Anthropic shipped Claude 4 with unusually little fanfare. There was no keynote, no countdown, no glossy demo reel — just a blog post, an updated API reference, and a quiet rollout to existing customers. Two months in, it is clear why: Claude 4 is less a flashy new product than a thoughtful refactor of how a frontier model should behave when given hard problems.
Reasoning as a dial, not a switch
The headline feature is configurable reasoning. Instead of a binary 'thinking mode,' Claude 4 exposes a continuous parameter that controls how much inference-time compute the model spends on a request. At the low end, it behaves like a fast chat model. At the high end, it produces multi-step reasoning traces that rival dedicated reasoning models from competitors.
What makes this useful in practice is that the dial can be set per request. A customer-support agent can route trivial questions to fast mode and escalate edge cases to deep mode without changing models. That granularity has been missing from most reasoning-model offerings until now.
Computer use, generally available
Claude's computer-use feature graduated from preview to general availability. The model can now control a virtual desktop — clicking, typing, scrolling, and reading screenshots — with significantly better reliability than the original release. In internal benchmarks, success rates on multi-step office tasks climbed from roughly thirty percent to over seventy.
The implications are large. Whole categories of automation that previously required brittle browser scripts or vendor APIs can now be handled with a single model and a screenshot. That includes legacy enterprise software, scanned documents, and the long tail of internal tools that no engineering team has time to wrap in a proper API.
Honesty and refusal behavior
Anthropic has long been the most explicit lab about model honesty. Claude 4 takes that further. The model will refuse fewer benign requests than its predecessor — a frequent complaint about Claude 3 — while becoming more willing to flag uncertainty in its own answers. In practice, you see more sentences like 'I'm not confident about this specific number, but here's how I'd verify it.'
That kind of calibrated uncertainty is exactly what production users have been asking for. It makes downstream verification cheaper and reduces the rate of confidently wrong answers, which is the single most expensive failure mode for any AI-powered product.
Long context, but with caveats
Claude 4 supports a one-million-token context window in beta. As with every long-context release, the practical question is not 'does it fit?' but 'does it use the information well?' Early evaluations suggest Claude 4 is noticeably better at needle-in-a-haystack retrieval than Claude 3, but still degrades on tasks that require synthesizing many distant pieces of information.
For most teams, the right pattern remains the same: retrieve aggressively, summarize when you can, and only fall back to dumping the whole corpus into context when retrieval is impossible. Long context is a useful escape hatch, not a replacement for a real retrieval pipeline.
How it compares
Versus GPT-5: Claude 4 is generally better at writing, coding refactors, and following nuanced instructions. GPT-5 is faster, cheaper at equivalent quality, and stronger at structured tool use. Versus Gemini 3: Claude 4 wins on text reasoning; Gemini wins on multimodal grounding, especially video.
Versus open-weights models: Claude 4 remains a clear step ahead on hard reasoning, but the gap on routine tasks is narrower than ever. If your workload is mostly summarization and classification, you can probably get acceptable quality from a self-hosted model at a fraction of the cost.
Should you switch?
If you're already on Claude 3, the upgrade is essentially free and the wins are real. If you're on a competitor, the answer depends entirely on your evaluation suite. Anthropic's pricing remains middle-of-the-pack, the API is stable, and the company's safety posture is unusually clear — all things that matter more in year three of a product than they did at launch.
The most honest summary: Claude 4 is the most polished frontier model on the market today. It is not a revolution. It is a careful, considered iteration that quietly raises the floor on what a production AI feature can do.