-cohere-and-aleph-alpha-to-form-global-ai-powerhouse
- Subject: Kolibri / PhariaAI

Cohere and Aleph Alpha announced plans to combine into a transatlantic AI company focused on sovereign AI for governments and regulated industries. The companies of Schwarz Group committed €500M (approx. $600M) in structured financing to back the venture, which will be anchored in Germany and Canada. The combined entity aims to provide an independent, enterprise-grade alternative for organizations that require control over their AI stack.

### DeepSeek — DeepSeek-V4 Preview: Entering the Era of Affordable Million-Token Context

- Company: DeepSeek (deepseek.com)
- Announced: 2026-04-24
- Category: new-model
- Coverage: not counted
- Announcement: yes
- Group: models
- Source: https://www.deepseek.com/en/news/v4-preview
- Record: https://forck.live/items/4319-deepseek-v4-preview-entering-the-era-of-affordable-million-token-context
- Subject: DeepSeek models

DeepSeek announces the preview release of DeepSeek-V4, including two models: V4-Pro (1.6T total / 49B active params) and V4-Flash (284B total / 13B active params), both open-sourced with a default 1M context window. The models feature enhanced agentic capabilities, world-class reasoning, and structural innovations for long-context efficiency. The API is updated and available today, and the older deepseek-chat and deepseek-reasoner models will be retired after July 24, 2026.

### Cursor — Multitask, Worktrees, and Multi-root Workspaces

- Company: Cursor (cursor.com)
- Announced: 2026-04-24
- Category: capability-change
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://cursor.com/changelog/04-24-26
- Record: https://forck.live/items/2618-multitask-worktrees-and-multi-root-workspaces
- Subject: Cursor

Cursor announces a new multitasking feature using async subagents, improved worktrees, and multi-root workspaces for cross-repo changes.

### Together AI — Accelerate RL rollouts by up to 50% with distribution-aware speculative decoding

- Company: Together AI (together.ai)
- Announced: 2026-04-24
- Category: not stated
- Coverage: not counted
- Announcement: no
- Group: routine
- Source: https://www.together.ai/blog/distribution-aware-speculative-decoding
- Record: https://forck.live/items/2356-accelerate-rl-rollouts-by-up-to-50-with-distribution-aware-speculative-decoding
- Subject: Inference platform

Distribution-aware speculative decoding (DAS) is a framework that reduces rollout time in RL post-training by up to 50% without altering model outputs. It uses a training-free adaptive suffix tree drafter and a length-aware scheduling strategy to handle stragglers and exploit historical trajectory data.

### Hugging Face — DeepSeek-V4: a million-token context that agents can actually use

- Company: Hugging Face (huggingface.co)
- Announced: 2026-04-24
- Category: new-model
- Coverage: not counted
- Announcement: yes
- Group: models
- Source: https://huggingface.co/blog/deepseekv4
- Record: https://forck.live/items/1512-deepseek-v4-a-million-token-context-that-agents-can-actually-use
- Subject: Platform

DeepSeek released two new MoE models, DeepSeek-V4-Pro (1.6T parameters, 49B active) and DeepSeek-V4-Flash (284B parameters, 13B active), both with a 1M-token context window. The models are designed for efficient long-context inference and agentic tasks, featuring hybrid attention (CSA and HCA) that reduces FLOPs and KV cache memory, plus post-training improvements for multi-turn tool use.

### Harvey — GPT-5.5: Research Preview Results

- Company: Harvey (harvey.ai)
- Announced: 2026-04-23T18:00:00+00:00
- Category: not stated
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://www.harvey.ai/blog/gpt-5-5-research-preview-results
- Record: https://forck.live/items/17578-gpt-5-5-research-preview-results
- Subject: Harvey

A first look at GPT-5.5 from Harvey. Today, we're sharing early access results for GPT-5.

### Sakana AI — Sakana Fugu: A Multi-Agent Orchestration System as a Foundation Model

- Company: Sakana AI (sakana.ai)
- Announced: 2026-04-23T15:00:00+00:00
- Category: availability-change
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://sakana.ai/fugu-beta
- Record: https://forck.live/items/4688-sakana-fugu-a-multi-agent-orchestration-system-as-a-foundation-model
- Subject: Sakana models

Sakana AI announces Sakana Fugu, a multi-agent orchestration system that coordinates frontier foundation models, now available for beta testing via API.

### OpenAI — Introducing GPT-5.5

- Company: OpenAI (openai.com)
- Announced: 2026-04-23T11:00:00+00:00
- Category: new-model
- Coverage: not counted
- Announcement: yes
- Group: models
- Source: https://openai.com/index/introducing-gpt-5-5
- Record: https://forck.live/items/170-introducing-gpt-5-5
- Subject: GPT / ChatGPT / API

OpenAI announced GPT-5.5, described as their smartest, faster, and more capable model for complex tasks such as coding, research, and data analysis across tools.

### OpenAI — Everyday work with ChatGPT Work

- Company: OpenAI (openai.com)
- Announced: 2026-04-23T10:00:00+00:00
- Category: not stated
- Coverage: not counted
- Announcement: no
- Group: routine
- Source: https://openai.com/academy/how-to-use-chatgpt-work-for-everyday-tasks
- Record: https://forck.live/items/4148-everyday-work-with-chatgpt-work
- Subject: GPT / ChatGPT / API

Practical use cases for ChatGPT Work, focusing on automating tasks, creating deliverables, and integrating real inputs into outputs across tools, files, and workflows.

### OpenAI — How to use ChatGPT Work for everyday tasks

- Company: OpenAI (openai.com)
- Announced: 2026-04-23T10:00:00+00:00
- Category: not stated
- Coverage: not counted
- Announcement: no
- Group: routine
- Source: https://openai.com/academy/how-to-use-chatgpt-work-for-everyday-tasks
- Record: https://forck.live/items/171-how-to-use-chatgpt-work-for-everyday-tasks
- Subject: GPT / ChatGPT / API

OpenAI published an educational guide on practical use cases for ChatGPT Work, focusing on automating tasks and creating deliverables.

### Harvey — And now, from Madrid

- Company: Harvey (harvey.ai)
- Announced: 2026-04-23T06:31:00+00:00
- Category: availability-change
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://www.harvey.ai/blog/and-now-from-madrid
- Record: https://forck.live/items/17579-and-now-from-madrid
- Subject: Harvey

Harvey opened a Madrid office with an eight-person team to support clients in Spain and Portugal. The company has been active in the Iberian market for over two years and lists partnerships with law firms including Cuatrecasas, Vieira de Almeida, PLMJ, and others. Harvey also expanded its academic partnership program to Portugal through a new collaboration with Católica Global School of Law.

### Ant Ling — Ling-2.6-flash Release: Faster Response, Stronger Execution, Higher Token Efficiency

- Company: Ant Ling (ant-ling.com)
- Announced: 2026-04-23
- Category: new-model
- Coverage: not counted
- Announcement: yes
- Group: models
- Source: https://developer.ant-ling.com/en/blogs/ling-2.6-flash-release
- Record: https://forck.live/items/22331-ling-2-6-flash-release-faster-response-stronger-execution-higher-token
- Subject: Ling / Ring / Ming

Ant Ling released Ling-2.6-flash, an Instruct model with 10.4 billion total parameters and 7.4 billion active parameters. The model uses a hybrid linear attention architecture (MLA + Lightning Linear in a 1:7 ratio) to achieve inference speeds up to 340 tokens/s on 4-H20 cards and prefill throughput 2.2 times that of Nemotron-3-Super. It consumed only 15 million tokens in an Artificial Analysis evaluation, roughly one-tenth the tokens used by Nemotron-3-Super, while achieving performance comparable to or state-of-the-art on agent benchmarks such as BFCL-V4, TAU2-bench, SWE-bench Verified, Claw-Eval, and PinchBench.

### Upstage — Upstage AI Honored with Two Stevie Awards in 2026 American Business Awards

- Company: Upstage (upstage.ai)
- Announced: 2026-04-23
- Category: funding-company-news
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://www.upstage.ai/news/upstage-stevie-awards-2026-american-business-awards
- Record: https://forck.live/items/18423-upstage-ai-honored-with-two-stevie-awards-in-2026-american-business-awards
- Subject: Solar Pro / Solar Mini

Upstage AI won two Stevie Awards in the 24th Annual American Business Awards: a Silver Stevie for Women in AI Leadership for US CEO Kasey Roh, and a Bronze Stevie for AI Company of the Year for the organization. The Bronze award recognized Upstage's proprietary LLM and OCR stack for transforming document processing in insurance and finance. Kasey Roh was honored for scaling Upstage's U.S. presence, securing $45M in funding, and driving 3x YoY growth.

### Factory AI — Automated QA: Testing Your App the Way a Real User Would

- Company: Factory AI (factory.com)
- Announced: 2026-04-23
- Category: capability-change
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://factory.com/news/automated-qa
- Record: https://forck.live/items/17816-automated-qa-testing-your-app-the-way-a-real-user-would
- Subject: Droid

Factory introduced Automated QA, a skill that drives an app like a real user—filling forms, typing into terminals, and calling endpoints—and posts a structured report with visual evidence directly to a pull request. The skill is available today in all Factory plans.

### Cua — Inside macOS window internals: how SkyLight enables multi-cursor background agents

- Company: Cua (cua.ai)
- Announced: 2026-04-23
- Category: not stated
- Coverage: not counted
- Announcement: no
- Group: routine
- Source: https://cua.ai/blog/inside-macos-window-internals
- Record: https://forck.live/items/16822-inside-macos-window-internals-how-skylight-enables-multi-cursor-background
- Subject: Cua Driver / Lume / Cua Bench / Cua Fleets

Published on April 23, 2026 by Francesco Bonacci TL;DR: cua-driver is an open-source macOS driver that lets any agent (Claude Code, Codex, your own harness) drive any Mac app in the background.

### Cognition — What We Learned Building Cloud Agents

- Company: Cognition (cognition.com)
- Announced: 2026-04-23
- Category: not stated
- Coverage: not counted
- Announcement: no
- Group: routine
- Source: https://cognition.com/blog/what-we-learned-building-cloud-agents
- Record: https://forck.live/items/7666-what-we-learned-building-cloud-agents
- Subject: Devin / SWE

Cognition shared lessons from over two years of building cloud agent infrastructure for Devin, covering technical challenges such as VM-level isolation for security, hypervisor-level snapshotting for persistence across async gaps, and the orchestration, governance, and integration layers needed to scale from one to hundreds of concurrent agent sessions.

### Anthropic — Built-in memory for Claude Managed Agents

- Company: Anthropic (anthropic.com)
- Announced: 2026-04-23
- Category: capability-change
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://claude.com/blog/claude-managed-agents-memory
- Record: https://forck.live/items/7408-built-in-memory-for-claude-managed-agents
- Subject: Claude

Anthropic announced built-in memory for Claude Managed Agents, available in public beta. The memory layer allows agents to learn from every session, with filesystem-based storage, audit logs, and API control. The company reports 97% fewer first-pass errors, 27% lower cost, and 34% lower latency.

### Anthropic — New connectors in Claude for everyday life

- Company: Anthropic (anthropic.com)
- Announced: 2026-04-23
- Category: capability-change
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://claude.com/blog/connectors-for-everyday-life
- Record: https://forck.live/items/7407-new-connectors-in-claude-for-everyday-life
- Subject: Claude

Anthropic expanded Claude's connectors to include everyday apps such as AllTrails, Instacart, Audible, and others. Connectors now dynamically appear in conversations based on user context. The feature is available on all plans, with mobile in beta.

### xAI — Grok Voice 1.0

- Company: xAI (x.ai)
- Announced: 2026-04-23
- Category: api-release
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://x.ai/news/grok-voice-think-fast-1
- Record: https://forck.live/items/3472-grok-voice-1-0
- Subject: Grok

xAI announces that Grok Voice Think Fast 1.0, their most capable voice agent, is now available via API.

### Hugging Face — How to Use Transformers.js in a Chrome Extension

- Company: Hugging Face (huggingface.co)
- Announced: 2026-04-23
- Category: developer-tool-release
- Coverage: not counted
- Announcement: no
- Group: routine
- Source: https://huggingface.co/blog/transformersjs-chrome-extension
- Record: https://forck.live/items/1513-how-to-use-transformers-js-in-a-chrome-extension
- Subject: Platform

Hugging Face published a guide for developers on how to use Transformers.js in a Chrome extension, based on their recently released demo extension powered by Gemma 4 E2B. The guide covers Manifest V3 architecture, model loading, messaging, and download/cache lifecycle.

### Google Research — It's all about the angle: Your photos, re-composed

- Company: Google Research (research.google)
- Announced: 2026-04-22T17:00:00+00:00
- Category: capability-change
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://research.google/blog/its-all-about-the-angle-your-photos-re-composed/
- Record: https://forck.live/items/1343-it-s-all-about-the-angle-your-photos-re-composed
- Subject: Research

Google Research introduces a new method for re-composing photos by changing the camera angle after the photo is taken, using 3D scene estimation and generative inpainting. The feature is now available in Google Photos' Auto frame feature.

### Replit — Introducing Replit Auto-Protect

- Company: Replit (replit.com)
- Announced: 2026-04-22T16:15:00+00:00
- Category: capability-change
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://replit.com/blog/auto-protect
- Record: https://forck.live/items/17908-introducing-replit-auto-protect
- Subject: Replit Agent

Replit introduced Auto-Protect, a feature that automatically checks projects against newly disclosed critical CVEs and, when opted in, has Replit Agent prepare and test a patch, then emails the user a link to apply it. Users can set the minimum severity level for automatic remediation and email notifications, both off by default. The feature is available to account admins via the Security Center pane.

### ByteDance — Seed3D 2.0 Released: Higher Precision and Greater Usability

- Company: ByteDance (bytedance.com)
- Announced: 2026-04-22T16:00:00+00:00
- Category: new-model
- Coverage: not counted
- Announcement: yes
- Group: models
- Source: https://seed.bytedance.com/en/blog/seed3d-2-0-released-higher-precision-and-greater-usability
- Record: https://forck.live/items/4664-seed3d-2-0-released-higher-precision-and-greater-usability
- Subject: Seed / Doubao

ByteDance released Seed3D 2.0, a next-generation 3D generative model with improved geometric precision and material quality, available via API on Volcano Engine.

### Tencent — Hy3 preview: The First Step in Rebuilding the Hy model

- Company: Tencent (tencent.com)
- Announced: 2026-04-22T16:00:00+00:00
- Category: open-weight-release
- Coverage: not counted
- Announcement: yes
- Group: models
- Source: https://hunyuan.tencent.com/research/hy3-preview
- Record: https://forck.live/items/4580-hy3-preview-the-first-step-in-rebuilding-the-hy-model
- Subject: Hunyuan

Tencent releases and open-sources Hy3 preview, a 295B-parameter MoE model with 21B active parameters, supporting three reasoning modes and a 256K context window. The model is the first trained on rebuilt infrastructure and shows improvements in reasoning, instruction following, coding, and agent tasks. It is available on GitHub, HuggingFace, ModelScope, and GitCode, and supports vLLM and SGLang. API pricing is competitive, and Token Plans start at 28 CNY/month.

### OpenAI — Making ChatGPT better for clinicians

- Company: OpenAI (openai.com)
- Announced: 2026-04-22T15:00:00+00:00
- Category: availability-change
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://openai.com/index/making-chatgpt-better-for-clinicians
- Record: https://forck.live/items/172-making-chatgpt-better-for-clinicians
- Subject: GPT / ChatGPT / API

OpenAI is making ChatGPT for Clinicians available for free to verified physicians, nurse practitioners, and pharmacists in the United States for clinical care, documentation, and research use.

### Lovable — Our response to the April 2026 incident

- Company: Lovable (lovable.dev)
- Announced: 2026-04-22T14:49:47+00:00
- Category: safety-policy-update
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://lovable.dev/blog/our-response-to-the-april-2026-incident
- Record: https://forck.live/items/12088-our-response-to-the-april-2026-incident
- Subject: Lovable / AI app builder

Lovable fixed a security vulnerability that allowed any authenticated user to access chat history and source code of public projects between February 3 and April 20, 2026. The company is making all current public projects private except official templates, redesigning project visibility UX, and restructuring its HackerOne triage process. Private projects and Lovable Cloud were never impacted.

### Harvey — Harvey in Practice: How Real Estate Teams Move Deals Forward

- Company: Harvey (harvey.ai)
- Announced: 2026-04-22T13:00:00+00:00
- Category: not stated
- Coverage: not counted
- Announcement: no
- Group: routine
- Source: https://www.harvey.ai/blog/harvey-in-practice-real-estate-teams
- Record: https://forck.live/items/17580-harvey-in-practice-how-real-estate-teams-move-deals-forward
- Subject: Harvey

Harvey's blog describes how real estate legal teams use its platform across the full lifecycle of a transaction, from drafting letters of intent to reconciling leases during diligence and finalizing agreements. The post highlights that at GSK Stockmann, teams using Harvey can generate red-flag summaries in hours instead of days and reduce diligence review time by up to 75% in unstructured data rooms. At Carrefour Spain, the legal team uses Harvey to accelerate real estate contract analysis and drafting, reducing turnaround times.

### OpenAI — Introducing workspace agents in ChatGPT

- Company: OpenAI (openai.com)
- Announced: 2026-04-22T10:00:00+00:00
- Category: capability-change
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://openai.com/index/introducing-workspace-agents-in-chatgpt
- Record: https://forck.live/items/175-introducing-workspace-agents-in-chatgpt
- Subject: GPT / ChatGPT / API

OpenAI introduced workspace agents in ChatGPT, which are Codex-powered agents that automate complex workflows, run in the cloud, and help teams scale work across tools securely.

### OpenAI — Speeding up agentic workflows with WebSockets in the Responses API

- Company: OpenAI (openai.com)
- Announced: 2026-04-22T10:00:00+00:00
- Category: capability-change
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://openai.com/index/speeding-up-agentic-workflows-with-websockets
- Record: https://forck.live/items/174-speeding-up-agentic-workflows-with-websockets-in-the-responses-api
- Subject: GPT / ChatGPT / API

WebSockets and connection-scoped caching in the Responses API reduce API overhead and improve model latency for agentic workflows.

### OpenAI — Workspace agents

- Company: OpenAI (openai.com)
- Announced: 2026-04-22T10:00:00+00:00
- Category: not stated
- Coverage: not counted
- Announcement: no
- Group: routine
- Source: https://openai.com/academy/workspace-agents
- Record: https://forck.live/items/173-workspace-agents
- Subject: GPT / ChatGPT / API

This is a course or guide on building workspace agents in ChatGPT for workflow automation and tool integration.

### Alibaba — Qwen3.6-27B: Flagship-Level Coding in a 27B Dense Model

- Company: Alibaba (alibaba.com)
- Announced: 2026-04-22T02:00:00+00:00
- Category: open-weight-release
- Coverage: not counted
- Announcement: yes
- Group: models
- Source: https://qwen.ai/blog?id=qwen3.6-27b
- Record: https://forck.live/items/4605-qwen3-6-27b-flagship-level-coding-in-a-27b-dense-model
- Subject: Qwen

Alibaba open-sources Qwen3.6-27B, a dense 27-billion-parameter multimodal model that supports both thinking and non-thinking modes and delivers flagship-level agentic coding performance, surpassing the previous-generation open-source flagship Qwen3.5-397B-A17B across all major coding benchmarks.

### Supabase — Supabase is now ISO 27001 certified

- Company: Supabase (supabase.com)
- Announced: 2026-04-22
- Category: not stated
- Coverage: not counted
- Announcement: no
- Group: routine
- Source: https://supabase.com/blog/supabase-is-now-iso-27001-certified
- Record: https://forck.live/items/21291-supabase-is-now-iso-27001-certified
- Subject: Supabase

Supabase has achieved ISO/IEC 27001:2022 certification for its information security management system, covering the entire platform including Database, Auth, Storage, Realtime, Edge Functions, and the Data API. The certification is valid for three years with annual surveillance audits. Team and Enterprise plan customers can request the certificate from their dashboard.

### Factory AI — Droid Computers

- Company: Factory AI (factory.com)
- Announced: 2026-04-22
- Category: capability-change
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://factory.com/news/droid-computers
- Record: https://forck.live/items/17830-droid-computers
- Subject: Droid

Factory launched Droid Computers, persistent cloud machines for remotely orchestrating Droids. Users can spin up a machine in Factory's cloud or register their own machine via Bring Your Own Machine. Managed Computers are available for Max, Growth, and Enterprise plans, while BYOM is free for all users.

### Laminar — Agent traces aren't backend traces. Stop reading them like they are.

- Company: Laminar (laminar.sh)
- Announced: 2026-04-22
- Category: not stated
- Coverage: not counted
- Announcement: no
- Group: routine
- Source: https://laminar.sh/blog/transcript-view-agent-traces
- Record: https://forck.live/items/16851-agent-traces-aren-t-backend-traces-stop-reading-them-like-they-are
- Subject: flow / agent observability

Laminar introduces a transcript view for agent traces that displays the agent's loop—LLM calls and tool calls in sequence—alongside a timeline strip and collapsible subagent cards, designed to make debugging agent runs easier than traditional span-tree views.

### Cognition — Multi-Agents: What's Actually Working

- Company: Cognition (cognition.com)
- Announced: 2026-04-22
- Category: capability-change
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://cognition.com/blog/multi-agents-working
- Record: https://forck.live/items/7667-multi-agents-what-s-actually-working
- Subject: Devin / SWE

Cognition describes how it has begun deploying multi-agent systems that work in practice, focusing on patterns where multiple agents contribute intelligence while writes stay single-threaded. The company reports that its Devin Review catches an average of 2 bugs per pull request, with roughly 58% being severe, and that Devin and Devin Review now natively iterate against each other so most bugs are resolved before a human opens the PR.

### OpenAI — Introducing OpenAI Privacy Filter

- Company: OpenAI (openai.com)
- Announced: 2026-04-22
- Category: open-weight-release
- Coverage: not counted
- Announcement: yes
- Group: models
- Source: https://openai.com/index/introducing-openai-privacy-filter
- Record: https://forck.live/items/176-introducing-openai-privacy-filter
- Subject: GPT / ChatGPT / API

OpenAI announces Privacy Filter, an open-weight model for detecting and redacting PII in text with state-of-the-art accuracy.

### Replit — Replit Wins 2026 Google Cloud Partner of the Year Award

- Company: Replit (replit.com)
- Announced: 2026-04-21T19:05:00+00:00
- Category: funding-company-news
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://replit.com/blog/google-cloud-partner-award-winner
- Record: https://forck.live/items/17910-replit-wins-2026-google-cloud-partner-of-the-year-award
- Subject: Replit Agent

Google Cloud awarded Replit its 2026 AI Tooling Partner of the Year, recognizing the company for helping customers leverage AI to achieve success through Google Cloud. Replit is available on the Google Cloud Marketplace, integrating with services like Cloud Run, Kubernetes Engine, and BigQuery.

### Replit — Meet Replit Security Agent

- Company: Replit (replit.com)
- Announced: 2026-04-21T18:02:07+00:00
- Category: capability-change
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://replit.com/blog/meet-replit-security-agent
- Record: https://forck.live/items/17911-meet-replit-security-agent
- Subject: Replit Agent

Replit introduced Security Agent, a tool that performs a comprehensive security review of an app's codebase in under an hour. It uses a customizable threat modeling plan, leverages Semgrep and HoundDog.ai, and generates a report of identified risks that can be passed to Replit Agent for remediation.

### Google Research — ReasoningBank: Enabling agents to learn from experience

- Company: Google Research (research.google)
- Announced: 2026-04-21T16:42:22+00:00
- Category: research-paper
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://research.google/blog/reasoningbank-enabling-agents-to-learn-from-experience/
- Record: https://forck.live/items/1344-reasoningbank-enabling-agents-to-learn-from-experience
- Subject: Research

ReasoningBank is a novel agent memory framework that uses successful and failed experiences to distill generalizable reasoning strategies, enabling agents to learn continuously after deployment. It also introduces memory-aware test-time scaling (MaTTS) to leverage exploration for better learning.

### Harvey — Rebuilding the Review Algorithm to Increase Accuracy and Speed

- Company: Harvey (harvey.ai)
- Announced: 2026-04-21T16:00:00+00:00
- Category: capability-change
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://www.harvey.ai/blog/rebuilding-harveys-review-algorithm
- Record: https://forck.live/items/17581-rebuilding-the-review-algorithm-to-increase-accuracy-and-speed
- Subject: Harvey

Harvey rebuilt the algorithm behind its review tables to produce more accurate answers, sentence-level citations, and faster results. The new algorithm replaces the previous summary and additional context fields with a single answer field and adds a reasoning field that explains the model's logic. Sentence-level citations improve verifiability and reduce latency, especially for large tables.

### Google DeepMind — Partnering with industry leaders to accelerate AI transformation

- Company: Google DeepMind (deepmind.google)
- Announced: 2026-04-21T14:54:15+00:00
- Category: partnership-acquisition
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://deepmind.google/blog/partnering-with-industry-leaders-to-accelerate-ai-transformation/
- Record: https://forck.live/items/1239-partnering-with-industry-leaders-to-accelerate-ai-transformation
- Subject: Gemini

Google DeepMind announces partnerships with global consultancies to help organizations adopt frontier AI.

### OpenAI — Introducing ChatGPT Images 2.0

- Company: OpenAI (openai.com)
- Announced: 2026-04-21T12:00:00+00:00
- Category: model-update
- Coverage: not counted
- Announcement: yes
- Group: models
- Source: https://openai.com/index/introducing-chatgpt-images-2-0
- Record: https://forck.live/items/177-introducing-chatgpt-images-2-0
- Subject: GPT / ChatGPT / API

OpenAI announces ChatGPT Images 2.0, a new image generation model with improved text rendering, multilingual support, and advanced visual reasoning.

### TII — QIMMA قِمّة ⛰: A Quality-First Arabic LLM Leaderboard

- Company: TII (tii.ae)
- Announced: 2026-04-21T10:09:58+00:00
- Category: research-paper
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://huggingface.co/blog/tiiuae/qimma-arabic-leaderboard
- Record: https://forck.live/items/1514-qimma-a-quality-first-arabic-llm-leaderboard

TII's QIMMA leaderboard applies a multi-stage quality validation pipeline to Arabic benchmarks before evaluating models, discarding samples with systematic issues such as false gold answers and cultural misalignment. It consolidates 109 subsets from 14 benchmarks into over 52,000 samples across seven domains, including the first Arabic leaderboard with code evaluation. The post reports discard rates per benchmark, with ArabicMMLU losing 3.1% of its samples.

### Replit — Defense in Depth: How Replit Secures Every Layer of the Vibe Coding Stack

- Company: Replit (replit.com)
- Announced: 2026-04-21T03:00:43+00:00
- Category: not stated
- Coverage: not counted
- Announcement: no
- Group: routine
- Source: https://replit.com/blog/defense-in-depth-how-replit-secures-every-layer-of-the-vibe-coding-stack
- Record: https://forck.live/items/17909-defense-in-depth-how-replit-secures-every-layer-of-the-vibe-coding-stack
- Title in English (translated by forck.live from en): Defense in Depth: How Replit Secures Every Layer of the Vibe Coding Stack
- Subject: Replit Agent

Replit published a technical walkthrough of its security architecture, covering isolated cloud sandboxes, structural separation of frontend and backend, supply chain protection via Determinate Nix, full separation of development and production environments, and continuous backups with append-only git. The post explains how these layers implement Zero Trust principles and defense in depth for the Replit platform.

### Odyssey — Introducing Odyssey-2 Max: Scaled World Simulation

- Company: Odyssey (odyssey.systems)
- Announced: 2026-04-21
- Category: new-model
- Coverage: not counted
- Announcement: yes
- Group: models
- Source: https://odyssey.systems/introducing-odyssey-2-max
- Record: https://forck.live/items/18598-introducing-odyssey-2-max-scaled-world-simulation
- Subject: Odyssey world models

Odyssey introduced Odyssey-2 Max, its largest and most powerful general world model, which advances the state of the art in physical accuracy of world models. The model simulates open-ended futures through continuous, interactive rollouts that evolve with actions in real time, a capability the company says bidirectional video models such as Sora, Veo, Kling, and Runway lack.

### LG AI Research — [AACR 2026] Spatial biology and AI research at LG AI Research-VUMC

- Company: LG AI Research (lgresearch.ai)
- Announced: 2026-04-21
- Category: research-paper
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://www.lgresearch.ai/blog/view?seq=645
- Record: https://forck.live/items/4751-aacr-2026-spatial-biology-and-ai-research-at-lg-ai-research-vumc
- Subject: EXAONE

LG AI Research and Vanderbilt University Medical Center presented joint research at the 2026 AACR covering spatial biology, 3D/4D image analysis, pathology AI, and foundation models for cancer research.

### Together AI — Capacity without conflict: A guide to multi-tenant GPU cluster design for AI-native teams

- Company: Together AI (together.ai)
- Announced: 2026-04-21
- Category: not stated
- Coverage: not counted
- Announcement: no
- Group: routine
- Source: https://www.together.ai/blog/multi-tenant-gpu-cluster-design-for-ai-native-teams
- Record: https://forck.live/items/2357-capacity-without-conflict-a-guide-to-multi-tenant-gpu-cluster-design-for-ai
- Subject: Inference platform

Together AI publishes a guide on designing multi-tenant GPU clusters for AI-native teams, covering core design principles, failure modes, and their own implementation.

### Hugging Face — AI and the Future of Cybersecurity: Why Openness Matters

- Company: Hugging Face (huggingface.co)
- Announced: 2026-04-21
- Category: not stated
- Coverage: not counted
- Announcement: no
- Group: routine
- Source: https://huggingface.co/blog/cybersecurity-openness
- Record: https://forck.live/items/1515-ai-and-the-future-of-cybersecurity-why-openness-matters
- Subject: Platform

Mythos is a frontier AI model that can process software code and is embedded in a system that can find and patch vulnerabilities.

### OpenAI — Scaling Codex to enterprises worldwide

- Company: OpenAI (openai.com)
- Announced: 2026-04-21
- Category: product-launch
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://openai.com/index/scaling-codex-to-enterprises-worldwide
- Record: https://forck.live/items/178-scaling-codex-to-enterprises-worldwide
- Subject: GPT / ChatGPT / API

OpenAI announced the launch of Codex Labs, a new product to help enterprises deploy and scale Codex across the software development lifecycle, along with partnerships with Accenture, PwC, Infosys, and others, and reported reaching 4 million weekly active users for Codex.

### Sakana AI — String Seed of Thought: Prompting LLMs for Distribution-Faithful and Diverse Generation

- Company: Sakana AI (sakana.ai)
- Announced: 2026-04-20T15:00:00+00:00
- Category: research-paper
- Coverage: not counted
- Announcement: yes
- Group: announcements
- Source: https://sakana.ai/ssot
- Record: https://forck.live/items/4689-string-seed-of-thought-prompting-llms-for-distribution-faithful-and-diverse
- Subject: Sakana models

Sakana AI announces a prompting technique called String Seed of Thought (SSoT) that reduces output bias and increases diversity in LLM generations, accepted at ICLR2026.

### Moonshot AI — Kimi K2.6

- Company: Moonshot AI (moonshot.ai)
- Announced: 2026-04-20
- Category: open-weight-release
- Coverage: not counted
- Announcement: yes
- Group: models
- Source: https://www.kimi.ai/blog/kimi-k2-6
- Record: https://forck.live/items/4623-kimi-k2-6
- Subject: Kimi

Moonshot AI is open-sourcing Kimi K2.6, a model with state-of-the-art coding, long-horizon execution, and agent swarm capabilities. It is available via Kimi.com, Kimi App, API, and Kimi Code. The model shows significant improvements over K2.5 in long-horizon coding tasks, as demonstrated by internal benchmarks and enterprise beta tests.

### OpenAI — OpenAI helps Hyatt advance AI among colleagues

- Company: OpenAI (openai.com)
- Announced: 2026-04-20
- Category: not stated
- Coverage: not counted
- Announcement: no
- Group: routine
- Source: https://openai.com/index/hyatt-advances-ai-with-chatgpt-enterprise
- Record: https://forck.live/items/179-openai-helps-hyatt-advance-ai-among-colleagues
- Subject: GPT / ChatGPT / API

Hyatt deploys ChatGPT Enterprise across its global workforce, using GPT-5.4 and Codex to improve productivity, operations, and guest experiences.