Claude Sonnet 5 Explained: What's New, Pricing, and Whether You Should Switch
Introduction
On June 30, 2026, Anthropic released Claude Sonnet 5 and made it the default model for every Free and Pro user starting the next day. Anthropic's most capable model is still Claude Opus 4.8. Sonnet 5 costs far less to run, and it is capable enough for serious agentic work. This is why the release matters: a much cheaper model that performs close to the flagship.
This post walks through what has actually changed under the hood, what it costs, and the situations where it makes sense to move up to Claude Opus 4.8 instead. Nothing here is speculation about the model's internals: the technical claims are drawn directly from Anthropic's own announcement and API documentation, with anything less certain clearly labeled as such.
What This Means If You Just Use Claude.ai
Most of this post is written for developers, because that is where the concrete behavior changes actually live. But the headline change, becoming the default model for every Free and Pro user, affects far more people than that. If you just chat with Claude in a browser or the app, here is the short version.
Two things you might actually notice. First, responses can occasionally take a little longer, because adaptive thinking means the model decides for itself when a question deserves more reasoning before answering, rather than always answering instantly. That tradeoff generally shows up as better answers on harder questions at the cost of a bit more latency on some of them. Second, Claude's knowledge is meaningfully fresher: Anthropic lists Sonnet 5's reliable knowledge cutoff as January 2026, compared to August 2025 for Sonnet 4.6, so it should be noticeably better informed about anything that happened in the back half of 2025 and into 2026.
Problem Statement: The Cost-Versus-Capability Tension
Every team building on top of an LLM eventually runs into the same tradeoff. The most capable model is usually the most expensive, and running every request through it, including simple ones, burns budget for no real gain. The cheapest model is usually the least reliable on multi-step, tool-using, agentic tasks, which is exactly the kind of workload more products are built around today.
Sonnet-class models exist specifically to sit in the middle of that tradeoff: priced well below the flagship tier, but capable enough that most production workloads do not actually need to reach for the most expensive model. Claude Sonnet 5 is Anthropic's latest attempt to push that middle tier closer to flagship-level capability without moving the price.
Core Concepts and Terminology
| Term | Plain English Definition |
|---|---|
| Adaptive thinking | A mode where the model decides on its own, per request, whether and how much to "think" before answering, rather than the developer manually setting a fixed thinking budget. |
| Effort parameter | A setting that lets a developer nudge how much reasoning effort adaptive thinking applies to a request, replacing the old manual thinking-budget system. |
| Tokenizer | The component that converts raw text into the numeric tokens a model actually processes. Changing it changes how many tokens the same sentence costs to send or generate. |
| Context window | The maximum amount of text (measured in tokens) a model can hold in a single conversation or request at once. |
| Agentic task | A task where the model has to plan, call tools, read results, and decide on next steps across multiple turns, rather than answering in a single pass. |
| Zero data retention (ZDR) | A contractual arrangement where a provider does not retain a customer's input or output data after processing the request. |
What's Actually New in Sonnet 5
Anthropic describes Claude Sonnet 5 as a drop-in upgrade for Claude Sonnet 4.6: the same API model naming pattern, the same tools and platform features, and no required code changes beyond swapping the model ID. Underneath that simple framing, three things genuinely changed.
1. A 1 million token context window, by default and by maximum. There is no smaller context variant to opt into. This is large enough to hold a sizeable codebase or a long document collection in a single request. Max output is capped at 128k tokens.
2. Adaptive thinking is on by default, and manual thinking budgets are gone. On Sonnet 4.6, a
request with no thinking field simply ran without thinking. On Sonnet 5, that same request now runs
with adaptive thinking automatically. Manually setting a fixed thinking budget
(thinking: {"type": "enabled", "budget_tokens": N}), which was already deprecated on Sonnet 4.6, is
fully removed on Sonnet 5 and returns an error. The replacement is the effort parameter, used alongside
thinking: {"type": "adaptive"}. In practice, migrating a request looks like this:
# Updating the model ID
model = "claude-sonnet-4-6" # Before
model = "claude-sonnet-5" # After
# Manual thinking budgets are no longer accepted on Sonnet 5
thinking = {"type": "enabled", "budget_tokens": 32000} # Returns a 400 error
# Adaptive thinking replaces it
thinking = {"type": "adaptive"}
On Sonnet 5, the effort parameter defaults to high when called through the Claude API or Claude Code;
set it explicitly if you want a different reasoning level. Anthropic confirms this same high default
for Opus 4.8 on claude.ai too, but its documentation doesn't say whether that also holds for Sonnet 5 on claude.ai.
3. A new tokenizer that produces about 30% more tokens for the same text. This is not a
cosmetic change. Anthropic's own migration notes are direct about the consequences: token counts measured against
Sonnet 4.6 do not carry over, the 1M token context window holds less actual text than the number implies, and
max_tokens limits tuned for Sonnet 4.6 may truncate equivalent output on Sonnet 5. None of this shows
up as an API shape change, so it is easy to miss until a budget or output limit quietly starts behaving differently.
A smaller but practically important change: setting temperature, top_p, or
top_k to a non-default value now returns an error, and assistant message prefilling remains
unsupported (a constraint carried over from Sonnet 4.6). Sonnet 5 is also the first Sonnet-tier model with
real-time cybersecurity safeguards: requests touching prohibited or high-risk cybersecurity topics may be refused,
and that refusal comes back as a normal HTTP 200 response with stop_reason: "refusal", not an error
code, which is worth handling explicitly rather than treating as a failure.
Sonnet 4.6 vs. Sonnet 5 vs. Opus 4.8 at a Glance
Anthropic's own models comparison table makes the tradeoffs concrete. One number worth pausing on: Sonnet 4.6 and Sonnet 5 both advertise a 1M token context window, but because of the new tokenizer, that window does not hold the same amount of real text. Anthropic's own token-to-word estimates put Sonnet 4.6's 1M tokens at roughly 750,000 words, versus roughly 555,000 words for the same 1M tokens on Sonnet 5, a concrete way to see the "30% more tokens for the same text" change in practice.
| Spec | Sonnet 4.6 | Sonnet 5 | Opus 4.8 |
|---|---|---|---|
| Pricing (input / output per million tokens) | $3 / $15 | $3 / $15 (intro $2 / $10 through Aug 31, 2026) | $5 / $25 |
| Context window (nominal) | 1M tokens | 1M tokens | 1M tokens |
| Real-world text capacity of that window | ~750,000 words | ~555,000 words | ~555,000 words |
| Max output | 128k tokens | 128k tokens | 128k tokens |
| Adaptive thinking | Supported, not on by default | Supported, on by default | Supported |
| Manual extended thinking | Supported | Removed (400 error) | Not supported |
| Reliable knowledge cutoff | August 2025 | January 2026 | January 2026 |
| Comparative latency | Fast | Fast | Moderate |
Benchmarks: How Much Better Is It, Really?
Anthropic's own materials state plainly that the largest capability gains over Sonnet 4.6 are in coding and agentic tasks, and point to the company's Transparency Hub for detailed evaluation results. At the time of writing, that public benchmark page had not yet been updated with Sonnet 5's figures, so this post is not going to present precise percentage scores as confirmed Anthropic numbers.
What can be said with more confidence: independent trackers and early hands-on write-ups that appeared after launch consistently describe Sonnet 5 as closing much of the gap to Opus 4.8 on coding and terminal-based agentic benchmarks specifically, with the largest single jump over Sonnet 4.6 showing up on agentic, multi-step tool-use evaluations rather than single-turn knowledge tests. If you need exact, citable numbers for a decision that matters (a procurement comparison, a technical write-up, a benchmark claim in your own documentation), check Anthropic's Transparency Hub directly rather than relying on secondary summaries, since third-party figures published in the first days after a release are the most likely to disagree with each other and with the eventual official numbers.
Pricing and Availability
Per-token pricing on Sonnet 5 is unchanged from Sonnet 4.6: $3 per million input tokens and $15 per million output tokens at standard rates. An introductory rate of $2 per million input tokens and $10 per million output tokens is in effect through August 31, 2026. Because the new tokenizer produces more tokens for the same text, the effective cost of an equivalent real-world request can still shift even though the listed per-token price has not.
Sonnet 5 is the default model for Free and Pro plans as of July 1, 2026, and is also available on Max, Team, and
Enterprise plans, in Claude Code, and through the Claude API to all customers. It is available through Claude in
Amazon Bedrock and Claude Platform on AWS (though not through the legacy Bedrock InvokeModel and
Converse APIs), through Claude on Google Cloud Vertex AI, and through Claude in Microsoft Foundry. It
supports zero data retention for organizations with a ZDR agreement. One notable gap: Sonnet 5 is not available on
Priority Tier at launch.
Should You Switch? Practical Guidance
If you are already on Sonnet 4.6, migration is genuinely close to a drop-in model ID swap, but do not skip the
token-budget review. Recount your prompts under the new tokenizer, revisit any max_tokens limit that
was sized close to your expected output length, and remove any non-default temperature,
top_p,
or top_k settings before switching, since those now return an error instead of being silently applied.
If you are deciding between Sonnet 5 and Opus 4.8 for a new project: Sonnet 5 is the reasonable default starting point for most agentic and coding workloads given the price difference, and you move up to Opus only when you have evidence that Sonnet is the bottleneck on a specific task class, rather than assuming the flagship model first. If your workload is genuinely simple, single-turn, and cost-sensitive at very high volume, it's still worth checking whether a smaller model in Anthropic's lineup is a better fit than either Sonnet or Opus.
Frequently Asked Questions
Is Claude Sonnet 5 a completely new model architecture, or an upgrade?
Anthropic positions it as a drop-in upgrade for Claude Sonnet 4.6 rather than a separate model family: same API model naming pattern, same tools and platform features, and no required code changes beyond updating the model ID string. The meaningful differences are behavioral (adaptive thinking on by default, manual thinking budgets removed) and infrastructural (new tokenizer, 1M token context window), not a change in how you integrate with the API.
Does Claude Sonnet 5 cost more than Claude Sonnet 4.6?
The listed per-token price is identical: $3 per million input tokens and $15 per million output tokens at standard rates (with a temporary introductory discount through August 31, 2026). However, because the new tokenizer produces roughly 30% more tokens for the same input text, the actual cost of processing the same real-world request can be higher even though the per-token rate has not changed. Recount your typical prompts before assuming costs are flat.
Can I still manually set a thinking budget on Claude Sonnet 5?
No. Manual extended thinking, where a developer sets a fixed budget_tokens value, was already
deprecated on Claude Sonnet 4.6 and is fully removed on Claude Sonnet 5, returning a 400 error if attempted. The
replacement is adaptive thinking combined with the effort parameter, which lets the model decide how much reasoning
a given request needs rather than the developer fixing a budget in advance.
What happens if I set a custom temperature or top_p value on Claude Sonnet 5?
Setting temperature, top_p, or top_k to any non-default value now returns a
400 error, a constraint that was already in place on Claude Opus 4.7 and now extends to Sonnet-tier models for the
first time. If you rely on sampling parameter tuning for a specific use case, that has to be redesigned around
system-prompt instructions instead before migrating.
References
- Anthropic. (2026). Introducing Claude Sonnet 5. Official product announcement, June 30, 2026. Available at anthropic.com/news/claude-sonnet-5.
- Anthropic. (2026). What's new in Claude Sonnet 5. Claude Platform Documentation, migration and behavior-change reference. Available at platform.claude.com/docs.
- Anthropic. Transparency Hub. Anthropic's public repository of model system cards and evaluation results, referenced by Anthropic as the source for detailed Sonnet 5 benchmark figures. Available at anthropic.com/transparency.
- Anthropic. Models Overview. Claude Platform Documentation, comparison table of specs and pricing across current and legacy Claude models. Available at platform.claude.com/docs.
Key Takeaways
- If you're on the Free or Pro plan, no action is needed: you were switched to Claude Sonnet 5 automatically on July 1, 2026, and its reliable knowledge cutoff (January 2026) is meaningfully fresher than Sonnet 4.6's (August 2025).
- Claude Sonnet 5 launched June 30, 2026, and became the default model for Free and Pro users on July 1, 2026, while remaining available on Max, Team, Enterprise, and via the API.
- It is a drop-in upgrade for Sonnet 4.6: a 1M token context window (default and maximum, no smaller variant), 128k max output tokens, and no required code changes beyond the model ID.
- Adaptive thinking is on by default, and manual thinking budgets are removed entirely, replaced by an effort parameter used alongside adaptive thinking.
- A new tokenizer produces roughly 30% more tokens for the same text; concretely, Sonnet 5's 1M token window holds about 555,000 words versus about 750,000 words on Sonnet 4.6's same-sized window.
- Standard pricing is $3/$15 per million input/output tokens (unchanged from Sonnet 4.6), with an introductory $2/$10 rate through August 31, 2026. Opus 4.8, by comparison, is priced at $5/$25.
- Precise benchmark percentages were not yet published on Anthropic's own Transparency Hub at the time of writing; treat early third-party benchmark numbers as provisional until confirmed there.
