Sunday, April 19, 2026

TLDR

Claude Haiku 3 retires TODAY (Apr 19) — API calls to claude-3-haiku-20240307 will error after today; migrate to claude-haiku-4-5 now
Claude Opus 4.7 launched (Apr 16) — 3× vision resolution, 1M-token context, new "xhigh" effort mode for complex agent workflows; now on API, Bedrock, Vertex, and GitHub Copilot
xAI storage billing starts April 20 — Files on the xAI platform incur charges starting tomorrow; audit your usage today
Physical Intelligence π0.7 (Apr 16) — Robot foundation model achieves compositional generalization: combines motor skills to tackle tasks it was never trained on
Stanford AI Index 2026 — U.S.–China AI gap compressed to 2.7%; SWE-benchA benchmark that tests whether a model can resolve real GitHub issues by editing code in popular Python repositories.Jimenez et al. 2023 ↗ scores jumped from 60% to ~100% in one year; model transparency scores declining sharply

Things your engineering team should look at or do this week.

Migrate from claude-3-haiku-20240307 today — Anthropic retires this model on April 19 (today). After today, API calls to this model return errors. Recommended replacement: claude-haiku-4-5. Scan your codebase for any hardcoded model strings before EOD. (Anthropic Docs | Migration Guide)
Audit xAI file storage before April 20 — xAI begins billing for files and collections stored on its platform starting tomorrow. If your team uses the xAI Files API, check the xAI Console, delete unused assets, and set spend caps now to avoid surprise charges. (xAI Docs)
Plan Gemini 2.0 Flash migration before June 1 — Google is deprecating Gemini 2.0 Flash and 2.0 Flash-Lite on June 1, 2026. Recommended path: Gemini 2.5 Flash or Gemini 3 Flash. Also note Gemini Pro models are now paid-only — the free tier was removed on April 1. (AI Pricing Guru | FindSkill)

Claude Opus 4.7 (Apr 16) Agents — Anthropic's new flagship is live across Claude.ai, the API, Amazon Bedrock, Google Vertex AI, and is rolling out on GitHub Copilot. Key upgrades: 3× higher vision resolution, a new "xhigh" effort level for long-running agent tasks, 1M-token context windowHow much text a model can see and use at once while writing a reply.Anthropic ↗, and automatic cyber safeguards. Same pricing as Opus 4.6. (GitHub Changelog | AWS Blog | Anthropic Docs)
Physical Intelligence π0.7 (Apr 16) Robotics/IoT — PI's new robot foundation model demonstrates compositional generalization: like an LLMLarge language modelA neural network trained on huge amounts of text that generates text, answers questions and summarizes.Anthropic ↗ combining concepts, π0.7 can combine trained motor skills to tackle tasks it was never explicitly trained on. Matched specialist model performance on coffee-making, laundry folding, and box assembly; cross-embodiment transfer to a UR5e industrial arm succeeded at 80% on t-shirt folding. Company being valued at ~$11B. Worth watching for teams exploring embodied AI or physical automation. (TechCrunch | PI Blog)

Anthropic: claude-3-haiku-20240307 retires today (Apr 19) — API calls to this model return errors after today. Migrate to claude-haiku-4-5. See Action Items above. (Anthropic Docs)
xAI: Storage billing starts April 20 Infra — Files and collections stored on the xAI platform will be billed starting tomorrow based on storage volume. Rate details are in the xAI Console. (xAI Docs)
Google Gemini: Pro models paid-only; Gemini 2.0 deprecation June 1 (Apr 1 / ongoing) — Gemini Pro models (3.1 Pro, 3 Pro, 2.5 Pro) were removed from the free tier on April 1; Flash models remain free with reduced daily quotas. Gemini 2.0 Flash and 2.0 Flash-Lite officially deprecate June 1. Migrate to Gemini 2.5 Flash or Gemini 3 Flash. New accounts also default to prepaid billing. (AI Pricing Guru | Google Gemini Pricing)

Novo Nordisk partners with OpenAI (Apr 14) — The Danish pharma giant (Ozempic/Wegovy) signed a broad strategic partnership to integrate AI across R&D, manufacturing, supply chain, and commercial operations. OpenAI will help upskill Novo's global workforce; pilot programs launch immediately with full rollout targeted by end of 2026. Structurally similar to other enterprise AI deployments but notably covers the full product lifecycle, not just drug discovery. (CNBC | FiercePharma)
Snap cuts 1,000 jobs (16% of workforce), cites AI productivity (Apr 15) Hiring — CEO Evan Spiegel confirmed AI now generates more than 65% of Snap's new code. This is one of the first major tech layoffs where a public company explicitly named AI coding productivity — not just macroeconomic conditions — as the primary rationale. Stock jumped 7% on the news. A signal worth tracking for engineering hiring plans. (TechCrunch | CNBC)

Stanford AI Index 2026: China gap at 2.7%, transparency collapsing (published Apr 16, trending) — Key findings for engineering leaders: (1) The U.S.–China AI performance gap compressed from double digits to just 2.7%; models have traded leads repeatedly since early 2025. (2) SWE-bench Verified scores rose from 60% to ~100% in a single year. (3) Foundation Model Transparency Index scores dropped to 40 from 58 — meaning companies are disclosing less about training data, compute, and risks over time. (4) AI researcher inflow to the U.S. fell 80% in the past year. Worth a read before your next roadmap conversation with leadership. (Stanford HAI Report | The Decoder | Fortune)

"Autogenesis: A Self-Evolving Agent Protocol" (Apr 17, arXiv) — Preprint proposes an agent framework in which AI systems iteratively rewrite their own architecture and toolset based on task performance — without human intervention. Early-stage work, but the direction (agents that evolve their own scaffolding at runtime) has direct implications for teams building long-running agenticDescribes AI systems where a model uses tools and decides its own steps instead of following a fixed script.Anthropic ↗ workflows. (arXiv)