Anthropic Forces Claude Model Migration as Multiple Versions Face 19 April Shutdown

This week delivered a stark reminder that AI provider changes wait for no one. Anthropic dropped multiple deprecation notices with immediate effect, forcing users to migrate away from Claude Haiku 3, 3.5, and Sonnet 3.7 models by 19 April 2026. Meanwhile, Google quietly extended its Vertex AI migration deadline and AWS rolled out quantum-resistant security measures.

The Big Moves

Anthropic's Aggressive Model Deprecation Strategy

Anthropic has pulled the trigger on a comprehensive model cleanup, deprecating three Claude variants with just days' notice. The Claude Haiku 3 (claude-3-haiku-20240307), Claude Haiku 3.5 (claude-3-5-haiku-20241022), and Claude Sonnet 3.7 (claude-3-7-sonnet-20250219) models all face retirement on 19 April 2026.

The migration paths are clear but urgent: Claude Haiku 3 and 3.5 users must move to Claude Haiku 4.5, whilst Claude Sonnet 3.7 applications need upgrading to Claude Sonnet 4.6. This isn't a gentle nudge towards newer models, it's a forced march. Applications still calling these deprecated endpoints will simply stop working after the sunset date.

What makes this particularly challenging is the timing coinciding with multiple service incidents across Claude's platform. Users experienced critical errors across Claude.ai, the API, and Claude Code earlier this week, followed by desktop app outages requiring immediate updates to v1.3109.0. The combination of forced migrations and platform instability creates a perfect storm for development teams already stretched thin.

The silver lining comes in the form of Claude Opus 4.7's release, which delivers substantial improvements in coding capabilities and multimodal understanding. Early reports suggest significant performance gains for complex software engineering tasks, though the timing feels more like compensation than coincidence. Teams planning their migration should evaluate whether the enhanced capabilities of the newer models justify the disruption.

Google's Vertex AI Endpoint Migration Gets Extended Deadline

Google has provided some breathing room for Vertex AI users facing endpoint migrations, extending the deadline to 30 June 2025 for Imagen and Veo generation endpoints. This affects a broad range of services including Gemma 4 26B A4B IT, Lyria 3, and Veo 3.1 Lite models.

The documentation updates reveal significant changes to the platform architecture, suggesting Google is consolidating its generative AI services under a more unified approach. The introduction of Gemini 3.1 Flash-Lite alongside these changes indicates a strategic shift towards lighter, more efficient models that can serve a broader range of use cases.

Unlike Anthropic's immediate deadlines, Google's approach provides adequate planning time. However, the scope of changes is substantial enough to warrant immediate assessment. Teams should audit their current Vertex AI integrations and begin testing against the new endpoints well before the June deadline. The risk isn't just service disruption, it's missing out on performance improvements and cost optimisations that the new architecture provides.

AWS Introduces Quantum-Resistant Security Measures

AWS Secrets Manager has quietly rolled out hybrid post-quantum TLS encryption, automatically protecting secrets against future quantum computing threats. This addresses the 'harvest now, decrypt later' attack vector where adversaries collect encrypted data today with the intention of decrypting it once quantum computers become capable enough.

The implementation is refreshingly transparent: customers can verify activation by monitoring CloudTrail logs for the 'X25519MLKEM768' key exchange algorithm. This automatic deployment across AWS services and SDKs demonstrates how security improvements should be handled, requiring no customer action whilst providing clear verification mechanisms.

Whilst quantum computing threats remain largely theoretical, the proactive approach signals AWS's commitment to long-term security planning. Organisations in regulated industries or handling sensitive data should note this capability and consider how it fits into their broader security posture.

Worth Watching

Weaviate's Database Performance Fixes

Weaviate has released multiple versions addressing critical performance issues with secondary index size accumulation. The fixes in v1.35.18, v1.36.12, and v1.37.1 specifically target startup speed improvements and bucket creation optimisation. These aren't feature releases, they're stability fixes that suggest previous versions had significant performance degradation under certain workloads. Users should prioritise these updates and monitor performance metrics closely.

NVIDIA's Nemotron OCR v2 Breakthrough

NVIDIA's Nemotron OCR v2 represents a significant leap in multilingual OCR capabilities, achieving 34.7 pages per second on an A100 GPU. The model's success stems from a synthetic data pipeline generating 12 million training images across six languages, addressing the previous version's limitations with non-English text. This capability could substantially reduce manual transcription costs for organisations processing multilingual documents.

OpenAI's Responses API Streaming Issues

OpenAI experienced streaming errors with the Responses API, highlighting potential instability in this relatively new service. Whilst resolved quickly, the incident suggests the API may not be ready for production workloads requiring high reliability. Teams using the Responses API should implement robust retry logic and consider fallback mechanisms.

Claude Design Enters Research Preview

Anthropic has launched Claude Design, an AI-powered visual creation tool available to Pro, Max, Team, and Enterprise subscribers. Powered by Claude Opus 4.7, the service enables collaborative design work for prototypes and presentations. The research preview status suggests limited availability and potential reliability issues, but early access could provide competitive advantages for design-heavy workflows.

Quick Hits

• AWS launched EC2 C8in and C8ib instances with up to 43% performance improvements over C6in, targeting network-intensive and database workloads • Deadline Cloud gained an AI-powered troubleshooting assistant to automate render job diagnostics • SageMaker HyperPod now supports flexible instance groups across multiple subnets and instance types • Elastic Security's resource-based pricing is driving MSSP adoption for cost management and AI-powered threat detection • AWS Elastic Disaster Recovery expanded to the European Sovereign Cloud in Germany for data sovereignty compliance • OpenAI released MLX Skill to streamline transformer model porting from transformers to mlx-lm • HCompany launched HoloTab AI browser companion for automating repetitive web tasks

The Week Ahead

The 19 April deadline for Anthropic's model deprecations dominates the immediate horizon. Teams still running Claude Haiku 3, 3.5, or Sonnet 3.7 have mere days to complete their migrations. This isn't optional maintenance, it's a hard cutoff that will break applications.

Beyond the immediate crisis, Google's 30 June 2025 Vertex AI endpoint migration deadline provides a more manageable timeline but affects a broader range of services. Start planning now rather than scrambling later.

Watch for potential follow-up incidents across Claude's platform, given this week's pattern of outages coinciding with major changes. The combination of forced migrations and service instability suggests Anthropic may be pushing updates faster than their infrastructure can handle reliably.