Claude Sonnet 5: near-Opus agentic capability at Sonnet pricing

Anthropic shipped Claude Sonnet 5 on June 30 — the first Sonnet-tier model in the new Claude 5 family. Opus-class releases set the frontier, but Sonnet is the tier most production systems actually run, which makes a new Sonnet generation one of the highest-impact releases Anthropic makes. This one is worth reading closely: the headline is near-Opus agentic capability at Sonnet pricing, and the details mostly back it up.

What shipped

The model is available everywhere today — default for Free and Pro plans, in Claude Code, and on the API as claude-sonnet-5. The spec sheet:

• Pricing: $3 per million input tokens and $15 per million output, with introductory pricing of $2/$10 through August 31, 2026. Opus 4.8, for reference, is $5/$25.
• Context and output: a 1M-token context window and 128K max output tokens.
• Effort control: adaptive thinking is on by default, and Sonnet 5 is the first Sonnet-tier model to support the xhigh effort level.
• A new tokenizer: the same change introduced with Opus 4.7 — better performance, but the same input can map to roughly 1.0–1.35× more tokens depending on content.
• Vision: high-resolution image support up to 2576px.

Anthropic's own framing is that Sonnet 5 "can make plans, use tools like browsers and terminals, and run autonomously at a level that, just a few months ago, required larger and more expensive models." On their reported evaluations it is a clear improvement over Sonnet 4.6 on reasoning, tool use, coding, and knowledge work.

The effort dial is the real story

The most useful charts in the announcement are the cost-performance curves on BrowseComp (agentic search) and OSWorld-Verified (computer use). Sonnet 4.6 fell well short of Opus 4.8 no matter how much you spent. Sonnet 5 instead covers a wide band: at medium effort it is substantially more cost-efficient than its predecessor, and at higher effort it matches Opus 4.8 on some tasks.

One model now spans most of the range that previously required routing between two tiers.

That shape matters more than any single benchmark number. In production agentic systems, the model choice used to be binary: Sonnet for volume, Opus for the hard steps. An effort dial on a single model turns that discrete routing decision into a continuous one — same model, same prompts, same tool definitions, just a different effort setting per task class. That is operationally much simpler than maintaining two model configurations.

The fine print

Two things deserve attention before you migrate. First, the tokenizer: because the same input can tokenize to up to 1.35× more tokens, the intro pricing is explicitly set so the transition from Sonnet 4.6 is roughly cost-neutral, not cheaper. Re-measure your token counts on real traffic before treating the $2/$10 window as a discount.

Second, the safety profile is genuinely differentiated by tier. Sonnet 5 shows lower hallucination and sycophancy rates than Sonnet 4.6 and better resistance to prompt injection — both directly relevant if you run it as an agent with tool access. On cybersecurity, Anthropic deliberately did not train it up: it never produced a working exploit in their Firefox evaluation (a test built with Mozilla), and it ships with the same cyber safeguards as Opus 4.7 and 4.8 enabled by default. The system card also notes it scores slightly higher on misaligned-behavior audits than Opus 4.8 — safer than its predecessor, but not the safest model in the family.

My take

I run agentic pipelines where cost per task is a first-class constraint, and this release changes the default answer. For most multi-step tool-use work — coding agents, retrieval agents, workflow automation — Sonnet 5 at medium effort is now the sensible starting point, with xhigh reserved for the steps where you would previously have escalated to Opus. The migration checklist is short: benchmark your token counts under the new tokenizer, A/B your hardest agentic tasks at xhigh against Opus 4.8, and do it before August 31 while the pricing makes the comparison cost-neutral. The interesting question for the rest of 2026 is how much traffic actually stays on Opus once teams internalize the effort dial.