Anthropic's Claude 4: What Changed, What It Means for Builders, and the Honest Capability Assessment

Anthropic's Claude 4 family — Haiku 4.5, Sonnet 4.6, and Opus 4.7 — represents a meaningful generational step forward from the Claude 3 series. The improvements are real and unevenly distributed across capability areas. For builders who have applications running on Claude 3.5 Sonnet or Claude 3 Opus, the migration decision is not obvious — and the standard benchmark comparisons don't provide enough signal to make it well. Here's the honest assessment.

What Genuinely Improved

Instruction following is the most consistently improved capability across the Claude 4 family. The gap between a precisely-worded prompt and a loosely-worded one producing different outputs has narrowed. This matters practically: prompts that required careful iterative refinement on Claude 3 often work on first attempt with Claude 4 Sonnet. For production applications where prompt reliability determines product quality, this improvement reduces maintenance overhead.

Extended context performance is substantially better. Claude 3 models, despite their large context windows, showed meaningful performance degradation on tasks requiring retrieval and synthesis from the middle of long documents. Claude 4 handles long-context reasoning more uniformly — important for applications that process large documents, long conversation histories, or extensive codebases.

Tool use and structured output reliability have also improved. Function calling and JSON output generation are more consistent, with fewer edge-case failures on complex schemas. For agentic applications with multi-step tool chains, this reduces the retry logic overhead required to achieve reliable outputs.

What Hasn't Changed as Much as Claimed

The creative writing and nuanced communication improvements are real but smaller than Anthropic's marketing suggests relative to Claude 3.5 Sonnet. For most business writing applications — email drafts, report summaries, content generation — Claude 3.5 Sonnet outputs are difficult to distinguish from Claude 4 Sonnet outputs without careful side-by-side evaluation.

Hallucination rates have improved but the character of hallucinations has not fundamentally changed. Claude 4 still produces confident, well-structured incorrect information on topics outside its training data or on questions that require recency. The improvement is in frequency, not in the fundamental epistemological limitation. Applications requiring high accuracy on domain-specific or time-sensitive information still need external knowledge retrieval regardless of model version.

The Model Tier Decision

Claude Haiku 4.5 is meaningfully better than Claude 3 Haiku and competitive with Claude 3.5 Sonnet on many tasks, at significantly lower cost. For high-volume, lower-complexity applications — classification, extraction, basic Q&A, simple summarisation — migrating to Haiku 4.5 can deliver simultaneous quality improvements and cost reductions. This migration is almost always worth doing.

Claude Sonnet 4.6 vs. Claude 3.5 Sonnet is a closer call. Sonnet 4.6 is better on complex reasoning, instruction following, and long-context tasks. For applications pushing these capabilities, the upgrade is justified. For applications that perform well on Claude 3.5 Sonnet, the marginal improvement may not justify the migration overhead and higher costs.

Claude Opus 4.7 is the clearest upgrade case over Claude 3 Opus. The capability improvement is substantial and the extended thinking mode delivers reasoning depth that has no equivalent in the Claude 3 family. For the tasks that genuinely require Opus-level capability, 4.7 is meaningfully better.

Migration Recommendation

Start by migrating Haiku applications to Haiku 4.5 — the economics are favourable and the quality improvement is consistent. Then evaluate Sonnet applications with a structured test: run your 20 highest-stakes prompts through both models and score the outputs. If Sonnet 4.6 produces noticeably better outputs on those specific prompts, migrate. If not, defer the migration until you have a specific capability need it addresses.

Practical takeaway: Migrate to Haiku 4.5 immediately for any high-volume Haiku application. Evaluate Sonnet on your specific use cases before migrating — the improvement is real but the decision should be data-driven, not driven by release timing.