OpenAI Ships GPT-5.5, DeepSeek Drops V4, Meta Cuts 8,000 Jobs for AI
OpenAI Ships GPT-5.5, DeepSeek Drops V4, Meta Cuts 8,000 Jobs for AI
This Week in AI: April 21–27, 2026
This week the gap between labs shipping models and labs shipping infrastructure effectively closed. Three events on the same day made the picture clear: OpenAI's GPT-5.5 landed with hard benchmark numbers, DeepSeek released its first major open-source upgrade in a year, and Meta announced 8,000 layoffs to fund a near-doubling of AI capex. The model race is now indistinguishable from the infrastructure race.
Key Takeaways
- OpenAI: GPT-5.5 scores 88.7% on SWE-bench and ships with 60% fewer hallucinations than GPT-5.4. If you're running evals, update your baselines now.
- DeepSeek: V4 Flash and V4 Pro introduce a Hybrid Attention Architecture with a 1-million-token context window. The best open-source option just got meaningfully more capable.
- Meta: 8,000 layoffs (10% of workforce) announced alongside capex guidance of $115–135B for 2026, nearly double 2025's spend. Meta is no longer hedging on AI.
- Google + Apple: Gemini confirmed to power Apple Intelligence and a redesigned Siri. One deal just handed Google distribution across over a billion Apple devices.
- Tencent: Hy3 Preview (295B parameters, open-source) dropped on the same day as DeepSeek V4. Chinese labs are now competing on release cadence, not just capability.
The Model Sprint
OpenAI Releases GPT-5.5
OpenAI announced GPT-5.5 on April 23, describing it as better at multi-step agentic work: coding, computer use, early-stage scientific research, and document creation. The model scores 88.7% on SWE-bench and 92.4% on MMLU, with OpenAI reporting 60% fewer hallucinations compared to GPT-5.4. API pricing starts at $5 per million input tokens, with a Pro variant at $30 per million. The model is live now for Plus, Pro, Business, and Enterprise users in ChatGPT and Codex.
DeepSeek Drops V4 Flash and V4 Pro
One day later, DeepSeek unveiled V4 Flash and V4 Pro, exactly one year after the V3 release that upended Silicon Valley. The new models use a Hybrid Attention Architecture designed to improve retention across long conversations, and both support a 1-million-token context window with claimed top-tier coding benchmark performance. DeepSeek is also raising external funding for the first time, targeting $300M at a $10B+ valuation, with Tencent and Alibaba reportedly in talks for a stake.
Infrastructure and Alliances
Meta Cuts 8,000 Jobs and Doubles AI Spend
Meta announced 8,000 layoffs on April 23, representing 10% of its workforce, with cuts starting May 20. The company is also closing 6,000 open roles. Its 2026 capex guidance is now $115–135B, nearly double the $72B it spent in 2025. Zuckerberg framed the savings as direct inputs into AI infrastructure and next-generation LLM development. Meta's custom MTIA chips are already deployed in data centers, with the 450 and 500 series targeted for mass rollout by 2027.
Google Lands the Apple Intelligence Deal
Google Cloud's CEO confirmed that Gemini will power Apple Intelligence and a redesigned Siri, running on Apple's Private Cloud Compute. For builders, this means Gemini is now the default reasoning layer across iOS and macOS at scale. Apple offloads model development; Google gains the distribution it could not win on its own.
Tencent Ships Hy3 Preview
Tencent launched Hy3 Preview on April 24, a 295-billion-parameter open-source model now running in its Yuanbao chatbot, replacing DeepSeek. The timing, releasing the same day as DeepSeek V4, signals that Chinese AI labs are treating weekly model drops as a deliberate competitive strategy, not just a product roadmap.
Quick Hits
- Cursor raises $2B: The AI coding tool is in talks to raise $2 billion at a $50B+ valuation. Developer tooling built on foundation models is now valued at frontier-lab scale.
- Anthropic at $800B: Investors have reportedly offered valuations of $800 billion, up from $350B in February, alongside the company's $100B, 10-year AWS commitment.
- AI energy efficiency: Researchers published results showing a 100x reduction in AI energy use by combining neural networks with symbolic reasoning. Still early-stage, but relevant if inference costs are on your roadmap.
- US AI legislation: 78 chatbot-related bills are currently active across 27 states. Consumer AI products shipping in the US now have a material compliance surface to monitor.
- Terafab + Intel: Elon Musk's hardware venture Terafab announced an AI chip partnership with Intel, positioning itself as a potential challenger to NVIDIA's data center dominance.
This Week in AI is published every Monday by the Lablab team.
Software Developer