Claude Sonnet 5 Ships: Near-Opus Power, Big Token Cost

Anthropic shipped Claude Sonnet 5 on June 30, making it the default model across all Free and Pro plans. Our verdict: recommended. The agentic capability jump from Claude 4 Sonnet is unambiguous — it finishes complex multi-step tasks its predecessor stalled on, checks its own output unprompted, and closes the gap with Opus 4.8 to a razor-thin margin. But budget for a ~30% token-count increase from the new tokenizer, and migrate any code relying on temperature, top_p, or top_k — those parameters are gone.

What happened

Claude Sonnet 5 narrows the performance gap with Opus 4.8 to the slimmest margin yet. At medium effort levels, it delivers the same output quality for less money. At higher effort settings — the kind agents use for sustained multi-step work — it can match Opus 4.8 on agentic search (BrowseComp: 84.7%) and computer use (OSWorld-Verified: 81.2%).

The headline number is a 13.4-point jump on Terminal-Bench 2.1 over Claude 4 Sonnet. Early-access partners echo the benchmarks: Stripe called it "a strong execution layer for multi-step software engineering work." Lovable's co-founder reported "same output quality, fewer steps to get there." One engineer described Sonnet 5 investigating a bug, writing a reproducing test, implementing the fix, then stashing it to confirm the bug came back — all in a single pass, unprompted.

Safety improved too: lower hallucination rates, better refusal of malicious requests, and 0.0% Firefox exploit success. Cyber safeguards are enabled by default at Opus 4.7/4.8 levels — less strict than Fable 5's classifiers, which Anthropic says is appropriate given Sonnet 5's "substantially poorer" offensive cyber capability.

The catch, disclosed honestly by Anthropic in a footnote: a new tokenizer maps the same input to roughly 1.0–1.35× more tokens. Simon Willison's analysis pegged it at ~1.4× for English and ~1.28× for Python.

The per-token pricing looks flat — $2/$10 introductory (through August 31), then $3/$15 standard — the same nominal rates as Claude 4 Sonnet. But when a call that used 5,000 tokens now needs 7,000 tokens, your bill climbs 40% even though the rate card stayed the same. Anthropic says introductory pricing makes the transition "roughly cost-neutral." After August 31, the effective increase lands.

In a clean break, Sonnet 5 no longer supports temperature, top_p, or top_k in the API or Workbench. Anthropic has been moving toward effort-based control (low/medium/high/extra-high) instead of traditional sampling parameters. If your code relies on these, migrate now.

Why it matters

This is the model Anthropic wants enterprises to run agents on. It is the Sonnet-class workhorse — close enough to Opus 4.8 that many workloads won't need the more expensive tier, while cheap enough to keep agent budgets from spiraling. The Claude Sonnet 5 directory verdict is recommended, and this launch reinforces that: the best cost-to-capability ratio in the Claude lineup.

The tokenizer twist is not a dealbreaker, but it is a budget trap for teams scaling agent workloads aggressively. Per-token rates are the wrong metric now — count tokens per task before and after migration.

The triple-ship day — Sonnet 5 + Fable 5 restoration + Claude Science launch — signals that Anthropic's pipeline is firing despite the June 12 export ban disruption. For teams building on Claude, Sonnet 5 is the default now.

What changes for you

If you're on Claude 4 Sonnet, upgrade — the capability jump is real and the introductory pricing makes it cost-neutral through August. If you're running agent workloads at scale, recalculate your token budgets before switching. If your API calls pass temperature, top_p, or top_k, remove them — those parameters are dead on Sonnet 5.

FAQ

Q: Why does the tokenizer change matter if per-token pricing stayed the same? A new tokenizer breaks text into more pieces — roughly 30% more for English input. Your per-call cost is token-count × per-token rate, so if the token count rises 30%, your bill rises 30% even though the rate card looks unchanged. Count tokens per task before and after migration, not just rates.

Q: What replaces temperature, top_p, and top_k? Anthropic's effort-based control: set thinking with effort levels (low/medium/high/extra-high) instead of traditional sampling parameters. This gives you deterministic control over reasoning depth rather than randomness.

Q: Lock in introductory pricing or wait for standard rates in September? Lock in now. The introductory $2/$10 per million tokens saves roughly 33% versus the $3/$15 standard rates. Factor in the tokenizer expansion and the effective cost gap widens further — early adoption with intro pricing is the cheapest path.