Sonnet 5 vs Sonnet 4.6: read the benchmark, not the backlash

The short answer

“Sonnet 5 is only good at beating Sonnet 4.6” is a sharp headline, but it hides the useful question: which model is better for your workload, under your settings, at an acceptable cost and latency?

Anthropic describes Sonnet 5 as its most agentic Sonnet, with improvements in coding, tool use, planning, and professional work. Its own system card is more nuanced than the launch summary. Sonnet 5 advances many headline evaluations, yet some specialized results are lower than Sonnet 4.6. That is normal for a new model and a reason to evaluate, not evidence that either marketing or community criticism is universally correct.

What the official evidence supports

The Sonnet 5 announcement says the model brings capabilities that recently required larger models to the Sonnet tier. It launched across Claude products and the API with a native one-million-token context window.

The system card documents the actual evaluation conditions and exposes mixed results. For example, in one biological sequence-design task, Sonnet 5's top-sequence design score was lower than Sonnet 4.6, while its prediction result was slightly better. Other evaluations use different effort settings, tools, scaffolds, token budgets, and numbers of trials.

That context is not fine print. A model can lead a coding benchmark and still feel worse on a particular repository because it edits too aggressively, follows local conventions less consistently, or spends more time reasoning. It can also lose a narrow scientific benchmark while being substantially better at multi-step software work.

Why launch charts and backlash both mislead

Launch material selects evaluations that explain a model's intended strengths. Community reports select memorable failures. Both are useful inputs, but neither is a workload-level decision.

Watch for four common comparison errors:

• Different harnesses — tool access, prompts, compaction, and retry policy can move agentic scores substantially.
• Single runs — nondeterministic models need repeated trials; one impressive success or failure is an anecdote.
• Changed defaults — effort level, thinking behavior, and Claude Code scaffolding may differ between releases.
• Benchmark substitution — a SWE-bench result does not measure your frontend style, data pipeline, or code-review precision.

Even a valid aggregate win can conceal regressions in a narrow task. Conversely, a handful of angry posts can come from migration surprises rather than a broad capability decline.

A practical migration test

Build a small evaluation set from work you already understand. Five to twenty tasks is enough to expose directional differences:

1. Include a bug fix, a scoped feature, a refactor, a test-writing task, and a code-review task.
2. Start each model from the same commit and the same instructions.
3. Match effort and tool permissions as closely as the products allow.
4. Run important tasks more than once.
5. Score correctness first, then unnecessary changes, test quality, instruction following, latency, and cost.

Keep the expected outcome outside the prompt so the model cannot simply mirror the rubric. Review the patch, run the same automated checks, and record whether a human had to intervene.

When to keep Sonnet 4.6

Stay on Sonnet 4.6 when it is already reliable on a stable production workflow and Sonnet 5 has not demonstrated enough improvement to justify migration risk. A known model with measured failure modes can be more valuable than a higher average benchmark score.

Move to Sonnet 5 when your paired evaluation shows better completion rate, fewer repair turns, or better cost-adjusted throughput on the work that matters. For mixed workloads, routing is a valid answer: keep the older model for a task where it wins and use Sonnet 5 for agentic coding where its improvements show up.

Takeaway

Sonnet 5 does not need to win every benchmark to be a worthwhile upgrade, and a launch chart does not make it the best choice for every task. Treat both the announcement and the backlash as hypotheses. Your repeated, versioned evaluation is the decision.

---

Claims checked against Anthropic's Sonnet 5 announcement, Sonnet model page, and Claude Sonnet 5 system card.