Google Loses 5 AI Researchers in a Week as GPT-5.6 Slips to July

Good morning. It's Friday, June 26, and we're covering Google's accelerating talent departure, a GPT-5.6 launch that slipped out of June, OpenCode crossing a threshold that signals the scale of open-source AI tooling adoption, and DeepSeek's permanent frontier-model price reset.

Five senior Google AI researchers announced departures for Anthropic and OpenAI this week. The Bloomberg piece broke Wednesday and drew immediate discussion - three went to Anthropic, one to OpenAI, a fifth followed Thursday.

In today's brief:

• Google loses 5 senior AI researchers this week: John Jumper, Jonas Adler, Alexander Pritzel, Arthur Conmy to Anthropic; Noam Shazeer to OpenAI
• GPT-5.6 prediction-market odds fall from 83% to 18% on a June 28 deadline
• OpenCode hits 160K GitHub stars and 7.5M monthly active developers
• DeepSeek V4-Pro's 75% price cut is now the permanent list price

THE BIG ONE

Google's AI Research Exodus: Five Departures in One Week

Bloomberg reported on June 24 that Jonas Adler, who led Google's AI coding work, and Alexander Pritzel, a model training specialist, are both leaving for Anthropic. That came days after reports that Nobel Prize-winning biologist John Jumper - co-creator of AlphaFold2, whose protein structure prediction work earned the 2024 Chemistry Nobel - is also heading to Anthropic. Noam Shazeer, a co-author of the foundational transformer architecture who returned to Google two years ago after co-founding Character.AI, is joining OpenAI. On Thursday, Arthur Conmy, a senior research engineer credited on Gemini 2.5 and AI safety work, wrote on X that he is joining Anthropic specifically to work on safety.

TechCrunch framed this as a pattern: Anthropic recently raised at a $965 billion valuation and is reportedly considering an IPO as early as autumn. Pre-IPO equity is the pull, not just research mission. Google has not confirmed a retention plan or competing offer.

Why it matters: When a Nobel laureate and a co-author of the transformer paper leave for competitors in the same week, the compounding effect on model capability shows up 18 to 36 months out - not in this quarter's benchmarks but in the next generation of flagship models.

Our coverage: Noam Shazeer's move to OpenAI and the June 2026 frontier model landscape for where Google Gemini stands heading into this shift.

PLATFORMS

GPT-5.6 Prediction Markets Price In a July Launch

Polymarket odds for a GPT-5.6 release by June 28 collapsed from roughly 83% to 18% over the past few days, according to model release trackers. An internal memo from OpenAI Chief Scientist Jakub Pachocki, reported by TechTimes in mid-June, described GPT-5.6 as "a meaningful improvement" over GPT-5.5. The rumored upgrade list includes a 1.5 million token context window (up from 1 million in GPT-5.5), sharper front-end code generation, and faster Codex response times. None of this is confirmed by an official OpenAI announcement.

OpenAI has not set a public launch date. The company's historical release pattern - ChatGPT first, then Codex, then API - means developer access trails even a consumer launch by days or weeks. A July slip means GPT-5.6 enters a market where Claude Fable 5 is already shipping, Gemini 3.5 Flash has computer use, and FrontierCode shows Claude Opus 4.8 leading on code mergeability at 13.4% diamond-tier score.

Why it matters: The longer GPT-5.6 is delayed, the more the current frontier benchmark landscape - where Opus 4.8 leads FrontierCode and Fable 5 holds the coding quality lead - becomes the decision context for teams choosing a primary AI stack.

TOOLS WORTH A LOOK

1. OpenCode - Terminal-based, model-agnostic AI coding agent from the SST/Anomaly team. 160K GitHub stars, 7.5M monthly active developers, 75+ AI providers supported. Runs locally, plan-then-build split so you review changes before execution. Our developer walkthrough: OpenCode setup and workflow guide. (OSS / bring your own API keys)

2. FrontierCode - Cognition's benchmark for AI coding agents scored on mergeability: scope control, regression safety, test quality, maintainability. Three tiers (Diamond, Main, Extended), tasks authored with OSS maintainers. Claude Opus 4.8 leads at 13.4% diamond score; GPT-5.5 at 6.3%. Our analysis: what FrontierCode changes about how you pick a coding agent. (Free leaderboard / OSS benchmark)

3. DeepSeek V4-Pro - The May promotional 75% price cut is now the permanent list price: $0.44/MTok input, $0.87/MTok output, $0.003625/MTok on cache hits. On output tokens, 34x cheaper than GPT-5.5 and 17x cheaper than Claude Opus 4.7. The economics for agentic workloads: DeepSeek V4 cost-quality analysis. (Paid API)

WHAT ELSE IS HAPPENING

• Anthropic's $965B valuation and IPO prep: Pre-IPO equity packages are the mechanism driving Google researcher recruitment. An autumn 2026 offering is reportedly on the table. (5 min read)
• OpenAI filed a confidential S-1 with the SEC on June 8, targeting a September 2026 IPO at up to $1 trillion valuation. Goldman Sachs and Morgan Stanley are lead underwriters. A confidential S-1 is a step toward an IPO, not a commitment to complete one. (3 min read)
• Google fired a DevRel engineer for building a Google Workspace CLI: Justin Poehnelt's open-source tool hit #1 on Hacker News. Google terminated him two days before announcing their own version. Our full breakdown: the DevRel firing and what it signals.
• Vulnerability reports are not special anymore: Cryptographer Filippo Valsorda argues LLMs have commoditized vulnerability discovery, breaking the coordinated disclosure model that gave researchers leverage. Our coverage: why LLM-era vuln reporting needs new norms.
• Arcade raised $60M for secure AI agent authorization: The MCP-native action layer for production agents closes a gap between OAuth flows and multi-agent workflows. Developer guide: Arcade agent authorization in practice.
• AI code tools market at $9.3B in 2026: Mordor Intelligence estimates 26% annual growth toward $30B by 2031. DeepSeek's permanent frontier-tier price reset changes build-vs-buy math for every team in that market. (report)

---

Every link above goes to a primary source or our sourced coverage. Browse past issues at /daily and subscribe to get it by email.