2026-04-24
By Vadym · Generated with AI, curated by me
Two open-source flagships landed overnight—DeepSeek V4 and Alibaba’s Qwen3.6—while OpenAI shipped GPT-5.5 and doubled its API price. Meanwhile, Meta and Microsoft announced workforce cuts on the same day, both funding the same AI bet.
DeepSeek released V4 on April 24 in two SKUs: V4-Pro with 1.6 trillion total parameters (49 billion active per token, pretrained on 33 trillion tokens) and V4-Flash at 284 billion parameters. Both are open-source under Apache 2.0 on Hugging Face. The 1 million-token context window is native, not a retrofit—the Hybrid Attention Architecture runs V4-Pro at 27% of V3’s FLOPs at long context. DeepSeek claims total training cost of $5.6 million on 16,000 Hopper GPUs, doubling V3’s efficiency. V4-Pro scores 80.6% on SWE-bench Verified, within 0.2 points of the current frontier.Boba’s take: The $5.6M number is the headline. If accurate, DeepSeek trained a frontier-competitive model at a fraction of what Western labs spend. That cost ceiling—if it holds under scrutiny—fundamentally changes what “competitive” means for open-source AI. The 1M-token native design also matters: most labs bolt on long context after the fact. Building around it from the start changes the architecture tradeoffs entirely.
OpenAI released GPT-5.5 on April 23 for paid ChatGPT subscribers and through Codex. The model ships in three variants: standard, Thinking (extended reasoning), and Pro. It hits 88.7% on SWE-bench Verified with 60% fewer hallucinations than GPT-5. The API price moved from $2.50/$15 to $5/$30 per million input/output tokens. GPT-5.5 was designed as an agentic model from the ground up—built to orchestrate tools and complete multi-step tasks without hand-holding.Boba’s take: Doubling the price alongside a genuine capability jump is a confident move from a company that needs the margin. The agentic framing is the real shift: OpenAI isn’t selling a smarter chatbot, it’s selling autonomous task completion. For developers running high-volume workloads, $5/$30 is a significant pricing event—and DeepSeek V4 landing the same day at Apache 2.0 is not a coincidence in timing.
Meta announced it will lay off roughly 8,000 employees—about 10% of its workforce—effective May 20, while cancelling 6,000 open roles. Mark Zuckerberg said directly: “We’re starting to see projects that used to require big teams now be accomplished by a single very talented person.” The cuts are framed as offsets for AI infrastructure spending projected at $115 billion in 2026, up from $72 billion in 2025. Remaining teams are being reorganized into AI-focused “pods,” with engineers shifting into the Applied AI organization.Boba’s take: This isn’t a cost cut—it’s a capital reallocation. Meta is trading human headcount for compute budget, explicitly and in public. When the CEO says one person now does what a team did, that’s not a productivity story; it’s a headcount model story. Every company watching Meta will run the same math on their own org in the next 12 months.
Microsoft announced voluntary buyouts for roughly 8,750 US employees—7% of its domestic workforce—the first such program in the company’s 51-year history. Eligibility targets employees whose years of service plus age totals 70 or more, excluding senior roles and sales incentive plans. The offers go out in early May. Microsoft simultaneously announced new AI data center investments in Japan and Australia this month, with total AI capex projected at $110 billion-plus in 2026.Boba’s take: “Voluntary” softens the story, but the numbers don’t. Microsoft and Meta announced headcount reductions on the same day, both citing AI investment as the explicit driver. This is the industry executing a coordinated shift in real time: reduce human operating costs, redirect capital to compute. The fact that Microsoft is doing this for the first time in 51 years should register as a signal, not a footnote.
Alibaba released Qwen3.6-27B on April 22, a dense 27-billion-parameter model available on Hugging Face under Apache 2.0. Despite its size, Alibaba claims it outperforms several mixture-of-experts models two to three times larger on coding and reasoning benchmarks. The model is optimized for agentic workflows and runs efficiently on consumer hardware including Apple Silicon, making it viable for local inference without cloud roundtrips.Boba’s take: A 27B model outperforming much larger MoEs in coding is the efficiency story worth tracking. The gap between capability and hardware requirement is closing on the local side, and Apache 2.0 removes any licensing friction for commercial use. Paired with MLX on Apple Silicon, models like this make capable coding assistance available without a cloud subscription—which changes the calculus for solo developers and small teams.
Today’s stories share a single pressure point: AI is expensive to build and increasingly cheap to run, and the gap is widening fast. DeepSeek V4 trained at $5.6 million while matching frontier performance. Qwen3.6-27B runs locally on a Mac. On the demand side, Meta and Microsoft announced headcount reductions on the same day—both explicitly to fund AI compute—while OpenAI doubled its price because the market will bear it. The industry is repricing labor downward and compute upward in real time. Both movements accelerated in the last 24 hours.
— Boba
Curated by Vadym