Claude Sonnet 5 for the agent loop: cheaper tokens, more of them

What it is

Claude Sonnet 5 is Anthropic's new default Sonnet model, positioned as its most agentic mid-tier release yet. Anthropic's Claude Sonnet 5 nears Opus 4.8 on agentic benchmarks and ships with introductory API pricing at $2/$10 per million tokens through August. The company calls Sonnet 5 its "most agentic Sonnet model yet," and in benchmark terms, its performance is close to Opus 4.8 and represents a notable improvement over Sonnet 4.6. It's a hybrid reasoning model with a 1M context window, callable as claude-sonnet-5 on the Claude API.

Quick correction on the brief: introductory pricing is $2/$10, not $3/$15 — that's the post-August standard rate.

The next-work-session test

You have an agent loop in production on Sonnet 4.6 — say, a code-review agent that pulls a diff, calls read_file / grep / run_tests tools in a loop, then writes a summary. The next work session is: swap the model string, re-run your eval harness, look at the cost-per-successful-trajectory column, decide whether the tool-call chain gets shorter (fewer retries, fewer clarifying turns) or just cheaper per token. That's the actual test. This is not an API change: requests, responses, and streaming events keep the same shape, and no code changes are required.

If your evals already exist, this is a one-afternoon experiment. If they don't, build them first — this model changes enough variables that vibes won't tell you.

Pricing

Verified via Anthropic's launch post and platform docs.

• Introductory (through Aug 31, 2026): $2 / 1M input tokens, $10 / 1M output tokens. Claude Sonnet 5 is available everywhere today at an introductory price of $2 per million input tokens and $10 per million output tokens through August 31, 2026.
• Standard (from Sept 1, 2026): $3 / 1M input, $15 / 1M output. It then moves to standard pricing at $3 per million input tokens and $15 per million output tokens.
• The catch nobody puts on the pricing page slide: Claude Sonnet 5 uses a new tokenizer. The same input text produces approximately 30% more tokens than on Claude Sonnet 4.6. Combine that with the September step-up and a "cost-neutral" migration can quietly become materially more expensive per task.

What we'd actually use it for

The narrow, honest use: replacing Sonnet 4.6 in agent loops where the bottleneck is tool-use reliability, not raw generation. Longer-horizon coding agents, retrieval loops with 5+ tool calls, anything where a single bad tool argument cascades into wasted turns. That's where the "more agentic" claim, if it holds in your evals, actually shows up as fewer tokens spent recovering from mistakes.

We would not rewrite our Opus 4.8 pipelines to save money on this until we've A/B'd on our own eval set — the "close to Opus 4.8" framing is Anthropic's, and on hard reasoning it's still their smaller model.

Limits

• Tokenizer inflation. The 30% token increase is real and applies to both prompt and output billing. Your existing cost dashboards will lie to you for a week if you don't recalibrate.
• Price cliff on Sept 1. Anything you commit to at $2/$10 economics needs to still pencil at $3/$15. Sonnet 5 is a real price-performance step forward. It also carries two cost mechanics that won't show up in a quick read of the launch post: a price step-up on September 1 and a new tokenizer that can turn the same text into up to 35% more billable tokens.
• Still not Opus. Unlike with some previous Sonnet launches, it doesn't quite outperform the most recent Opus release on the hardest tasks.
• Context ≠ recall. The 1M window is available; whether your agent actually uses the middle of it well is an eval question, not a spec-sheet answer.
• Migration ergonomics. Change the model ID and you're done — but token budgets, max-output caps, rate-limit math, and any prompt-caching heuristics need re-tuning because of the new tokenizer.

Try it if

• You run an agent loop today on Sonnet 4.6 and have an eval harness to measure change.
• Your workload is tool-heavy and you're paying for retries and recovery turns.
• You want Opus-adjacent quality for coding/tool-use tasks without Opus-tier spend.
• You can complete a migration + benchmark before Aug 31 to lock in real cost data.

Skip it if

• You budgeted on nominal per-token price and can't rerun cost models with a 30% tokenizer adjustment. See the platform pricing docs before you commit.
• You're on Sonnet 4.6 and have no evals — you'll be flying blind on whether "more agentic" is true for your loop.
• Your pipeline is Opus 4.8 for a reason (deep reasoning, long chains) — downgrading to save money without measuring is how regressions ship.
• You need a stable price to forecast Q4 spend; wait until post-Sept 1 rates settle in your billing.

Anthropic's own migration notes are the honest starting point: what's new in Sonnet 5. Read the tokenizer section twice.