Changelog

MachinaCheck is a multi-agent CNC manufacturability system built on AMD MI300X. It matters because it applies multi-agent AI to manufacturing validation, suggesting faster automated checks for whether parts can be machined.

Open-source AI agent from Nous Research with a built-in learning loop — creates skills from experience, persists knowledge, builds a model of its user across sessions. ~60K stars in two months.

Open-source autonomous AI agent (formerly Clawdbot, Moltbot — lobster theme). Self-hosted; supports Claude, GPT, KIMI, MiMo, Qwen 3, Llama 4, Mistral, plus local models via Ollama. ~347K GitHub stars.

In-browser AI coding agent powered by StackBlitz WebContainers — generates and runs full Node.js apps without ever leaving the tab.

Conversational app builder — build a full-stack web app (database, auth, UI, deploys) in plain English. Stripe integration for monetized SaaS.

Vercel's UI-generation agent. Describe a component or app, get production-ready React + Tailwind + shadcn/ui code, deploy in one click.

Perplexity's agentic web browser — search, summarize, take actions across tabs, all powered by their answer engine.

OpenAI's browser-using agent. Books flights, fills forms, runs research tasks autonomously inside a sandboxed browser.

Anthropic's screen-control capability — Claude clicks, types, and scrolls through any GUI app via screenshots and keyboard/mouse primitives.

General-purpose autonomous agent that decomposes a goal into parallel sub-agents and executes browser, code, and analysis tasks end-to-end.

Cognition Labs' autonomous AI software engineer — runs sessions for hours, planning and executing complex multi-step engineering work.

Replit's in-browser coding agent. Agent 4 (Mar 2026) introduced parallel task forking that auto-resolves merge conflicts ~90% of the time.

Open-source CLI coding agent. ~147K stars by April 2026. Official GitHub Copilot partnership lets paid Copilot subscribers auth directly into it.

VS Code extension AI coding agent. 5M+ installs — the most-installed open-source coding agent.

The most widely-adopted AI coding tool, ~15M developer accounts. Multi-model since 2024 — pick your preferred backbone per task.

AI-native code editor, fork of VS Code. Hit $2B ARR in February 2026. Composer mode for natural-language multi-file refactors.

OpenAI's coding agent — runs in CLI, IDE, and cloud-hosted runtimes. Drives the Plus/Pro coding workflow.

Anthropic's CLI coding agent. Top of the SWE-bench Verified leaderboard at 80.8% on Opus 4.6. Lives in the terminal — reads files, executes commands, navigates repos, applies multi-file edits autonomously.

Alibaba's first dense open-weight in the 3.6 family. Strong agentic-coding scores (77.2 SWE-bench Verified, matching Claude 4.5 Opus on Terminal-Bench 2.0). Supports 201 languages and multimodal text/image/video input.

Smaller, faster sibling to DeepSeek-V4-Pro. Same 1M context window with a much lighter 284B / 13B-active MoE.

DeepSeek's flagship open-weight MoE. 1.6T parameters with 49B activated, 1M-token context, and a hybrid attention scheme (CSA + HCA) that delivers long-context inference at ~27% of V3.2's FLOPs.

Google DeepMind's flagship multimodal model. Best-in-class for multimodal understanding and agentic / vibe coding workflows. A "Deep Think" reasoning mode is available to AI Ultra subscribers.

OpenAI's smartest and most intuitive model — successor to GPT-5, with major coding, research, and document workflow gains. GPT-5.5 Pro variant also available for heavier reasoning.

Predecessor to the Qwen3.6 family.

Reasoning-focused open-weight model.

Large MoE open-weight model. Predecessor to DeepSeek-V4.

Mixture-of-experts open-weight model from Meta.

Predecessor to Gemini 3 Pro.

Fast, cost-effective Claude model.

Anthropic's most capable model, optimized for complex reasoning and coding. Improved software engineering, long-running coding tasks, and higher-resolution vision over Claude 4.6.

Reasoning-focused model in the o-series.

OpenAI's flagship general-purpose model. Superseded by GPT-5.5 in April 2026.

Nvidia has already committed $40 billion to equity AI deals this year, continuing to expand its investment footprint across the AI ecosystem. The scale of that commitment underscores how central Nvidia has become to financing and shaping the industry around its GPUs, even as demand for AI infrastructure keeps rising.

The latest updates show AI data centers expanding at breakneck speed, with projects like OpenAI, Oracle, and SoftBank’s five new Stargate sites, Anthropic’s $50 billion US buildout, Microsoft’s 15-data-center plan in Mount Pleasant, Wisconsin, and a 40,000-acre Utah campus moving forward alongside talk of space-based facilities. They matter because the boom is already driving fights over power grids, electricity bills, water use, pollution, and health, with reports of 43% of Americans blaming data centers for rising bills, possible rolling blackouts, and communities seeing costs as high as 267% above pre-data-center levels.

Sierra raised $950 million at a $15 billion valuation to lead the week’s 10 biggest U.S. startup funding rounds, followed by Astranis’s $455 million, Anagram Therapeutics’s $250 million, and Blitzy’s $200 million, with the rest spanning insurance, renewable energy, AI infrastructure, vertical AI, and gaming. The list shows enterprise AI still commanding the largest checks, but space tech, biotech, and energy-linked computing are also attracting major capital from firms including Google Ventures, Tiger Global, Snowpoint Ventures, Franklin Templeton, Blackstone Life Sciences, TCV, Peter Thiel, KKR, and Andreessen Horowitz.

Court documents from the Musk v. Altman trial show Microsoft executives discussing early OpenAI investment and worrying the startup might “storm off to Amazon” and “shit-talk” Azure while the companies were negotiating their AI partnership. The emails offer a rare look at how seriously Microsoft treated OpenAI’s leverage in 2017, right after OpenAI’s Dota 2 bot beat a professional player, and how quickly the relationship moved beyond a simple congratulatory exchange into a major strategic deal.

Venture funding for sales, marketing and CRM startups is on track to be roughly flat in 2026 at about $3.7 billion, versus a 2021-2022 peak and more than $20 billion in boom-era years, but a larger share is now going to AI-focused companies. Recent standouts include Sierra’s $950 million round at a $15 billion valuation, Hightouch’s $150 million Series D at $2.75 billion, Netomi’s $110 million raise, Actively’s $45 million Series B, and Parloa’s $350 million Series D at a $3 billion valuation, while exits like Adyen’s roughly $880 million Talon.One deal and NICE’s $955 million Cognigy acquisition show some M&A despite muted IPO activity.

Pit, a new AI startup led by the co-founders of European scooter giant Voi, is raising a $16 million seed round led by a16z. The backing signals continued investor appetite for Stockholm-based AI startups and gives Pit a high-profile launch with major venture support.

SpaceX is planning to invest at least $55 billion in its “Terafab” AI chip plant in Austin, Texas, according to a public hearing notice tied to a tax-break request, with the total potentially reaching $119 billion if later phases are built. Musk originally said in March that the factory could eventually produce enough chips to support up to 200 gigawatts per year of compute, making it one of the most ambitious chip manufacturing plans in the U.S.

Elon Musk and Sam Altman are in a jury trial over Musk’s 2024 lawsuit seeking to remove Altman and Greg Brockman, force OpenAI to stop operating as a public benefit corporation, and win up to $150 billion for OpenAI’s nonprofit, with testimony already heard from Musk, Jared Birchall, Greg Brockman, Shivon Zilis, and Mira Murati. The case matters because it could reshape OpenAI’s governance and ChatGPT’s future, and it has also put Microsoft CEO Satya Nadella and former chief scientist Ilya Sutskever on the witness list while highlighting the clash between OpenAI’s nonprofit origins and Musk’s xAI/Grok rivalry.

Version 2.36.0 was released on 2026-05-07, with two listed features: `api: manual updates` and `api: realtime 2`. The update suggests continued API development, especially around real-time capabilities, though the excerpt provides no further implementation details.

Fazeshift raised a $17 million Series A led by F-Prime Capital, with participation from Gradient Ventures, Y Combinator, Wayfinder Ventures, Pioneer Fund, Ritual Capital and angels, bringing total funding to $22 million since its 2023 founding. The startup says its AI agents automate more than 90% of manual accounts receivable work, and its 12x revenue growth over the past year suggests investors see AR as a wedge into a broader autonomous finance suite.

American Express Ventures says its investment bar has shifted toward agentic commerce systems that can understand preferences, constraints, and intent, orchestrate end-to-end workflows, and complete authorized transactions, while it backs startups like Palm, Bluefish, and Candex. The notable detail is that Amex is framing AI as a path from credit cards to a “global agentic concierge,” but Tsang said the firm cares less about any single model or trust layer than about how those pieces combine into seamless, secure customer experiences.

v2.35.1, released on 2026-05-06, fixes an API regression in the image generation size enum. The change is a targeted bug fix that restores expected size handling without adding new features, which can prevent broken imagegen requests in clients relying on that enum.

Version 2.35.0, dated 2026-05-06, adds API features for “update image 2” and manual updates, updates the top_logprobs parameter description across chat and responses, and includes chores to remove the legacy Python CLI and rename its entrypoint. The release is notable more for cleanup and documentation consistency than major new capability, with the legacy Python CLI explicitly phased out.

v0.100.0, released on 2026-05-06, adds API support for Managed Agents multiagents and outcomes, plus webhooks and vault validation, alongside a webhook configuration fix. The update expands agent orchestration capabilities while tightening API integration details for webhook handling.

AI tools have lowered the cost of starting a company, but investors are under more pressure to back fewer, stronger seed bets, so the focus has shifted from raw technical ability to founder-market fit, domain expertise, and evidence of customer discovery. Carta says the average seed-stage company had just over 6 employees last year, down from more than 10 in 2021, and the piece argues that AI also makes it easier to fake startup credibility, especially in software, while deep tech remains harder to spoof.

Crunchbase says 28 companies joined the Unicorn Board in April, with AI-related startups making up 26 of them and robotics and frontier labs leading for a second straight month. London-based Ineffable Intelligence raised a $1.1 billion seed at a $5.1 billion valuation, Recursive Superintelligence raised a $500 million Series A at $4.5 billion, and six humanoid robotics companies from China and Japan also crossed $1 billion valuations.

Blitzy, an autonomous software development startup, raised $200 million in a round led by Northzone at a $1.4 billion valuation, bringing its total funding to more than $204.4 million. The Cambridge, Mass.-based company says its platform autonomously completes months of enterprise software development and has been adopted by dozens of Global 2000 companies across 10 industries, with claims of 5x engineering velocity.

v0.99.0, released on 2026-05-05, adds a client feature to target a workspace for OIDC federation token exchange. This matters because it enables more granular authentication flows across workspaces, which can simplify federated access management.

The piece says M&A deals often fail before they start because founders and investors anchor on prior fundraising marks or rare AI outlier transactions, while buyers price to demonstrated revenue, growth, retention, and strategic fit. It cites examples like a company that raised $10 million at a $40 million valuation with no revenue, plus Microsoft’s roughly $650 million Inflection AI licensing-and-talent deal, to show why expectations must align with how M&A actually works.

Global venture funding hit $56 billion in April, up 100% year over year, with Anthropic raising $15 billion and Jeff Bezos’s Project Prometheus raising $10 billion, while AI overall drew $37 billion, or 66% of all venture investment. The month shows how concentrated the AI boom has become, since those two rounds alone accounted for 45% of April VC and nearly 60% of year-to-date venture capital has gone to just five companies.

OpenAI introduced MRC (Multipath Reliable Connection), a new supercomputer networking protocol released through OCP to improve resilience and performance in large-scale AI training clusters. The protocol is aimed at large-scale AI training networks where reliability and throughput are critical, and its OCP release suggests broader hardware interoperability for cluster operators.

v0.98.1 was released on 2026-05-04 as a small maintenance update after v0.98.0, with the only listed change being a typo fix in an example (#1754). It matters mainly as a housekeeping release, showing no functional changes beyond cleanup.

Version 2.34.0 adds `external_key_id` to projects, `email`/`metadata` params to users, support for Admin API Keys per endpoint, and env-based header configuration, along with several API and type updates. It also fixes Azure auth headers, admin auth handling, timestamp/type issues, and multipart file array formatting, while improving multipart file copying performance and updating file rate-limit/vector-store docs.

Version 0.98.0 adds improved Managed Agents APIs, client support for Workload Identity Federation, interactive OAuth, and auth profiles, plus the ability to set headers via environment variables. It also fixes streaming stop_details propagation, multipart file array field naming, and missing us/eu multi-region base_url branches in the Vertex async client, which should reduce auth and integration issues.

True Anomaly led the week’s U.S. startup funding with a $600 million Series D led by Eclipse and Riot Ventures, while Rogo raised $160 million, Hightouch $150 million, Avoca $125 million, Netomi $110 million, and Parallel and Scout AI each took in $100 million. The round sizes show defense tech drawing unusually large venture checks alongside AI startups in fintech, marketing, customer service, and developer tools, and True Anomaly’s $1.1 billion total funding underscores investor interest in space security amid geopolitical tension.

A timeline of GPT-5 “goblin outputs” tracks how a personality-driven quirk spread through model behavior, identifies its root cause, and outlines fixes. The notable detail is that the issue appears tied to model personality dynamics rather than a single isolated bug, which makes diagnosing and correcting it more complex.

OpenAI is scaling Stargate to build the compute infrastructure for AGI, adding new data center capacity to meet growing AI demand. The notable detail is that this is about expanding the underlying infrastructure rather than a specific model release, signaling a larger push to support future AI workloads.

v2.33.0, released on 2026-04-28, adds an API update and includes a bug fix that changes the `prompt_cache_retention` enum value from `in-memory` to `in_memory`, along with a CI chore removing the release-doctor workflow. The enum correction is the notable change because it can affect clients that rely on the exact API value, while the rest of the release is a small maintenance update.

v0.97.0, released on 2026-04-23, adds CMA Memory as a public beta in the API, fixes errors in the API spec, and restores missing API features. It also improves multipart request file-structure copying performance on the client side and includes a few maintenance updates, including a steadier test dependency bumped to v0.22.1.

Google released a new video explaining how its TPUs power increasingly demanding AI workloads. It highlights the role of TPUs in handling heavier AI training and inference demands, though the excerpt provides no specific performance numbers or model names.

Google is launching two specialized eighth-generation TPU chips for the “agentic era” of AI. The notable detail is that these are purpose-built accelerators, signaling a shift toward hardware optimized for AI agents rather than general-purpose model training alone.

v0.96.0, released 2026-04-16, adds `claude-opus-4-7`, token budgets, and `user_profiles` in the API, and removes the `release-doctor` workflow from CI. The update expands model and account-management capabilities while simplifying the CI pipeline.

v2.32.0, released 2026-04-15, adds `detail` to `InputFileContent`, introduces an `OAuthErrorCode` type, and expands websocket client support with event handlers, enqueueing before connection, and reconnection. The update also fixes file uploads so file data are sent as a single parameter, which should reduce API misuse and make websocket-driven clients more resilient.

v0.95.0, released on 2026-04-14, marks Sonnet and Opus 4 as deprecated and changes the Bedrock Mantle client to use an auth header. Deprecating two Claude 4 models signals a shift in supported model tiers, while the Bedrock auth change affects how clients authenticate in production integrations.

v0.94.1, released on 2026-04-13, is a patch update from v0.94.0 that fixes streaming by adding missing events. The change is narrow but important because missing stream events can break real-time outputs or downstream consumers relying on complete event sequences.