Anthropic AI Updates: April 17, 2026
1. Claude Opus 4.7 Released with Coding, Vision, and Agentic Upgrades
Anthropic. The new flagship lands on Claude products, the API, Amazon Bedrock, Google Cloud Vertex AI, and Microsoft Foundry at the same price as Opus 4.6: $5 per million input tokens and $25 per million output tokens. Anthropic reports a 13% improvement on coding benchmarks, 10-15% lift on complex task workflows, and a 21% reduction in document reasoning errors versus 4.6. Vision inputs now support images up to ~3.75 megapixels (roughly 3x the prior ceiling), and a new xhigh effort level sits between high and max for finer reasoning control. Task budgets entered public beta, and a /ultrareview command ships for code review in Claude Code. Note that the updated tokenizer maps inputs at 1.0-1.35x the prior token count, so existing cost models need re-calibration. Source
2. Opus 4.7 Deliberately Scales Back Cyber Offense While Improving Coding
Anthropic. The Decoder summarizes Anthropic’s approach: Opus 4.7 preserves improved software engineering and long-horizon autonomy while intentionally reducing certain offensive cybersecurity capabilities during training. That mirrors the policy trajectory set after the Mythos evaluation — Anthropic is willing to trade some raw capability on high-risk vectors for a cleaner safety profile, and is transparent about where it did so. Honesty and prompt-injection resistance also improved; harm-reduction advice specificity slipped slightly. Source
3. CPO Mike Krieger Resigns from Figma Board Ahead of Competing Design Tools
Anthropic. Anthropic’s chief product officer Mike Krieger stepped off Figma’s board on April 14 after The Information reported that Opus 4.7 would ship with design tools competing with Figma’s core offering. Krieger joined Figma’s board less than a year ago. The move adds concrete evidence to what investors are calling the “SaaSpocalypse” — frontier labs moving directly into application-layer markets that their own models threaten to commoditize. Source