Claude Sonnet 5 Makes Model Routing the Real Decision

Anthropic released Claude Sonnet 5 on June 30, pitched as the most agentic Sonnet so far, able to make plans, use tools like browsers and terminals, and run on its own, with performance Anthropic describes as close to Opus 4.8 at lower prices. It is now the default model on the Free and Pro plans, available on Max, Team and Enterprise, and on the API as claude-sonnet-5. If you run coding agents all day, this release mostly changes one decision: which work still needs the flagship model.

What was actually announced#

• Price at launch: $2 per million input tokens and $10 per million output tokens. TechCrunch reports this holds through August 31, after which it is scheduled to move to $3 and $15.
• Benchmarks: TechCrunch cites 63.2% on one agentic coding benchmark, against 69.2% for Opus 4.8 and 58.1% for Sonnet 4.6. On a knowledge-work benchmark, Sonnet 5 slightly outperforms Opus 4.8.
• Effort levels: Anthropic says Sonnet 5 gives substantially improved cost efficiency at medium effort, and at higher effort it can match Opus 4.8 on some tasks. Rate limits went up across Chat, Cowork, Claude Code and the Claude Platform to cover the extra tokens that higher effort uses.
• Copilot: GitHub made it generally available the same day for Pro, Pro+, Max, Business and Enterprise. Business and Enterprise admins turn it on in the Copilot model policy settings, and usage is billed at provider list pricing.

Routing now has two settings: model and effort#

I already route work across models by cost. Until now my question was mostly "Sonnet or Opus?" With Sonnet 5, the effort level matters as much as the model, and the two overlap. Opus still leads on agentic coding by six points on the benchmark TechCrunch cites, but Sonnet 5 is slightly ahead on knowledge work. A blanket "Opus for anything important" rule now overpays, and "Sonnet for everything" will likely fall short on the hardest tasks.

The split I plan to test:

• Sonnet 5, medium effort: scoped bug fixes that come with a failing test, refactors inside one module, migration boilerplate, docs, first-pass review comments.
• Sonnet 5, high effort: the same tasks when medium effort fails, before switching models.
• Opus 4.8: long multi-file changes where a wrong plan is expensive, architecture decisions, and code on security-sensitive paths.

Escalation needs a trigger you can check, not a habit: two failed test runs, a plan that touches more modules than the ticket suggests, or a reviewer FAIL. "This feels important" is not a trigger.

The tokenizer footnote cuts into the discount#

Footnote 2 of Anthropic's post matters more than the benchmark numbers: Sonnet 5 uses an updated tokenizer, similar to the change that came with Opus 4.7, and the same input can map to roughly 1.0 to 1.35 times as many tokens, depending on content type.

Text that came to 1M input tokens on Sonnet 4.6 can come to as much as 1.35M on Sonnet 5. At the launch price, that costs $2.70 instead of $2.00. At the $3 rate TechCrunch says follows August 31, it costs up to $4.05. The footnote is about input; until I measure it, I'd assume output behaves the same way.

It also affects more than the bill:

• A system prompt that fit under your context budget or compaction threshold may not anymore.
• max_tokens caps tuned on the old model can start cutting off answers.
• Per-token prices across models now compare different units; compare cost per completed task.

The ratio depends on content type, so measure your own traffic (code, stack traces, JSON and, in my case, Turkish text): run the same real requests through both models and compare reported usage.

A checklist for this week#

• Pin model IDs. Sonnet 5 became the default for Free and Pro users in the Claude apps. For anything automated, name the exact model in config so a switch happens because you chose it.
• Decide on Copilot on purpose. Admins have a new model to enable or leave off, and at provider list pricing the tokenizer change should show up there too.
• Re-run your eval set. On launch day Anthropic replaced its BrowseComp chart because the original used a simpler methodology that underestimated Sonnet 5. A fair correction, and a reminder that charts depend on method.
• Test security tooling. Sonnet 5 ships with cyber safeguards on by default, the same ones used for Opus 4.7 and 4.8. They detect and block dangerous cyber usage in real time. If your pipeline does legitimate security work, check what gets blocked before you rely on it.

What I'm doing about it#

I run several coding agents in parallel every day, so for me this is a routing-table change: Sonnet 5 at medium effort as the default worker, Opus 4.8 for planning and for escalations that meet the triggers above. I also want to try Sonnet 5 as one of the reviewers in Quorate. Quorate applies deterministic PASS/WARN/FAIL gates, so a cheaper reviewer changes who looks at a diff, not the thresholds it has to clear. Before I believe the price cut, I'll count tokens on my own fixtures, and check again before August 31, when the price is scheduled to change.

Update, August 10: Anthropic later made the $2/$10 introductory price permanent, so the $3/$15 increase described above no longer applies.