April 2026: The Biggest Month Ever for Open-Source AI — Complete Breakdown (Apr 1–22)
📑 Table of Contents
🎯 The 30-Second Summary
Between April 1 and April 22, 2026, at least 19 major open-source AI models, tools, or infrastructure updates shipped. Three separate Chinese open-weight models topped SWE-Bench Pro at different points in the month. Apache 2.0 became the dominant license of the wave. And the infrastructure layer — vLLM, Haystack, CrewAI, Open Interpreter — shipped major upgrades to actually run all of it. This was not a coincidence. It was a phase transition.
Nobody planned for all of this to drop in the same three weeks. And yet here we are.
Between April 1 and April 22, 2026, the open-source AI ecosystem produced a release wave with no precedent in the field’s history. By the time Kimi K2.6 landed on April 20 — claiming the new #1 SWE-Bench Pro score of 58.6 — three separate open-weight models from Chinese labs had each held the global coding leaderboard top spot within the same calendar month. The closed-model labs were shipping too: Claude Opus 4.7, Claude Design, and the new OpenAI Codex all dropped in week three. But the story of April 2026 is the open-source ecosystem finally arriving — not as an alternative — but as the competition.
The Big Picture: Why This Month Is Different
Open-source AI has had good months before. What makes April 2026 exceptional isn’t any single release — it’s the sustained density. A new significant release, on average, every 1.5 days for three weeks straight. Not previews. Not research papers. Downloadable weights, working inference code, and production-ready APIs.
Three structural forces converged simultaneously. First, the MoE architecture — pioneered in the open ecosystem by models like GLM and Qwen — reached mainstream adoption, enabling frontier-class capability at a fraction of the compute cost. Second, Apache 2.0 licensing became the default competitive move: Google, Alibaba, and Ai2 all shipped under fully permissive terms, forcing the rest of the field to respond. Third, Chinese AI labs — Z.ai, MiniMax, Moonshot AI, Alibaba Qwen — proved capable of iterating at the same speed as Western frontier labs while pricing at 5–50x lower per-token costs.
The 2026 Databricks survey found that 75%+ of enterprises now use two or more LLM families in production. April 2026 is the month that preference became a genuine technical argument, not just a cost hedge.
Week 1 — April 1–7: The Opening Salvo
The month opened with three coordinated frontier drops in five days that collectively rewrote what “open source” means at the capability tier.
🔷 Gemma 4 — April 2 (Google, Apache 2.0)
Google’s most important open model release since the original Gemma. Four variants from E2B (~2B effective parameters, runs offline on phones) to 31B Dense, all under Apache 2.0 — the most permissive license any Google AI model has ever carried. Built from the same research as Gemini 3, Gemma 4’s 31B Dense outperforms Llama 4 Maverick on AIME 2026 Math (89.2% vs 88.3%), LiveCodeBench v6 (80.0% vs 77.1%), and GPQA Diamond (84.3% vs 82.3%), despite having 13x fewer total parameters. Context window up to 256K tokens. Day-0 vLLM, transformers, llama.cpp, and Google Cloud TPU support. And the E2B/E4B edge variants run completely offline on a phone with near-zero latency — the first genuinely capable on-device models from a major Western lab. 400M+ Gemma downloads to date; this version opened the ecosystem to anyone who previously couldn’t navigate Llama’s user cap restrictions.
📦 OLMo 2 32B — April 3 (Ai2, Apache 2.0)
The Allen Institute for AI shipped OLMo 2 32B — the only frontier-scale model in this entire wave where the training data, training code, and evaluation pipeline are completely open alongside the weights. For researchers who need full reproducibility (not just usable weights), OLMo 2 32B under Apache 2.0 is the only choice. It’s not the highest performer in this list, but it’s the most transparently built model at anything approaching this size.
🦙 Llama 4 Scout + Maverick — April 5 (Meta, Custom License)
Meta’s first MoE model family and first natively multimodal release. Scout (109B total, 17B active, 16 experts) and Maverick (400B total, 17B active, 128 experts) both share a 17B active parameter count at inference, keeping compute costs far below their total sizes. Scout fits on a single H100 with int4 quantization — and consumer hardware with the right GPU. Scout’s 10M token context window is the largest of any open-weight model. Maverick positions directly against GPT-4o and Gemini 2.0 Flash. Both are natively multimodal across text, images, and video. The custom Llama community license restricts companies over 700M MAUs and EU-domiciled deployments — a caveat that matters for enterprise evaluation.
🤖 Qwen 3 72B — April 5 (Alibaba, Apache 2.0)
Alibaba’s top dense model in its Q2 open-source push — 72B parameters, Apache 2.0, with reasoning and coding performance that exceeded GPT-4o on MMLU-Pro. Supports both a “thinking” mode for harder problems and a fast mode for everyday tasks. The first of several Qwen releases this month, with more to come.
🐉 GLM-5.1 (Z.ai, MIT) — April 7
The shot heard round the developer world. Z.ai dropped open weights for GLM-5.1 — 754B MoE (40B active), MIT license, trained entirely on Huawei Ascend 910B chips — with a 58.4 SWE-Bench Pro score that edged past GPT-5.4 (57.7) and Claude Opus 4.6 (57.3). It was the first time an open-source model held the #1 spot on the most respected real-world coding benchmark. The MIT license means download, modify, self-host, and commercialize with zero restrictions. At $1.00/$3.20 per million tokens via API — 5–8x cheaper than Claude — the cost argument became impossible to ignore. The geopolitical footnote: China’s top AI lab, on the US Entity List, produced a frontier-class model without a single Nvidia chip. That’s not a technical curiosity. That’s a strategic data point for every government assessing AI sovereignty.
Week 2 — April 8–14: Infrastructure Catches Up
The second week delivered the tooling and frameworks needed to actually deploy and orchestrate everything that dropped in week one. This is the layer most press coverage misses — and it’s arguably what makes the models usable at scale.
🤖 Qwen 3 MoE 235B — April 8 (Alibaba, Apache 2.0)
Alibaba followed up the 72B Dense with the heavier MoE variant — 235B total, 22B active, Apache 2.0. Near-frontier performance at low active parameter cost. With the 72B Dense already available, teams could now choose between a compact reasoning model (72B) and a cost-efficient frontier-adjacent model (235B MoE) from the same family, under the same license, with the same ecosystem support.
💻 Codestral 2 — April 8 (Mistral, Apache 2.0)
Mistral shipped their second-generation code model — 22B parameters, Apache 2.0, optimized for fill-in-the-middle and code generation workflows. The Apache 2.0 license is a significant upgrade from prior Mistral code models that carried commercial restrictions. For teams building coding assistants or CI/CD integration tools, Codestral 2 became the cleanest commercial-use option in its size class.
🔍 Haystack 2.9.0 — April 8 (deepset, Apache 2.0)
deepset’s production RAG framework shipped multi-modal pipeline support as a first-class feature. Images, audio, and documents as native pipeline inputs — not text-only workarounds. Given that every model in week one has multimodal capabilities, this was the critical piece that made those capabilities usable in structured pipelines without custom adapter code.
📱 Gemma 3n — April 9 (Google, Gemma License)
Google’s next-generation on-device model — 4B effective parameters, 2B memory footprint — built specifically for phones and tablets. Ships via Google AI Edge Gallery, a mobile app that runs Gemma models directly on device hardware. No API key. No network dependency. No data leaving the device. Go into airplane mode and your AI still works. The 128K context window and native system role support turn a modern phone into a self-contained AI workstation for the first time at this capability level.
🤝 CrewAI 0.9.1 — April 9 (MIT)
The most widely adopted multi-agent orchestration framework in the Python ecosystem shipped native support for OpenRouter, DeepSeek, Ollama, vLLM, Cerebras, and Dashscope as first-class providers. No custom adapters. Direct configuration. Security patches across three dependencies. The framework now covers virtually every model released in this wave without integration boilerplate.
⚙️ vLLM v0.18 + v0.19 — April 10 (Apache 2.0)
Two major releases in the same cycle. The infrastructure that powers a significant fraction of production open-source LLM deployments globally shipped: gRPC serving (direct integration, no proxy layer), GPU-accelerated speculative decoding with zero-bubble overlap, full Gemma 4 day-0 support across all four variants, Model Runner V2 maturation, and a critical security patch (CVE-2026-0994). For MiniMax M2.7 specifically, NVIDIA and vLLM collaboratively developed QK RMS Norm and FP8 MoE kernel optimizations that deliver up to 2.5x throughput on Blackwell Ultra GPUs. Hardware support expanded to Intel XPU, AMD ROCm 7.2.1, ARM CPU, and Huawei Ascend NPU.
🤖 MiniMax M2.7 — April 11 (Research License ⚠️)
MiniMax’s self-evolving agentic model — 230B total, 10B active, the smallest active-parameter footprint in the Tier-1 coding class. During training, M2.7 ran 100+ rounds of autonomous scaffold optimization, analyzing its own failure trajectories and improving its agent harness without human prompt engineering, achieving a 30% performance gain. On GDPval-AA (economic knowledge work), M2.7 scored ELO 1495 — highest among all open-weight models, ahead of GPT-5.3. On the MM Claw end-to-end benchmark, 62.7% — approaching Claude Sonnet 4.6 levels. API priced at $0.30/$1.20 per million tokens — roughly 50x cheaper than Claude Opus on input. Critical caveat: the weights prohibit commercial use without MiniMax authorization. Not OSI-certified open source. Research and evaluation use is permitted; production deployment requires a separate agreement.
💻 Open Interpreter 0.5.3 — April 12 (MIT)
Improved sandboxed execution with better host isolation, and expanded multi-model routing letting different open models handle different task types in the same interactive session. With Llama 4 Scout running on consumer hardware and Open Interpreter 0.5.3 providing the execution layer, a fully local AI coding assistant became practical on hardware most developers already own.
🌙 Kimi K2.6 Code Preview — April 13 (Moonshot AI)
Beta testers started running K2.6 Code Preview on April 13, with Moonshot quietly confirming access by email. Early partner results were striking: Vercel saw 50%+ improvement on their Next.js benchmark vs K2.5. Factory.ai reported +15% on internal benchmarks. CodeBuddy: +12% code accuracy, +18% long-context stability, 96.6% tool success rate. The GA release was just days away.
🔧 llama.cpp b8779 — April 13
The backbone of local LLM inference for hundreds of thousands of developers pushed a meaningful technical update: a Vulkan flash attention DP4A shader for quantized KV cache — enabling efficient quantized attention on AMD, Intel, and mobile GPUs without CUDA dependency. For the large population running open models on non-Nvidia hardware, this is a meaningful inference performance improvement.
Week 3 — April 15–22: The Second Wave
If week one was about models and week two was about infrastructure, week three was about efficiency breakthroughs and agentic endurance — two themes that are redefining what the open-source tier can actually do in production.
🧠 Qwen 3.6-35B-A3B — April 16 (Alibaba, Apache 2.0)
The most technically surprising release of the entire month. Alibaba’s Qwen team shipped a 35B parameter MoE model that activates only 3B parameters per query — a 90% compute reduction versus models it outperforms. On SWE-Bench Verified: 73.4%, beating Gemma 4 31B by 21.4 points and matching Claude Opus 4.6 territory. On Terminal-Bench 2.0: 51.5% vs Gemma 4 31B’s 42.9%. The model is natively multimodal (text, image, video), supports 224K video tokens, scored 83.7% on VideoMMU, and runs locally on consumer hardware at approximately 21GB quantized. Apache 2.0 licensed.
The efficiency story is the real headline: 256 total expert sub-networks, only 9 activated per layer, producing 3B active parameters from a 35B model. That’s compute efficiency at a level that makes the “you need a server farm” objection to open-source models increasingly obsolete. A developer with an RTX 4090 can now run a model that scores in Claude Opus 4.6 territory on coding benchmarks.
🌙 Kimi K2.6 GA — April 20 (Moonshot AI, Modified MIT)
Eight days after the preview, Moonshot AI shipped Kimi K2.6 as generally available — and immediately claimed the new SWE-Bench Pro #1 with 58.6 (beating GLM-5.1’s 58.4 that had held the top spot since April 7). This is a 1-trillion-parameter MoE model with 32B active parameters, 384 experts, built specifically around one conviction: stamina. The ability to run autonomously for hours, across thousands of tool calls, without drifting, stalling, or looping.
Two showcase runs published alongside the release: K2.6 implemented local LLM inference from scratch in Zig (a niche systems language) over 12 hours and 4,000+ tool calls, boosting throughput from ~15 to ~193 tokens per second — 20% faster than LM Studio. And given an 8-year-old financial matching engine near its limits, K2.6 worked for 13 hours, modified 4,000+ lines of code, and extracted a 185% median throughput gain. These are not demos. These are the kinds of engineering runs that define what “autonomous AI” actually means in production.
On benchmarks: 58.6 on SWE-Bench Pro (new #1), 80.2 on SWE-Bench Verified, 89.6 on LiveCodeBench v6, and 54.0 on Humanity’s Last Exam with tools — leading every model in that comparison including GPT-5.4 (52.1), Claude Opus 4.6 (53.0), and Gemini 3.1 Pro (51.4). The Agent Swarm mode scales to 300 sub-agents coordinating across 4,000 steps. “Claw Groups” enables heterogeneous swarms where humans and agents on any device collaborate in a shared operational space. Available on Kimi.com, the API, Kimi Code CLI, and Ollama. Weights on HuggingFace under a Modified MIT License.
🔧 Google ADK Python 1.0 — April 21 (Apache 2.0)
Google’s Agent Development Kit hit 1.0 — a modular open-source framework for building and deploying AI agents, natively compatible with Gemma 4’s function calling, structured output, and reasoning capabilities. The 1.0 release hit 8,200+ GitHub stars quickly, signaling that the framework ecosystem for Gemma 4 is maturing rapidly. For teams building agentic Gemma 4 workflows on Google Cloud or locally, ADK Python 1.0 is the official orchestration layer.
Full Release Table: April 1–22, 2026
| Date | Release | Org | License | Key Highlight |
|---|---|---|---|---|
| Apr 2 | Gemma 4 (E2B/E4B/26B/31B) | Apache 2.0 ✅ | 31B beats Llama 4 on math & coding; offline phone/edge support | |
| Apr 3 | OLMo 2 32B | Ai2 | Apache 2.0 ✅ | Only fully open training data, code & weights at this scale |
| Apr 5 | Llama 4 Scout + Maverick | Meta | Custom ⚠️ | First native multimodal MoE; 10M context (Scout); consumer GPU viable |
| Apr 5 | Qwen 3 72B | Alibaba | Apache 2.0 ✅ | Top dense model on reasoning; thinking + fast modes |
| Apr 7 | GLM-5.1 | Z.ai | MIT ✅ | #1 SWE-Bench Pro (58.4); no Nvidia chips; $1/$3.20 per 1M tokens |
| Apr 8 | Qwen 3 MoE 235B | Alibaba | Apache 2.0 ✅ | 235B/22A MoE; near-frontier at low active param cost |
| Apr 8 | Codestral 2 | Mistral | Apache 2.0 ✅ | 22B code model; fill-in-the-middle; unrestricted commercial use |
| Apr 8 | Haystack 2.9.0 | deepset | Apache 2.0 ✅ | First-class multi-modal RAG pipeline support |
| Apr 9 | Gemma 3n | Gemma ⚠️ | 4B effective / 2B footprint; fully offline on phones; 128K context | |
| Apr 9 | CrewAI 0.9.1 | CrewAI | MIT ✅ | Native vLLM, OpenRouter, DeepSeek, Ollama providers; security fixes |
| Apr 10 | vLLM v0.18 + v0.19 | vLLM Project | Apache 2.0 ✅ | gRPC, zero-bubble spec decode, Gemma 4 day-0, MRV2, CVE patch |
| Apr 11 | MiniMax M2.7 | MiniMax | Research ⚠️ | Self-evolving; 230B/10B active; #1 GDPval-AA open-weight; $0.30/$1.20 |
| Apr 12 | Open Interpreter 0.5.3 | OI | MIT ✅ | Sandboxed execution; multi-model routing; local AI coding stack |
| Apr 13 | llama.cpp b8779 | ggml | MIT ✅ | Vulkan flash attention DP4A; faster inference on AMD/Intel GPUs |
| Apr 16 | Qwen 3.6-35B-A3B | Alibaba | Apache 2.0 ✅ | 73.4% SWE-Bench Verified; 3B active / 35B total; runs on RTX 4090 |
| Apr 20 | Kimi K2.6 GA | Moonshot AI | Modified MIT ⚠️ | New #1 SWE-Bench Pro (58.6); 1T/32B MoE; 300-agent swarms; 12-hr runs |
| Apr 21 | Google ADK Python 1.0 | Apache 2.0 ✅ | Official agentic framework for Gemma 4; 8,200+ GitHub stars |
What It All Means
Three weeks. Seventeen tracked releases. The open-source AI narrative didn’t just evolve in April 2026 — it inverted.
Apache 2.0 won the licensing war. Google, Alibaba, Ai2, and Mistral all shipped under fully permissive terms. The old pattern — restrict commercial use, add user-count caps, impose data residency obligations — is now the exception rather than the rule at the frontier-adjacent tier. For enterprises that couldn’t touch Llama due to the 700M MAU cap, or couldn’t use closed models for GDPR reasons, April 2026 opened the door.
MoE architecture is the new baseline. Every major model in this wave — Llama 4, GLM-5.1, Qwen 3.6, MiniMax M2.7, Kimi K2.6 — uses Mixture-of-Experts. The result: 35B-parameter models performing at 100B+ levels, 400B-class models running on a single GPU, trillion-parameter models pricing at $0.30 per million input tokens. The parameter count arms race is over. The architecture efficiency race is what matters now.
Agentic endurance became the new battleground. The benchmark that mattered most in week three wasn’t a single-turn eval. It was: how long can your model work autonomously on a real engineering problem without drifting, looping, or giving up? GLM-5.1 claimed 8 hours. Kimi K2.6 demonstrated 12–13. MiniMax M2.7 trained itself over 100+ autonomous rounds. This is the capability that turns an AI assistant into infrastructure — and it’s now arriving at open-weight prices.
What the closed models still hold: Claude Opus 4.7 and GPT-5.4 retain meaningful advantages in safety alignment, highest-stakes one-shot reasoning, multimodal quality at the highest resolution, and enterprise support guarantees. The open-source ecosystem has closed the gap on most tasks that appear in developer and enterprise workloads. But “most tasks” is not “all tasks” — and for organizations where AI reliability and alignment are non-negotiable, the closed-model tier still has a defensible position.
The practical read for developers and enterprises is simple: you now have genuine options at every point of the cost-capability curve. April 2026 is the month that menu became complete.
🚀 Read Our Individual Deep-Dives
We’ve published full reviews on the biggest individual launches from this wave.
Browse All Articles →GLM-5.1 · Claude Opus 4.7 · Claude Design · OpenAI Codex Update
❓ Frequently Asked Questions
Which open-source model released in April 2026 currently holds SWE-Bench Pro #1?
As of April 22, 2026, Kimi K2.6 (Moonshot AI) holds the top SWE-Bench Pro score at 58.6, released April 20. It edged out GLM-5.1 (58.4, April 7) which previously held the top spot.
Which April 2026 open-source models use Apache 2.0?
Gemma 4, OLMo 2 32B, Qwen 3 72B, Qwen 3 MoE 235B, Qwen 3.6-35B-A3B, Codestral 2, Haystack 2.9.0, CrewAI 0.9.1, vLLM v0.18/v0.19, and Google ADK Python 1.0 all use Apache 2.0 — fully permissive, no commercial restrictions.
What makes Qwen 3.6-35B-A3B significant?
It activates only 3B parameters per query from a 35B model — a 90% compute reduction — while scoring 73.4% on SWE-Bench Verified, beating Gemma 4 31B by over 21 points. At approximately 21GB quantized, it runs on consumer hardware like an RTX 4090 under Apache 2.0, with full multimodal support including video.
What is Kimi K2.6’s most notable capability?
Long-horizon autonomous execution — the ability to work on a real engineering task for 12+ hours across thousands of tool calls without human intervention. In showcase runs, K2.6 implemented a Zig inference engine over 12 hours (achieving 20% faster throughput than LM Studio) and overhauled a financial matching engine over 13 hours, achieving a 185% throughput gain.
Do any open-source models from this wave now match closed frontier models?
On coding and agentic benchmarks specifically, yes — Kimi K2.6 (58.6) and GLM-5.1 (58.4) both lead or tie GPT-5.4 (57.7) on SWE-Bench Pro. Qwen 3.6 leads on SWE-Bench Verified. On general reasoning, safety alignment, and highest-resolution multimodal tasks, closed models like Claude Opus 4.7 retain meaningful advantages.
Latest Articles
Browse our comprehensive AI tool reviews and productivity guides
Musk v. OpenAI Trial: The Case That Could Reshape the Entire AI...
Musk v. OpenAI Trial: The Case That Could Reshape the Entire AI Industry
Musk called himself "a fool" on the stand. Altman appeared by prerecorded video from AWS while being sued. The judge reprimanded both sides. And the AI industry's most consequential legal battle is just getting started.
Big Tech Q1 2026 Earnings: The $665 Billion AI Bet — Winners,...
Big Tech Q1 2026 Earnings: The $665 Billion AI Bet — Winners, Losers, and What It Means
Five tech giants reported Q1 2026 earnings in 48 hours. Combined AI capex: $665 billion — 75% more than 2025. Alphabet and Amazon won. Meta spooked investors. Here's every number that matters.
AI Is Replacing Developers — The Real Numbers (2026)
AI Is Replacing Developers — The Real Numbers (2026)
Snap fired 1,000. Google generates 75% of new code with AI. Entry-level developer jobs fell 20%. But 1.3M new AI roles were created and India's AI hiring surged 59.5%. Here's what's actually happening.
I Used Claude Free for 3 Months Instead of ChatGPT and Gemini...
I Used Claude Free for 3 Months Instead of ChatGPT and Gemini — Here’s What Happened
I launched and grew NivaaLabs on Claude's free tier for 3 months. I also used ChatGPT and Gemini. Here's the honest, task-by-task breakdown of what each AI actually does well — and which one I'd recommend for someone building something real on a $0 budget.
Runway Gen-3 Turbo: Real-Time Video Tested (2026)
Runway Gen-3 Turbo: Real-Time Video Tested (2026)
Runway Gen-3 Turbo's real-time video generation capabilities are put to the test, examining quality, speed, and value.
Best AI Coding Tools 2026: Every Major Tool Ranked — Cursor, Claude...
Best AI Coding Tools 2026: Every Major Tool Ranked — Cursor, Claude Code, Copilot, Windsurf & More
85% of developers now use AI coding tools daily. AI writes 46% of all new code. The market has 10+ serious tools and most developers end up using two or three. Here's how every major AI coding tool in 2026 ranks — with real benchmark data, honest pricing, and a verdict for every workflow type.
DeepSeek V4 Review: V4 Flash & V4 Pro — Almost Frontier, a...
DeepSeek V4 Review: V4 Flash & V4 Pro — Almost Frontier, a Fraction of the Price (April 2026)
DeepSeek V4 arrived April 24, 2026 — one year after R1 shook Silicon Valley. V4 Pro is the world's largest open-weight model at 1.6 trillion parameters. V4 Flash is cheaper than GPT-5.4 Nano. And both run on Chinese chips. Here's everything you need to know.
GPT-5.5 vs Claude Opus 4.6 (2026): Which AI Model Wins for Your...
GPT-5.5 vs Claude Opus 4.6 (2026): Which AI Model Wins for Your Work?
OpenAI's GPT-5.5 arrived April 23 claiming to be the smartest model yet. Anthropic's Claude Opus 4.6 still holds the top Chatbot Arena ELO. Both cost real money. Which one actually wins for your workflow? Here's the full data-driven comparison.
GPT-5.5 Review: OpenAI’s Smartest Model Yet — Agentic Coding, Computer Use &...
GPT-5.5 Review: OpenAI’s Smartest Model Yet — Agentic Coding, Computer Use & More (April 2026)
GPT-5.5 landed April 23 — seven weeks after 5.4. OpenAI calls it a "new class of intelligence for real work." It's faster per token, stronger at agentic coding, computer use, and scientific research, and comes with the strongest safety guardrails yet. Here's everything you need to know.
Project Glasswing: Anthropic’s “Too Dangerous to Release” AI and the Cybersecurity Reckoning
Project Glasswing: Anthropic’s “Too Dangerous to Release” AI and the Cybersecurity Reckoning
Anthropic built an AI so capable at hacking that they won't release it publicly. Claude Mythos Preview found a 27-year-old OpenBSD zero-day for under $50. Project Glasswing is what happens next.
Google Cloud Next 2026: Every Major Announcement — Agents, TPU 8, Virgo...
Google Cloud Next 2026: Every Major Announcement — Agents, TPU 8, Virgo Network & More
Google Cloud Next 2026 just happened. Here's everything: new 8th-gen TPUs, the Gemini Enterprise Agent Platform, A2A protocol in production at 150 orgs, Workspace Studio for no-code agents, and a $185B infrastructure bet. One article, all the details.
OpenAI ChatGPT Ads Review: The $100 Billion Bet That’s Already Getting Messy
OpenAI ChatGPT Ads Review: The $100 Billion Bet That’s Already Getting Messy
OpenAI launched ads in ChatGPT, faced user backlash, fired back at Anthropic's Super Bowl jabs, and just flipped to cost-per-click pricing. Here's what's actually happening.