AI weekly digest: The Model Wars Escalate

This was the biggest model-drop week of 2026 so far. Anthropic shipped Opus 4.7, OpenAI countered with GPT-Rosalind and open-sourced its first weights, Google released Gemma 4 under Apache 2.0, and the Mythos saga continued to pull in the White House. Meanwhile, the infrastructure side saw a record-breaking compute deal and a new automation primitive landed in Claude Code.

1. Anthropic launches Claude Opus 4.7 with new tokenizer, 3x vision, and xhigh effort

Anthropic released Claude Opus 4.7 on April 16, succeeding Opus 4.6 as the new top-tier model. The headline numbers: 87.6% on SWE-bench Verified (+7 pts), 64.3% on SWE-bench Pro (+11 pts), and #1 on GDPval-AA with 1753 Elo. A new tokenizer (signaling a possible new pretrain) can inflate token counts up to 35%, but improved reasoning efficiency means net token usage is still down ~50% at equivalent effort levels. The model introduces an xhigh effort tier (now the Claude Code default), high-resolution vision up to 2,576 px (~3.75 MP, 3x previous), and differential cyber-capability reduction from the Mythos safety work. Cursor reported its internal benchmark jumped from 58% to 70%; Notion saw a 14% lift with one-third fewer tool errors. Pricing stays at $5/$25 per million input/output tokens. Source: Anthropic Claude Opus 4.7 (CNBC)

2. OpenAI debuts GPT-Rosalind for life sciences and a new Codex architecture

OpenAI shipped two major releases on the same day. GPT-Rosalind is a gated, domain-specific model fine-tuned for biochemistry, drug discovery, and genomics — the first time OpenAI has released a vertical model. Early partners include Amgen, Moderna, the Allen Institute, and Thermo Fisher. Separately, OpenAI revealed the architecture of its new Codex agent: a shared Rust-based "harness" powering CLI, web, VS Code, and macOS surfaces through a bidirectional JSON-RPC protocol, with built-in prompt caching and context compaction. Source: OpenAI launches GPT-Rosalind (VentureBeat)

3. OpenAI open-sources gpt-oss-120b and gpt-oss-20b under Apache 2.0

OpenAI released its first open-weight models: gpt-oss-120b (117B total params, 5.1B active via MoE) and gpt-oss-20b (21B total, 3.6B active). The 120b model runs on a single 80 GB GPU and matches or exceeds o4-mini on coding (Codeforces), reasoning (MMLU, HLE), and tool calling (TauBench). Apache 2.0 license, available on Hugging Face and LM Studio. This is a direct response to Meta's Llama, DeepSeek, and Google's Gemma — OpenAI is now in the open-weight race. Source: Introducing gpt-oss (OpenAI)

4. Google releases Gemma 4 — open models with 256K context, function calling, and multimodal input

Google shipped Gemma 4 on April 2 under Apache 2.0, with four variants: E2B, E4B, 26B MoE, and 31B Dense. The 31B dense model outperforms models 20x its size on several benchmarks and features a 256K token context window, native function calling, video/audio input, and support for 140+ languages. With over 400M cumulative Gemma downloads, this is Google's strongest play yet to compete with Llama and now gpt-oss in the open-weight space. The 26B MoE variant activates only 3.8B parameters, making it practical for edge and local deployment. Source: Gemma 4 (Google Blog)

5. Anthropic briefs the White House on Mythos; government seeks access

Anthropic co-founder Jack Clark confirmed the company briefed the Trump administration on Claude Mythos, its unreleased 10-trillion-parameter model with advanced cyber-offensive and defensive capabilities. VP Vance and Treasury Secretary Bessent held calls with CEOs from Anthropic, OpenAI, Google, xAI, and major cybersecurity firms. Treasury CIO Sam Corcos is preparing his team for eventual access. Meanwhile, Mythos remains locked behind a 50-company firewall — Opus 4.7 ships with differentially reduced cyber capabilities inherited from Mythos safety work. The situation is complicated by Anthropic's ongoing legal dispute with the DoD. Source: Anthropic co-founder confirms briefing (TechCrunch)

6. OpenAI launches GPT-5.4-Cyber via Trusted Access for Cyber program

OpenAI released a cyber-permissive variant of GPT-5.4 called GPT-5.4-Cyber, expanding its Trusted Access for Cyber (TAC) program to thousands of individual defenders and hundreds of teams. Unlike Anthropic's approach of restricting Mythos, OpenAI is going the opposite direction: making cyber-specialized models as widely available as possible to vetted security professionals. Companies can now authenticate as cybersecurity defenders, with the highest tiers getting GPT-5.4-Cyber access. This is the first time a frontier lab has released a model explicitly optimized to be permissive about security research. Source: Trusted Access for Cyber Defense (OpenAI via TLDR)

7. Jane Street commits $6B to CoreWeave, takes $1B equity stake

Jane Street signed a $6 billion AI cloud agreement with CoreWeave and invested $1 billion in equity at $109/share. The deal gives Jane Street access to next-gen compute infrastructure including NVIDIA Vera Rubin systems. This is CoreWeave's third multi-billion-dollar deal this month, following commitments from Meta and Anthropic. The deal signals that quantitative trading firms now view AI compute as critical infrastructure — Jane Street is betting that AI-driven trading returns will justify the spend. Source: Jane Street-CoreWeave deal (CoreWeave)

8. Claude Code ships Routines — scheduled, API-triggered, and event-driven automations

Anthropic launched Routines in research preview on April 14, adding a new automation primitive to Claude Code. A routine is a configured prompt + repo + connectors combination that runs on a schedule, via API (HTTP POST with bearer token), or in response to GitHub events (PRs, releases). Routines execute on Anthropic's web infrastructure, not your laptop. Pro gets 5 runs/day, Max 15, Team/Enterprise 25. This positions Claude Code as more than a coding assistant — it's becoming a lightweight CI/CD and ops automation layer. Source: Introducing routines in Claude Code (Anthropic)

9. Gemini 3.1 Flash TTS and native Mac app arrive

Google had a busy week beyond Gemma. Gemini 3.1 Flash TTS launched with an Elo score of 1,211 on the Artificial Analysis TTS leaderboard, supports 70+ languages, and introduces audio tags for granular vocal style control via natural language — all outputs watermarked with SynthID. Separately, the native Gemini Mac app shipped with system-wide Cmd+Space-style access, screen context sharing, and support for image/video generation via Nano Banana and Veo. Google also introduced "Skills" in Chrome for saving and reusing AI prompts across websites. Source: Gemini 3.1 Flash TTS (Google Blog)

10. Five hyperscalers now control two-thirds of global AI compute

Epoch AI reported that Google, Microsoft, Meta, Amazon, and Oracle now control over two-thirds of the world's AI compute capacity. Most AI labs depend almost entirely on these hyperscalers for training runs. Combined with this week's infrastructure deals (Jane Street/CoreWeave, Meta/Broadcom's 1 GW custom chip commitment through 2029, Microsoft securing 30,000 Vera Rubin GPUs at Norway's former Stargate site), the AI compute supply chain is consolidating fast. If you're not one of these five — or contracting with CoreWeave — accessing frontier-scale compute is becoming increasingly difficult. Source: Five hyperscalers own two-thirds of AI compute (Epoch AI)