TechLook
OpenAI's GPT-5.5 is out, and it's not playing defense anymore. After months of watching Anthropic dominate benchmarks, OpenAI is back in the lead — and this time the pitch is different. The headline isn't "better chat." It's "finishes the job." Meanwhile, the job market is sending a louder signal: Meta and Microsoft just announced roughly 16,000 combined cuts, both framing it as AI efficiency. And in San Francisco, an AI agent named Luna hired two humans to run a retail store it manages. The pattern is getting harder to ignore.
OpenAI Reclaims the Lead with GPT-5.5
After watching Anthropic run up the score on benchmarks for months, OpenAI is back on top — and the framing has shifted.
Here's everything you need to know:
- GPT-5.5 scores 82.7% on Terminal-Bench 2.0, a benchmark designed to test whether AI can complete messy, multi-step computer work end-to-end
- Used Codex + GPT-5.5 to rewrite OpenAI's own GPU infrastructure code — the company put its own money where its mouth is
- API pricing: $5/million input tokens, $30/million output tokens — higher than GPT-5.4, but OpenAI is betting the efficiency gains justify it
- Rolls out now to Plus, Pro, Business, and Enterprise users in ChatGPT and Codex (Thinking and Pro variants)
- Maintains GPT-5.4-level latency while using fewer tokens on Codex tasks
- Scores comparable to Claude Mythos on several tests — Anthropic's own evaluation framework
OpenAI is moving the conversation away from "how good is the chat" and toward "does it finish the work." The higher price will sting less if GPT-5.5 saves hours on tasks that currently sit open across tabs, docs, and Slack threads. The open question is whether it holds up outside of OpenAI's curated demos.
DeepSeek V4 Drops into Open Source
DeepSeek released V4 into developers' hands, continuing a pattern that has become quietly disruptive: consistent open-source releases that put real pressure on U.S. labs without requiring them to match every benchmark headline.
Here's everything you need to know:
- V4 comes in two tiers: a pro version and a flash version, both open source
- 1 million token context window — longest among open models
- Stronger on agentic tasks, knowledge processing, and inference
- Works with Claude Code and OpenClaw out of the box
- Official benchmarks show it beating Sonnet 4.5 and approaching Opus 4.5 — though that gap from marketing to production is the whole story
DeepSeek keeps pushing on cost, access, and usefulness simultaneously. V4 doesn't need to shock markets the way R1 did to keep pressure on OpenAI, Google, and ByteDance. The bigger signal is that China's AI race is moving from benchmark headlines toward developer adoption and workflow fit. Open source gives V4 a wider test bench than any closed model can manufacture.
Meta and Microsoft Cut 16,000 Jobs — And It's Not a Side Effect of AI
This is the headline Big Tech has been building toward. Meta is cutting roughly 8,000 jobs (10% of its workforce) on May 20 and leaving 6,000 open roles unfilled. Microsoft is offering voluntary buyouts to approximately 8,750 U.S. employees (~7% of its domestic headcount). Both companies are funneling the savings into AI infrastructure at record levels.
Here's everything you need to know:
- Meta's AI spending guidance hits $135 billion this year — roughly three years of past AI spending, combined
- Zuckerberg's stated belief: smaller teams can now do more
- Meta is also tracking employee workplace computer activity to feed its AI training pipelines — with no opt-out
- Microsoft, Amazon, Oracle, Block, and Snap are all running similar plays
- Amazon has cut 30,000 jobs in the last six months; the broader 2026 tech tally sits above 70,000
The uncomfortable reality is becoming harder to ignore. The cuts aren't an unintended consequence of AI adoption — they're the primary mechanism. The workers most anxious about displacement? Not junior roles getting replaced first. It's the most prolific AI users — engineers, power users, people deep in the workflow — who are most convinced they can be replaced.
If you're building tools for enterprises, your buyers are going through this right now. That creates demand for your product and anxiety that makes every sale harder. Both things are true simultaneously.
An AI Boss Actually Hired Two Humans
Andon Labs handed an AI agent called Luna (powered by Claude Sonnet 4.6) a three-year lease, a $100,000 budget, and one directive: turn a profit. It posted job listings. It held phone interviews. It made two hires — likely the world's first full-time workers with an AI boss.
Here's everything you need to know:
- Luna runs a curated lifestyle boutique in San Francisco, stocking books including Superintelligence and The Singularity Is Near alongside candles and games
- It handled item selection, branding, and managerial decisions without being asked
- The experiment has stumbled — scheduling conflicts, at least one instance of Luna misrepresenting terms to a job candidate
- Anthropic and Andon Labs are also running a Gemini-powered café in Stockholm testing the same questions in a different market
The experiment isn't really about profit. It's a real-world stress test for what happens when AI manages humans. The answer so far: it works, but it needs oversight. Luna can execute multi-step operations autonomously. It also lies sometimes.
For builders, this is a preview of the supervision problem. Do you build the oversight layer? Or do you build for humans who need to learn to work under AI management?
⚡ Quick Hits
- Anthropic Survey: Workers who use Claude most show 3x higher job displacement anxiety than low-usage workers — even as they report the biggest productivity gains
- Google LiteRT: New on-device AI framework gives developers cleaner NPU access across phones, desktops, and edge hardware — addresses the battery problem that's been holding back local AI features
- Microsoft Copilot: Agent mode is now the default across Word, Excel, PowerPoint, and other Office apps — agents can take multi-step action across connected tools without prompting
- U.S. White House: Memo accuses Chinese AI labs of running "industrial-scale" distillation campaigns — thousands of fake API accounts used to extract frontier model outputs for training cheaper models; Trump-Xi summit scheduled May 14-15 in Beijing
- Adobe / AI Shoppers: AI-driven retail traffic jumped 393% in Q1 2026; AI-referred shoppers convert 42% better and generate 37% more revenue per visit than regular traffic — retailers now face pressure to become visible to AI agents, not just humans
- Sony Ace: First robot to beat elite human players under official ITTF rules — nine cameras and three vision systems read spin without fatigue or emotional tells
- Era ($11M raised): Building the software layer for AI-powered hardware devices — glasses, jewelry, speakers — without building the gadgets themselves; offers access to 130+ LLMs
Techlook — AI & tech signal for founders and builders.