Anthropic published Claude Sonnet 5 this morning. The company’s "what's new" developer documentation provides the most actionable details about the release, including technical and safety context that clarifies why the model could be made available.
Capabilities and safety posture
Anthropic says Sonnet 5’s performance is "close to that of Opus 4.8, but at lower prices." The model’s system card notes that Sonnet 5 is significantly less capable on cyber tasks than Mythos 5, and therefore the safety controls applied to Sonnet 5 are similar to those used for Opus 4.7 and Opus 4.8 (models that are stronger than Sonnet 5 but much less capable than Mythos 5).
API changes and technical specs
- Sampling parameters temperature, top_p and top_k are no longer supported.
- Context window: 1,000,000 tokens.
- Maximum output: 128,000 tokens.
- Sonnet 5 offers the same set of tools and platform features as Claude Sonnet 4.6.
- Adaptive thinking is enabled by default; to disable it you must set "thinking": {type: "disabled"} in requests.
Pricing and the new tokenizer’s impact
The published pricing matches Sonnet 4.6: $3/million input and $15/million input, with an introductory discount to $2/$10 until August 31, 2026. However, Sonnet 5 uses a new tokenizer: Anthropic states that the same input text produces approximately 30% more tokens than on Claude Sonnet 4.6. In practice this means roughly a 30% increase in effective cost for the same text under per-token billing.
Measurements: tokenizer differences on sample documents
Using the "Claude Token Counter" tool, the following token counts were observed for several larger documents (Sonnet 4.6 vs Sonnet 5 and the approximate ratio):
- Universal Declaration of Human Rights (English): Sonnet 4.6 = 2,356 tokens; Sonnet 5 = 3,341 tokens (≈1.42×)
- Universal Declaration of Human Rights (Spanish): Sonnet 4.6 = 3,572 tokens; Sonnet 5 = 4,747 tokens (≈1.33×)
- Universal Declaration of Human Rights (Chinese, Simplified Mandarin): Sonnet 4.6 = 3,334 tokens; Sonnet 5 = 3,360 tokens (≈1.01×)
- sqlite_utils/db.py (4,279 lines of Python): Sonnet 4.6 = 44,014 tokens; Sonnet 5 = 56,113 tokens (≈1.27–1.28×)
From these samples, Sonnet 5’s tokenizer yields roughly 1.4× token counts for English, 1.33× for Spanish, ~1.28× for Python code, and effectively the same count for Simplified Mandarin.
Additional notes
The release removes some familiar API knobs (sampling parameters) while delivering a very large context window and high output limits, which may enable new use cases. At the same time, the increased tokenization density can materially raise effective costs compared with Sonnet 4.6 despite unchanged per-token prices during the introductory period.
A minor community anecdote: when shown an image of a pelican, Sonnet 5 described it as looking like a goose — a small example of model output shared by users.
Why this matters
Sonnet 5 combines substantial context and output capacities with platform-level changes that affect developers' integration choices. The new tokenizer’s higher token counts are particularly relevant for organizations processing large amounts of text or code, because they directly influence billing and operational cost even when nominal per-token prices remain the same.
The article is based on Anthropic’s "what's new" developer documentation and the model’s system card.
