Claude Sonnet 5: agentic capability moves down the cost curve

On June 30, Anthropic released Claude Sonnet 5, positioned as the most agentic Sonnet model to date — built to make plans, drive tools like browsers and terminals, and run autonomously at a level that, until recently, took a larger and more expensive model. I run generative and agentic systems in production, so I read these launches through one question: does it change how I route work? This one does.

What shipped

Sonnet 5 is a direct upgrade to Sonnet 4.6, with gains on the parts of agentic work that actually matter in production: reasoning, tool use, coding, and knowledge work. It is the default model on Free and Pro, available to Max, Team, and Enterprise, and exposed as claude-sonnet-5 in Claude Code and on the Claude API. Introductory pricing runs at $2 per million input tokens and $10 per million output tokens through August 31, 2026, then settles to $3 and $15. For reference, Opus 4.8 sits at $5 and $25.

The headline claim is narrow and worth stating precisely: Sonnet 5's performance is close to Opus 4.8 on agentic tasks, at a lower price. Not equal — close. Anthropic is explicit that Opus 4.8 remains the higher-accuracy choice. What changed is the shape of the tradeoff.

The real change is routing economics

For most of the last year, the Sonnet-versus-Opus decision was a step function. Sonnet 4.6 was the affordable option, but on agentic search (BrowseComp) and computer use (OSWorld-Verified) it fell well short of Opus. If you needed the accuracy, you paid for Opus; if you did not, you accepted a real gap. There was empty space between them.

Sonnet 5 fills that space. On the cost-performance curves Anthropic published, Sonnet 5 and Opus 4.8 now sit on a single continuous frontier — you move along it by turning the effort level up or down rather than by swapping models. That is a different kind of decision. Instead of "which model," the question becomes "how much effort does this task justify," and the same family answers it from cheap-and-fast to expensive-and-careful.

There is a subtlety in the pricing that is easy to miss. Sonnet 5 ships with an updated tokenizer, so the same input can map to roughly 1.0–1.35× more tokens than before. The introductory pricing is set so the migration is roughly cost-neutral — not cheaper on paper. The practical consequence: compare models on cost-per-task, not cost-per-token. Token price alone will mislead you here.

What "more agentic" actually means

The capability I care about is not a benchmark number; it is follow-through. The failure mode of earlier models in long agentic runs was stopping short — doing four of six steps and declaring victory. The early-access feedback Anthropic published is consistent on this point, and one example matches exactly what I look for when evaluating a model for autonomous work: handed a bug, it wrote a reproducing test, implemented the fix, then stashed the fix to confirm the bug came back without it — all in one pass, unprompted.

That last move is the tell. Verifying a regression by reverting your own change is what a careful engineer does, not what a model does to satisfy a prompt. A model that checks its own work without being asked is one you can leave running on brownfield code — race conditions, hidden tests, the parts of a codebase nobody volunteers to touch — and trust to trace a failure to its root cause instead of patching the symptom. That is the difference between a model that drafts and one that finishes.

Safer overall, with a cyber caveat

On Anthropic's pre-deployment evaluations, Sonnet 5 is an improvement over 4.6: better at refusing malicious requests, more resistant to prompt-injection hijacks, and showing lower rates of hallucination and sycophancy. On the automated behavioral audit it scored safer overall than 4.6 — though somewhat less safe than the more capable Opus 4.8 and the Mythos Preview, a useful reminder that "more capable" and "safer" do not move in lockstep.

The cyber story is the one to read carefully. Sonnet 5 was not trained for cybersecurity tasks and has substantially weaker offensive-cyber ability than the Opus models. On a Firefox exploit-development evaluation built with Mozilla, neither Sonnet model produced a working exploit (both scored 0.0%); Sonnet 5 showed a slightly higher partial success rate than 4.6, which Anthropic attributes to general-intelligence gains rather than targeted training. Because it is marginally stronger here, it ships with cyber safeguards on by default — the same ones used for Opus 4.7 and 4.8. For cyber work that needs reduced guardrails, Anthropic still points you to Opus 4.8.

How I would use it

My read: Sonnet 5 becomes the default execution layer, and Opus 4.8 becomes the deliberate escalation. Route the bulk of multi-step agentic work — coding, tool use, retrieval, automation — to Sonnet 5 at a moderate effort level, and reserve Opus 4.8 (or Sonnet 5 at xhigh) for the tasks where the last few points of accuracy pay for themselves. The effort dial means you can make that call per task instead of per system. Benchmark on your own workload before you commit — but read the launch announcement and the system card it links to first, because the interesting claim here is not a new ceiling. It is a smoother floor.