Anthropic released Claude Sonnet 5 on June 30, 2026. As of today it's the default model for all claude.ai users — free plan, Pro, Team. On the API it's available immediately at $2 per million input tokens and $10 per million output tokens, introductory pricing that holds through August 31.
For context: Claude Sonnet 4.6 was $3/$15 per million tokens. Claude Sonnet 5 is a more capable model at 33% less cost — for now. Whether to act on that window depends on what you're running, which is worth working through.
What's new in Sonnet 5
Sonnet 5 is a drop-in replacement for Sonnet 4.6 in terms of API shape, with three behavior changes and one engineering change worth knowing:
- Adaptive thinking is on by default. The model now reasons through problems before answering, automatically. On Sonnet 4.6 you had to opt in; on Sonnet 5 you opt out if you don't want it.
- Sampling parameters aren't accepted. Setting
temperature,top_p, ortop_kto non-default values returns a 400 error. Remove them when migrating; use system-prompt instructions to shape outputs instead. - Manual extended thinking is removed. Use adaptive thinking with the
effortparameter instead. - New tokenizer. The same input text produces approximately 30% more tokens than on Sonnet 4.6. More on this below — it matters for your cost math.
The capability improvements are largest in coding and agentic tasks. Sonnet 5 scores 63.2% on agentic coding benchmarks, up from Sonnet 4.6's 58.1%, and approaching Opus 4.8's 69.2% — at less than half the cost. For customer service routing, lead qualification, content drafts, and knowledge-base retrieval, it handles what most businesses need.
The four-tier breakdown
The Claude lineup now has four practical tiers, each priced for a different complexity level:
| Model | Input / Output (per MTok) | Context | Best for |
|---|---|---|---|
| Claude Haiku 4.5 | $1 / $5 | 200k | High-volume, simple classification, routing, short responses |
| Claude Sonnet 5 | $2 / $10 (intro)* | 1M | Most business workflows — drafting, analysis, lead qual, agents |
| Claude Opus 4.8 | $5 / $25 | 1M | Complex multi-step reasoning, long-horizon agentic coding |
| Claude Fable 5 | $10 / $50 | 1M | Frontier capability; justified only for the hardest tasks |
*Introductory pricing through August 31, 2026. Standard price is $3/$15 per MTok.
The decision tree is simpler than the number of options suggests. Haiku is for tasks where latency and cost dominate — classification, simple routing, templated responses at scale. Sonnet 5 handles most real business use cases. Opus 4.8 is for genuinely complex reasoning tasks where the 6% benchmark gap between it and Sonnet 5 shows up in practice. Fable 5 is a specialist tool, not a default.
Sonnet 5 scores 63.2% on agentic coding — close to Opus 4.8 (69.2%) for most workflows, at less than half the price.
The tokenizer gotcha
The new tokenizer is the most important operational detail in the Sonnet 5 release, and it's easy to miss.
Claude Sonnet 5 converts text to approximately 30% more tokens than Sonnet 4.6 did. The per-token price is the same, but the same block of text now requires 30% more of them. During the promotional window, the discount more than compensates: at $2/$10 versus Sonnet 4.6's $3/$15, you're paying roughly 13% less per equivalent text request. That math holds through August 31.
After the promo ends, the situation reverses. At the standard rate of $3/$15 — the same nominal per-token price as Sonnet 4.6 — the same input text costs about 30% more to process on Sonnet 5. The price tag looks identical on paper. The actual bill isn't.
The practical action: recount your prompt tokens against Sonnet 5 before you budget anything. Counts measured against Sonnet 4.6 don't transfer. If your agents process long documents, CRM notes, or email threads, the token delta compounds quickly.
Running the cost math
We've covered the full framework for what an AI agent actually costs to run, so I won't repeat all of it here. But Sonnet 5 at introductory pricing changes the numbers enough to run a quick example.
Take a lead-qualification agent handling 10,000 conversations per month, processing 2,000 input tokens and generating 500 output tokens per conversation:
- Sonnet 5 at intro pricing ($2/$10): (10K × 2K × $2/M) + (10K × 500 × $10/M) = $40 + $50 = $90/month
- Sonnet 5 at standard pricing ($3/$15): $60 + $75 = $135/month
- Opus 4.8 ($5/$25): $100 + $125 = $225/month
The $90 number assumes you've already accounted for the tokenizer — if those 2,000 tokens are measured against Sonnet 4.6, the real Sonnet 5 equivalent is closer to 2,600 input tokens, putting the intro-price total around $105/month. Still a clear win versus Opus 4.8 for this kind of task.
The point isn't to obsess over $15. It's that at any of these price points, a well-scoped agent pays for itself inside a month for most businesses once you account for the time it replaces.
Most businesses overshoot on model tier. The right model is the cheapest one that reliably handles the task — and right now, Sonnet 5 is a very good answer to that question.
What to do before August 31
If you're running Sonnet 4.6 workflows in production, migration is worth doing before the promotional window closes. The steps:
- Update your model ID from
claude-sonnet-4-6toclaude-sonnet-5. - Remove any
temperature,top_p, ortop_kparameters. - Recount your prompt tokens using Anthropic's token-counting API against the new model.
- Revisit any
max_tokenslimits sized close to your expected output — the new tokenizer means the same response may consume more of your limit than before.
If you're evaluating AI automation for the first time, this is a better-than-usual entry point. Two months at promotional pricing is enough time to build a real workflow, measure its results, and understand your ongoing costs before rates normalize. Our AI automation services start with an audit that sizes the right model to your actual workload — not the most expensive one.
And if you're deciding between Sonnet 5 and Opus 4.8 for something more complex, the Opus 4.8 breakdown we published in May is still accurate on capability — the pricing comparison has shifted in Sonnet 5's favor, but the use-case framing holds.
The promotional window closes August 31. After that, Sonnet 5 is still a good model at a fair price — just not quite as easy to say yes to from a standing start.
— Cole