# Agent Platform Research Briefing — April 25, 2026

Welcome to the agent platform research briefing for Saturday, April 25th, 2026. Big week wrapping up with three major stories.

**Google Commits $40 Billion to Anthropic** — The headline of the weekend: Google-parent Alphabet will invest up to $40 billion in Anthropic, confirmed Friday April 24th. The deal breaks down as $10 billion upfront with an additional $30 billion contingent on milestones. This is on top of Google's existing $3 billion investment and 14 percent stake in Anthropic. The investment is primarily about compute capacity — closing the gap between exploding Claude Code demand and available inference power. The timing matters: Anthropic hit a $30 billion annual revenue run-rate, up from $9 billion at the end of 2025. Claude Code alone is driving most of that growth. Combined with Amazon's $25 billion deal announced earlier this week, Anthropic now has roughly $65 billion in committed infrastructure backing from just two hyperscalers. Google framed it as "spreading its bets" — competing with its own Gemini models while simultaneously funding the biggest threat to those models. The strategic tension is remarkable: Google is betting against itself and still winning either way.

**OpenAI Ships GPT-5.5 "Spud" — API Coming Soon** — OpenAI released GPT-5.5 on April 23rd, six weeks after GPT-5.4 and just one week after Anthropic's Opus 4.7. The model scores 82.7 percent on Terminal-Bench 2.0 — that's state-of-the-art among publicly available models — Opus 4.7 scored 69.4 percent and Gemini 3.1 Pro scored 68.5 percent on the same benchmark. OpenAI claims 40 percent better token efficiency than 5.4, meaning more capability per compute dollar. But pricing doubled: $5 per million input tokens and $30 per million output. API access is not yet available — it's rolling out to ChatGPT and Codex for paid subscribers first, with the API coming "very soon." No Pro variant in wide release yet. The GPT-5.5 codename "Spud" leaked in a memo last week, but the model nickname appears to have been dropped from the official launch messaging. Greg Brockman called this a step toward OpenAI's "super app" — a combined ChatGPT-Codex-browser experience for enterprise users.

**DeepSeek V4 Officially Released — V4-Pro Hits 80.6% SWE-bench** — DeepSeek officially released V4-Pro (1.6 trillion parameters, 49 billion active per token) and V4-Flash (284 billion parameters, 13 billion active) on April 24th after months of delays. V4-Pro scores 80.6 percent on SWE-bench Verified — within 0.2 points of Claude Opus 4.6 — and DeepSeek claims SOTA for agentic coding among open-source models. The efficiency story is the headline: at 1 million token context, V4-Pro uses only 27 percent of the compute and 10 percent of the KV cache compared to V3.2. It's optimized for Huawei Ascend chips. DeepSeek is also deprecating the legacy deepseek-chat and deepseek-reasoner API endpoints in 90 days. For context: GPT-5.5 hit Terminal-Bench 82.7 percent the day before, so we now have three models — GPT-5.5, Mythos Preview, and V4-Pro — crowding the top of every agentic benchmark. The gap between frontier and open-source is officially closing.

That's the briefing for today.