Claude Sonnet 5: Near-Opus Agents at Sonnet Prices — With a Tokenizer Catch

Claude Sonnet 5 shipped on June 30, 2026, and the pitch is simple: near-Opus agentic performance at Sonnet prices. List pricing stays at $3 per million input tokens and $15 per million output — the same sticker as Sonnet 4.6 — with an introductory $2/$10 running through August 31. Anthropic calls it the most agentic Sonnet yet, and it is now the default model on claude.ai Free and Pro plans.

What most launch coverage skips is the tokenizer. Sonnet 5 uses a new tokenizer, and the same input now maps to more tokens than it did on Sonnet 4.6 — up to 42% more for English prose. That one detail decides whether this release saves you money, and the answer splits along a line most people are not checking: which model you are migrating from. Coming from Opus 4.8, the 40% saving is real. Coming from Sonnet 4.6, the same-price story is only true on the price page.

What changed

claude-sonnet-5 is available on the API, Amazon Bedrock, Google Vertex AI, and Microsoft Foundry. The headline specs: a 1M-token context window and 128K max output, extendable to 300K output on the Batch API with the output-300k-2026-03-24 beta header. Otherwise Sonnet 5 keeps the same tool and platform feature set as Sonnet 4.6.

The API surface changed enough that this is a migration, not a model-string swap:

• Adaptive thinking is on by default. On Sonnet 4.6, omitting the thinking parameter meant thinking off. On Sonnet 5, omitting it runs adaptive thinking — a silent behavior and cost change for any caller that never set the param. To turn it off, pass thinking: {type: "disabled"}.
• Manual thinking budgets are gone. Passing budget_tokens returns a 400.
• Sampling knobs are rejected. Non-default temperature, top_p, or top_k return a 400, matching the surface Opus 4.7 and 4.8 already have.
• effort now goes up to xhigh and defaults to high on the API and in Claude Code.

If your integration predates the Opus 4.7 conventions, budget an afternoon with the migration guide rather than five minutes for a config change.

The tokenizer catch

Anthropic's own guidance says the new tokenizer maps the same input to roughly 1.0–1.35x the tokens. Simon Willison measured it against Sonnet 4.6 and got: English 1.42x, Spanish 1.33x, Python 1.27x, Mandarin 1.01x.

Against Sonnet 4.6 the sticker price is unchanged, so those multipliers translate directly into effective cost: roughly 42% more for English prose, 27% more for Python, and near-flat for CJK text.

The intro pricing muddies this in an interesting way. $2/$10 is a 33% discount off list (you pay two-thirds of the old rate), and 0.667 × 1.42 ≈ 0.95 — for English text, the discount and the tokenizer penalty nearly cancel, so Sonnet 5 costs about what 4.6 does through August 31. Code does slightly better: 0.667 × 1.27 ≈ 0.85. Then the discount expires and the penalty stays. From September 1, an English-heavy workload pays roughly 42% more than it would on Sonnet 4.6 for identical text.

Two second-order effects are easy to miss. The 1M context window measured in new tokens holds less actual text — about 555k English words versus roughly 750k under the old tokenizer, per the official docs. And max_tokens limits tuned on 4.6 may now truncate output mid-response. If you track spend per token or alert on token counts, those baselines shift too. The count_tokens endpoint is the cheap way to measure your own corpus before committing.

The Opus comparison has no asterisk

Here is the nuance most coverage misses: the tokenizer penalty is a Sonnet-4.6 problem, not an Opus one. The official docs list Sonnet 5 and Opus 4.8 with the same context measurement — roughly 555k words to a 1M-token window on both — so the same prompt costs the same number of tokens either way. That makes comparing $3/$15 against Opus's $5/$25 apples to apples: the 40% price gap is exactly what you save. For teams running Opus 4.8 on coding agents and tool-use pipelines, that is the cleanest cost story in this launch.

How it compares

Capability-wise, Anthropic says Sonnet 5's performance is close to Opus 4.8, with the biggest gains over 4.6 in reasoning, tool use, coding, and knowledge work. Those are vendor numbers; independent benchmarks were still sparse at launch, so treat near-Opus as a claim to verify on your own workload rather than a settled fact.

Should you switch

From Sonnet 4.6: the capability upgrade is real, and the timing matters. During the intro window, English workloads are roughly cost-neutral and code is slightly cheaper, so migrating now means you absorb the API breaking changes and discover your actual cost delta while the discount covers it. If your pipeline is English-prose-heavy and cost-sensitive, run count_tokens on a representative sample first and budget for roughly 42% higher effective spend on English text after August 31 — then decide whether the quality gain covers it.

From Opus 4.8: step down for mainstream coding and agentic work. The 40% saving carries no tokenizer asterisk, and if near-Opus is enough for your tasks, it is the easy call. Keep Opus where its extra headroom demonstrably pays for itself — the longest autonomous runs, the hardest debugging sessions.

CJK workloads: the best case in this launch. At a ~1.01x multiplier for Mandarin, the tokenizer penalty essentially vanishes, making Sonnet 5 close to a strict upgrade over 4.6 at the same effective price.

Limitations

The performance claims are Anthropic's own; third-party evaluations were still landing at publication time.

Security research is the one clear do-not-step-down. Anthropic's system card reports that neither Sonnet model could develop a working exploit against Firefox 147 (both scored 0.0% full success, versus 8.8% for Opus 4.8), and the company notes substantially poorer performance on advanced cyber tasks than Opus 4.8. Separately, Sonnet 5 is more cyber-capable than 4.6, which means safeguard refusals can now surface on legitimate security-adjacent work that 4.6 handled without friction.

Adaptive-thinking-by-default is a cost consideration, not just a behavioral one: callers who previously omitted thinking now pay for thinking tokens they were not paying for before, and effort defaulting to high compounds that. Anthropic also reports lower hallucination rates and fewer undesirable behaviors than Sonnet 4.6 — likewise self-reported.

References

• Introducing Claude Sonnet 5 — Anthropic announcement, June 30, 2026
• Models overview — Claude platform docs (specs, pricing, intro-pricing footnote)
• Model migration guide — Claude platform docs (API breaking changes, tokenizer)
• Claude Sonnet 5 System Card — Anthropic (cyber-capability evaluations, Firefox 147 exploit results)
• Claude Sonnet 5 — Simon Willison (tokenizer measurements, cost analysis)