STEVEN STEINBECK
← ALL DISPATCHESDISPATCH 69

WEEK OF APR 18 – 24, 2026

Google bets $40B on Anthropic, GPT-5.5 lands, and Cloud Next turns everything into agents

A week of two things: OpenAI shipping its strongest agentic coding model yet, and Google going all-in on the agentic stack — money, silicon and a rebuilt platform.

Alphabet confirmed it will invest up to $40 billion in Anthropic, days after Amazon agreed to put in up to $25 billion. Anthropic's run-rate revenue passed $30 billion this month, up from roughly $9 billion at the end of 2025.

Google Cloud Next split its TPU line in two and rebranded Vertex AI into an agent platform with purpose-built runtimes, identity and sandboxing — while OpenAI shipped GPT-5.5 and workspace agents the same week.

FRI · Apr 24, 2026industryfundingcomputeanthropicgoogle

Google commits up to $40 billion to Anthropic

Alphabet confirmed it will invest up to $40 billion in Anthropic: $10 billion in cash now at a $350 billion valuation, with another $30 billion contingent on the Claude maker hitting performance targets.

The deal lands days after Amazon agreed to invest up to $25 billion, and comes as Anthropic's annual run-rate revenue passed $30 billion this month, up from roughly $9 billion at the end of 2025.

Google commits up to $40 billion to Anthropic
image.cnbcfm.com

WHY IT MATTERS

The single largest hyperscaler bet on a frontier lab this year, and further evidence that compute access — not just capital — is the scarce resource in the model race.

ReutersCNBC
THU · Apr 23, 2026model releaseopenaiagentic codingbenchmarks

OpenAI releases GPT-5.5, built for agentic coding and computer use

OpenAI shipped GPT-5.5 to Plus, Pro, Business and Enterprise users in ChatGPT and Codex, calling it its strongest agentic coding model yet; GPT-5.5 and GPT-5.5 Pro reached the API the next day with an updated system card.

OpenAI's published evals put it at 82.7% on Terminal-Bench 2.0 against 75.1% for GPT-5.4, 58.6% on SWE-Bench Pro and 78.7% on OSWorld-Verified, with roughly 20% faster token generation at matching per-token latency. OpenAI says it meets its High but not Critical cybersecurity risk threshold.

OpenAI releases GPT-5.5, built for agentic coding and computer use
image.cnbcfm.com

WHY IT MATTERS

Terminal-Bench and OSWorld both measure long-horizon tool use rather than chat quality, and both moved materially — the benchmark that matters for agents.

OpenAI — Introducing GPT-5.5CNBC
WED · Apr 22, 2026computechipsagent harnessesgoogle cloudsandboxing

Google Cloud Next: a split TPU line and an agent platform to govern agent fleets

Google split its eighth-generation TPU into two chips for the first time: TPU 8t for training, with up to 9,600 TPUs and 2 PB of shared memory in a superpod and 3x the processing power of Ironwood, and TPU 8i for inference, with 1,152 TPUs per pod, 3x the on-chip SRAM and 80% better performance per dollar.

In the same keynote, Vertex AI was rebranded the Gemini Enterprise Agent Platform, adding a re-engineered Agent Runtime for long-running agents, Memory Bank, Agent Identity, Agent Registry and Agent Gateway, plus GKE Agent Sandbox for executing untrusted agent code.

WHY IT MATTERS

This is the harness news of the week: purpose-built agent runtimes, identity and sandboxing are now first-class products, not glue code you write yourself.

Google Cloud — AI infrastructure at Next '26Google Cloud — Introducing Gemini Enterprise Agent Platform
WED · Apr 22, 2026agent harnessesopenaienterprisetool-calling

OpenAI launches workspace agents in ChatGPT, a Codex-powered successor to GPTs

OpenAI introduced workspace agents: shared, cloud-run agents powered by Codex that hold files, code, tools and memory, and keep working across multi-step workflows after you leave.

They connect to Slack, Google Drive, Microsoft apps, Salesforce, Notion and Atlassian Rovo, are governed by role-based admin controls, and ship in research preview for ChatGPT Business, Enterprise, Edu and Teachers plans — free until May 6, then credit-based pricing.

WHY IT MATTERS

OpenAI is productizing the same primitives you build by hand — persistent memory, tool connections, approval gates — into a managed per-seat product.

OpenAI — Introducing workspace agents in ChatGPTVentureBeat
TUE · Apr 21, 2026lawcopyrightdatabrickstraining data

Judge refuses to dismiss authors' copyright suit against Databricks over DBRX

Judge Charles R. Breyer of the Northern District of California denied Databricks and MosaicML's motion to dismiss and strike in In re Mosaic LLM Litigation, holding that authors plausibly alleged unauthorized copying of their books during development.

The writers allege roughly 196,000 pirated titles tied to the Books3 dataset surfaced in the chain that produced the DBRX model. The order was signed April 21 and reported April 23.

WHY IT MATTERS

Databricks has not yet argued fair use the way Meta and Anthropic did, so this is the first test of the theory that copying books anywhere in the pipeline, not just in final training data, is infringement.

Court order (Justia docket)Bloomberg Law
MON · Apr 20, 2026open weightsmoonshot aiagent swarmscoding agents

Moonshot AI open-sources Kimi K2.6, a trillion-parameter agent-swarm model

Moonshot AI released Kimi K2.6 with open weights on Hugging Face under a modified MIT license — a roughly 1T-parameter MoE with about 32B active parameters and a 256K context window.

Its Agent Swarm layer scales to 300 sub-agents executing 4,000 coordinated steps, up from 100 and 1,500 in K2.5, and the model reports 58.6% on SWE-Bench Pro, 54.0% on Humanity's Last Exam with tools and 80.2% on SWE-Bench Verified. Moonshot showed multi-hour autonomous runs, including a 13-hour session that reworked an open-source matching engine.

WHY IT MATTERS

An open-weight model matching or edging GPT-5.4 on coding and agentic evals, with a swarm orchestration layer and an OpenAI- and Anthropic-compatible API, is the most consequential open release of the window.

Hugging Face — Kimi-K2.6 model cardKimi — K2.6 tech blog