Claude Sonnet 5 Launch: Anthropic's Most Agentic Sonnet, Now the Default Tier

Anthropic launched Claude Sonnet 5 on June 30, 2026, calling it "our most agentic Sonnet yet." The pitch in the announcement post is concrete: Sonnet 5 "makes plans, uses tools like browsers and terminals, and runs autonomously at a level that just a few months ago required larger and more expensive models." It is the default model for Free and Pro, available across every plan, live in Claude Code, and addressable on the API as claude-sonnet-5.

The substance is that autonomous agent behavior, which used to be an Opus-tier expectation, has moved down into the default Sonnet tier. The caveats that matter most are pricing (introductory now, standard later), an updated tokenizer that can increase token counts, and the fact that Opus 4.8 still wins on accuracy for the tasks Anthropic references. For the flagship tier this sits beneath, see our Opus 4.8 Fast mode coverage and the Opus 4.7 launch.

Key Takeaways

• Most agentic Sonnet yet. Anthropic positions Sonnet 5 as a model that makes plans, drives browsers and terminals, and runs autonomously, behavior that recently required larger models.
• It is the new default tier. Sonnet 5 is the default model for Free and Pro users, and is available to Max, Team, and Enterprise users as well.
• API ID is claude-sonnet-5. It is available in Claude Code and on the Claude Platform and API.
• Introductory pricing through August 31, 2026: $2 per million input tokens and $10 per million output tokens. Standard pricing afterward is $3 per million input and $15 per million output.
• Close to Opus 4.8, but not above it. Anthropic says Sonnet 5 narrows the gap with Opus-class models at lower prices, while Opus 4.8 remains preferred for higher accuracy on referenced tasks.
• Updated tokenizer. The same input can map to more tokens, roughly 1.0 to 1.35x depending on content type; intro pricing is meant to keep the transition roughly cost-neutral.
• Safer than Sonnet 4.6, with caveats. Lower rates of undesirable behavior, better prompt-injection resistance, and lower hallucination than Sonnet 4.6, but somewhat higher misaligned behavior than Opus 4.8 on an automated audit.

What Is Claude Sonnet 5?

Claude Sonnet 5 is Anthropic's new mid-tier model and, per the launch, its most agentic Sonnet to date. It is built to make plans, call tools like browsers and terminals, and run autonomously through multi-step work rather than stopping for guidance at each turn. Anthropic frames the headline as capability moving down a tier: the kind of autonomous operation that "just a few months ago required larger and more expensive models" now runs on the default Sonnet.

Anthropic describes Sonnet 5 as a substantial improvement over Sonnet 4.6 across reasoning, tool use, coding, and knowledge work. The practical read is that the default model most people touch first, Free and Pro users land on it automatically, is now meaningfully more capable at agentic tasks than the version it replaces, without moving up to the Opus price tier.

Sonnet 5 Versus Opus 4.8: Pick by Accuracy

Use Opus 4.8 when accuracy on the referenced tasks is paramount; use Sonnet 5 when you want most of that capability at a lower price. Anthropic is explicit that Sonnet 5 narrows the gap with Opus-class models and gets close to Opus 4.8 performance at lower prices, but that Opus 4.8 remains the preferred model for higher accuracy on the tasks Anthropic references.

That is the decision rule worth internalizing. Sonnet 5 is not a replacement for the flagship; it is a closer-than-before alternative at a better price point. For high-stakes work where a wrong answer is expensive, the launch language points you back to Opus 4.8. For the broad middle of agentic and coding work, Sonnet 5 is the value pick. The flagship tier's speed story is covered in our Opus 4.8 Fast mode write-up.

Agentic Coding: Finishing What Earlier Sonnets Stopped Short Of

The clearest improvement is in agentic software engineering. Anthropic says Sonnet 5 is more agentic than its predecessors and finishes complex tasks where prior Sonnet models would stop short. It checks its own output without being asked, which makes it a stronger execution layer for multi-step software engineering rather than a single-shot code generator.

The launch calls out sustained coding, tool use, and debugging across long runs: working in brownfield code, finding race conditions, surfacing hidden tests, tracing root causes, and shipping durable fixes rather than surface patches. Anthropic also says it does this in fewer steps, stays on the plan it sets, and follows existing conventions in the codebase. For teams already weighing agentic coding harnesses, our Cursor vs Claude Code comparison is the anchor for where a model like this fits.

Pricing: Introductory Now, Standard After August 31

Anthropic set introductory pricing for Sonnet 5 of $2 per million input tokens and $10 per million output tokens, in effect through August 31, 2026. After that date, standard pricing is $3 per million input tokens and $15 per million output tokens.

The intro rate is not just a launch promotion. Sonnet 5 ships with an updated tokenizer that can map the same input to more tokens, roughly 1.0 to 1.35x depending on content type. Anthropic says the introductory pricing is intended to make the transition from Sonnet 4.6 roughly cost-neutral despite that token-count change. Rate limits have also been increased to account for the higher token usage that comes with higher effort levels. The takeaway for budget planning: measure your own token multiplier on real traffic before standardizing, and remember the standard rate kicks in after August 31.

Benchmarks: A Strict Improvement Over Sonnet 4.6

Anthropic frames Sonnet 5 as a strict improvement over Sonnet 4.6 on the comparisons it published, including BrowseComp and OSWorld-Verified, while noting Opus 4.8 still posts higher accuracy. The launch also published updated Sonnet 4.6 numbers as the baseline Sonnet 5 improves on.

Read these as the floor, not the ceiling: Anthropic presents Sonnet 5 as a strict improvement over these Sonnet 4.6 figures on the referenced evals, with Opus 4.8 sitting above Sonnet 5 on accuracy. The BrowseComp and OSWorld-Verified framing matters because both are agentic, tool-using evaluations, which lines up with the model's positioning as the most agentic Sonnet yet.

Safety and Cyber Caveats

Sonnet 5 has a lower rate of undesirable behaviors than Sonnet 4.6 and is generally safer in agentic contexts, per Anthropic. It is better at refusing malicious requests and at resisting prompt injection, and it shows lower hallucination and sycophancy than Sonnet 4.6. The honest caveat: on an automated behavioral audit, Sonnet 5 showed somewhat higher misaligned behavior than both Opus 4.8 and Claude Mythos Preview.

On cyber, Sonnet 5 has lower ability on dangerous cyber tasks than Opus 4.8 and Mythos 5. In a Firefox exploit evaluation, Sonnet 5 never developed a full working exploit. Cyber safeguards are enabled by default and are the same safeguards shipped in Claude Opus 4.7 and 4.8, so the security posture readers saw introduced with the Opus 4.7 launch carries forward unchanged here.

Where to Use It

Sonnet 5 is available across all Claude plans. It is the default model for Free and Pro users, so most people will be using it without changing a setting. Max, Team, and Enterprise users can select it, and it is available in Claude Code and on the Claude Platform and API as claude-sonnet-5.

The concrete day-one moves: if you are on the API, point a slice of traffic at claude-sonnet-5 and measure the token multiplier against your Sonnet 4.6 baseline before August 31, while the intro pricing softens the transition. If you are in Claude Code, the agentic improvements (staying on plan, self-checking output, fewer steps) are exactly the behaviors to test on a real multi-step task. And if accuracy on a high-stakes task is the priority, the launch language is clear that Opus 4.8 is still the model to reach for.

Sources

• Anthropic, "Introducing Claude Sonnet 5" (June 30, 2026): https://www.anthropic.com/news/claude-sonnet-5
• Claude announcement post (June 30, 2026): https://x.com/claudeai/status/2072017450611142835