Claude Sonnet 5 Is Here — But That ‘Same Price’ Claim Hides a 30% Cost Surprise
Claude Sonnet 5 launched on 30 June 2026 with near-Opus 4.8 performance, a 1 million token context window, and adaptive thinking enabled by default — but a new tokenizer makes English-language inputs roughly 30 to 42 percent more expensive in real terms than the headline price suggests. Simon Willison's analysis documents the tokenizer gap empirically, and Indian professionals using Claude-powered tools should audit their costs before the introductory discount expires on 31 August 2026.
Claude Sonnet 5 launched on 30 June 2026. If you work with Claude through the API, through a product built on it, or simply pay attention to which AI model powers your day-to-day work, this release matters — and not entirely for the reasons Anthropic’s announcement emphasises.
Simon Willison’s analysis at simonwillison.net cuts straight to the developer documentation rather than the marketing copy, and his findings reveal a release that is more complicated than it first appears. This post unpacks what changed, what it really costs, and what it means for non-technical professionals in India who rely on Claude-powered tools.
What Anthropic Is Claiming
According to Anthropic, Sonnet 5’s performance is “close to that of Opus 4.8, but at lower prices.” That is the headline. Opus 4.8 is Anthropic’s frontier model — the most capable version available. If Sonnet 5 truly matches that capability at a lower price point, it would be a significant step forward for anyone using Claude through the API or through third-party products.
Anthropics system card for Sonnet 5 also addresses why the model could be released without regulatory delay. The relevant passage notes that Sonnet 5 is significantly less capable at cyber tasks than a more powerful internal model called Mythos 5, which means its safeguards sit in the same tier as Opus 4.7 and Opus 4.8. That context matters if you work in legal, finance, or any regulated sector where AI governance is becoming part of the compliance conversation.
The API Changes That Actually Affect Your Work
For developers integrating Claude into products, the “what’s new” section of Anthropic’s developer documentation contains several notable changes that Simon Willison highlights.
Sampling Parameters Are Gone
Temperature, top_p, and top_k — three settings that developers traditionally used to control how predictable or creative a model’s responses are — are no longer supported in Sonnet 5. If you are a non-technical user, this will not affect you directly. But if your company uses a custom Claude integration built by a developer, it is worth knowing that any product relying on those controls will need updating. Outputs may feel different in tone or consistency until the underlying integration is adjusted.
A Vastly Larger Context Window
Sonnet 5 supports a 1 million token context window with up to 128,000 maximum output tokens. To make that concrete: a 1 million token context can hold roughly 700 to 750 pages of dense English text all at once. This is the kind of capacity that allows a CA firm in Pune to feed an entire year’s worth of client correspondence, ledger notes, and audit trails into a single Claude session and ask questions across all of it without losing context between messages.
Adaptive Thinking Is On by Default
Sonnet 5 enables adaptive thinking automatically unless you specifically turn it off. Adaptive thinking means the model allocates more computational effort to harder problems before responding. For most users this is a net positive — responses to complex questions should be more considered. Developers who want faster, lighter responses for simpler tasks will need to explicitly disable this in their API calls.
The Pricing Picture — Read This Carefully
This is where the story gets interesting, and where Willison’s analysis is most valuable to share with a broader audience.
At face value, Sonnet 5 is priced the same as its predecessor Sonnet 4.6: $3 per million input tokens and $15 per million output tokens (approximately ₹255 and ₹1,275 per million tokens respectively at current rates). There is also an introductory discount running until 31 August 2026 that drops those rates to $2 and $10 per million tokens (roughly ₹170 and ₹850).
However, Sonnet 5 uses a new tokenizer. According to Anthropic’s own documentation, “the same input text produces approximately 30% more tokens than on Claude Sonnet 4.6.” Simon Willison put this to the test using his Claude Token Counter tool, running several real-world documents through both the old and new tokenizer. His results:
- The Universal Declaration of Human Rights in English: 1.42 times more tokens on Sonnet 5 versus Sonnet 4.6
- The same document in Spanish: 1.33 times more tokens
- A large Python source file: 1.28 times more tokens
- The same document in Simplified Mandarin Chinese: effectively the same token count
What this means in plain language: if you work primarily in English, you are not paying the same price as before. You are paying approximately 30 to 42 percent more for the same amount of text, because the new tokenizer counts that text as more tokens. The introductory discount partially offsets this during the discount window, but after 31 August 2026 the effective cost increase for English-language tasks is real.
For Indian professionals who work in Hindi, Tamil, Telugu, Bengali, or other Indic languages, the tokenizer impact is not yet clearly documented. Willison’s data shows that Mandarin Chinese is barely affected, which suggests languages with denser scripts may fare better. However, most Indian-language content also involves a significant amount of English mixing, so the impact is unlikely to be zero.
A Concrete Indian Scenario: A Legal Research Team in Delhi
Consider a legal research team at a mid-sized law firm in Connaught Place, Delhi. They have been using a Claude-powered document review tool to process contracts, case summaries, and regulatory filings. Their vendor charges per API call, and those costs pass through to the firm’s monthly invoice.
With Sonnet 5, two things change simultaneously. First, the model’s reasoning quality improves — closer to Opus 4.8 performance means more accurate clause extraction, better summarisation of complex judgments, and fewer obvious errors that require a paralegal to manually correct. Second, the tokenizer change means that every contract they upload now registers as roughly 30 to 40 percent more tokens than it did under Sonnet 4.6.
If the firm was spending ₹40,000 per month on AI-assisted document review, they should expect that figure to rise to somewhere between ₹52,000 and ₹56,000 once the introductory discount period ends — assuming the same volume of work and the same document mix. That is not a small change for a firm that treats AI tooling as a cost-controlled line item.
The right question to ask your vendor between now and August 2026 is: are you passing on the introductory discount, and what is your pricing plan after it expires?
What Sonnet 5 Still Cannot Do (And What Remains Unclear)
Willison’s post notes that Sonnet 5 “features the same set of tools and platform features as Claude Sonnet 4.6,” meaning there are no new tool integrations or expanded capabilities beyond what already existed. It is a capability and efficiency upgrade, not a feature expansion.
The removal of temperature, top_p, and top_k controls is a genuine constraint for developers who relied on those parameters for specific output styles — for instance, teams generating structured data at low temperature for reliability, or teams needing higher temperature for creative brainstorming use cases.
There is also no region-specific availability information in the sources reviewed here. Enterprise and API access policies for India have historically followed global rollout timelines with occasional delays, so teams relying on Claude through regional cloud deployments should verify availability directly with their vendor or through Anthropic’s documentation.
What to Watch Between Now and September 2026
The 31 August 2026 introductory discount expiry is the most immediate date to track. Before that deadline, audit how much English-language text your team or your vendor’s tool is processing through Claude. The tokenizer change from Sonnet 4.6 to Sonnet 5 is not hypothetical — Willison’s empirical testing confirms it is real and material for English content.
If you are evaluating whether Sonnet 5’s quality improvement justifies the effective price increase, the most useful experiment is to run a representative sample of your actual work through the model and compare output quality against Sonnet 4.6. Quality gains that reduce human review time can more than offset a cost increase — but that calculation only works if you measure it for your specific use case rather than assuming the improvement will be uniform.
Anthropics own framing of Sonnet 5 as near-Opus 4.8 performance at lower prices is technically accurate on the published per-token rate. Simon Willison’s analysis at simonwillison.net is the more complete picture: the tokenizer change means the effective price for English text is higher than the headline number suggests, and that is the number your finance team should be planning around.
