Opus 4.7 just dropped

The upgrade is live. Anthropic released Claude Opus 4.7—SWE-bench Verified climbed from 80.8% on Opus 4.6 to 87.6%, and Anthropic’s internal 93-task coding benchmark showed a 13% improvement—the kind of gains engineering teams actually feel. Vision got a 3x resolution bump to roughly 3.75 megapixels (2,576 pixels on the long edge), and visual-acuity accuracy moved from 54.5% to 98.5%. Enterprise document analysis saw 21% fewer errors. Pricing stays at $5/$25 per million tokens input/output, which means the capability-per-dollar ratio just shifted again without touching the invoice. Developer tooling shipped alongside the model: a new xhigh reasoning effort level between high and max, task budgets in public beta for guiding token spend across longer runs, and a /ultrareview command in Claude Code for dedicated bug-finding sessions. The model is live on the API, Claude.ai, Bedrock, Vertex AI, and Microsoft Foundry. One catch: an updated tokenizer may inflate token counts 1.0–1.35x on existing prompts.

Anthropic upgraded horizontally; OpenAI went vertical. OpenAI debuted GPT-Rosalind, a frontier reasoning model purpose-built for biology, drug discovery, genomics, and protein reasoning—named after Rosalind Franklin. It’s not a general release: access is gated through OpenAI’s Trusted Access program, limited to Enterprise customers in the U.S. after a safety review. Launch partners include Amgen, Moderna, and Thermo Fisher Scientific. This follows the same identity-gated pattern as GPT-5.4-Cyber, which shipped earlier this week for vetted security professionals. Two gated models in one week, both locked behind credentials. The frontier is splitting: general-purpose on one track, industry-specific on the other—and the industry-specific track requires knowing who you are before you get access.

Not everyone is building behind gates. Mozilla’s MZLA Technologies launched Thunderbolt, an open-source, self-hostable enterprise AI client released under MPL 2.0 with native apps across Linux, macOS, Windows, iOS, and Android—5 platforms at launch. It integrates deepset Haystack for RAG, supports MCP servers and ACP agents, and lets teams bring any model provider including Ollama for fully local inference. The positioning is direct: a data-sovereignty alternative to Copilot, ChatGPT Enterprise, and Claude Enterprise. On the enterprise platform side, Salesforce used TDX 2026 to declare everything an API—Headless 360 opens the entire platform via API, MCP tool, or CLI. AgentExchange now hosts 13,000+ listings backed by a $50 million Builders Fund, and the free Developer Edition ships with Agentforce Vibes IDE, Claude Sonnet 4.5 as the default coding model, and 110 free Sonnet requests per month (1.5 million tokens through May 31). The agentic stack is hardening into infrastructure—the question now is whether it runs on your servers or theirs.