Claude Sonnet 5: Anthropic's New Agentic AI Model Explained

On June 30, 2026, Anthropic released Claude Sonnet 5, calling it "the most agentic Sonnet model yet." The launch comes at a pivotal moment — Anthropic is racing toward a landmark IPO, and the AI industry has shifted from simple chatbots to autonomous agents that plan, use tools, and complete multi-step tasks with minimal supervision. Sonnet 5 is positioned squarely at the center of that shift: near-flagship performance, at a fraction of the cost.

Here's everything you need to know about the model, what changed, and whether it's worth switching to.

What is Claude Sonnet 5?

Claude Sonnet 5 is the latest release in Anthropic's mid-tier "Sonnet" model line, sitting between the lightweight Haiku models and the flagship Opus tier. It succeeds Sonnet 4.6 (released in February 2026) and is built specifically to close the gap with Opus-class models on agentic work — planning, tool use, coding, and long-running autonomous tasks — while staying significantly cheaper to run.

Anthropic describes it as being able to "make plans, use tools like browsers and terminals, and run autonomously at a level that, just a few months ago, required larger and more expensive models."

Availability: where can you use it?

Sonnet 5 rolled out everywhere at once:

• Claude.ai (web, iOS, Android) — it is now the default model for Free and Pro plans, and is also available on Max, Team, and Enterprise plans
• Claude Code — Anthropic's coding agent
• Claude Platform / API — for developers, using the model string claude-sonnet-5
• Cloud platforms — Amazon Web Services (Bedrock), Google Cloud (Vertex AI), and Microsoft Foundry

Anthropic also increased rate limits across Chat, Cowork, Claude Code, and the API to accommodate the higher token usage that comes with running the model at higher "effort" (reasoning) levels.

Pricing: how much does Sonnet 5 cost?

This is where Sonnet 5 makes its strongest pitch. It launches with introductory pricing that undercuts Opus 4.8 significantly:

Tier	Input (per million tokens)	Output (per million tokens)	Valid until
Sonnet 5 (introductory)	$2	$10	August 31, 2026
Sonnet 5 (standard, from Sept 1)	$3	$15	—
Opus 4.8 (for comparison)	$4–5	$25	—

That works out to roughly 40% cheaper than Opus 4.8 at standard pricing, and about 60% cheaper during the introductory window. Anthropic is also offering up to 90% cost savings with prompt caching and 50% savings via batch processing, same as with other current-generation models. US-only inference is available for workloads that need it, at 1.1x standard pricing.

Read Now: How Kimi K2 Made Me Quit Claude Forever (OpenRouter)

Performance: how good is it, really?

Anthropic and independent testers ran Sonnet 5 through a battery of benchmarks. The headline numbers:

• Agentic coding: Sonnet 5 scores 63.2%, compared to Opus 4.8's 69.2% and the previous Sonnet 4.6's 58.1% — a solid jump over its predecessor, though it doesn't fully catch up to Opus
• Knowledge work: Sonnet 5 actually slightly outperforms Opus 4.8 on this benchmark category, which is notable given Opus is generally considered the stronger model for deep reasoning and judgment-heavy tasks
• Computer use / browsing (OSWorld-Verified, BrowseComp): at its highest "Extra High" reasoning setting, Sonnet 5 performs roughly in line with Opus 4.8 running at a medium-to-high setting — though running Sonnet 5 that hard can end up costing more than the equivalent Opus run

In short: Sonnet 5 doesn't beat Opus 4.8 outright, but it gets close enough on most tasks that it becomes the more practical default for teams watching their token budget.

What actually changed from Sonnet 4.6

The biggest shift isn't a single benchmark number — it's behavior. Early access partners consistently reported the same thing: Sonnet 5 finishes tasks that previous Sonnet models would abandon halfway through.

A few examples Anthropic and partners shared:

• Zapier handed the model a two-part job — update Salesforce account tiers, then send a launch announcement to enterprise contacts — and it completed the entire workflow end to end, something that used to stall midway
• In internal testing, Anthropic asked the model to investigate a bug. Without being prompted to, it wrote a test to reproduce the issue, implemented a fix, then temporarily reverted the fix just to confirm the bug actually came back — all in a single pass
• Cursor's team noted that with Sonnet 5, "agents stay on plan, follow our conventions, and ship clean multi-step changes"

The model also checks its own output without being explicitly told to — a pattern testers flagged repeatedly as a meaningful shift from earlier Sonnet generations.

Safety and alignment

Anthropic's system card for Sonnet 5 reports a few key safety findings:

• Lower rate of undesirable behaviors overall compared to Sonnet 4.6, including lower hallucination and sycophancy rates
• Better resistance to prompt injection and cleaner refusals of unsafe requests
• Much weaker cybersecurity capability than Opus-class models — on a Firefox exploit-development test run with Mozilla, Sonnet 5 could not produce a working exploit (0% success), far below Opus 4.8's 68.8% and the restricted Mythos 5 model's 88.4%
• Cyber safeguards are enabled by default, similar to Opus 4.7/4.8, but less restrictive than the tighter controls placed on Anthropic's Mythos-class models
• On Anthropic's Responsible Scaling Policy evaluations, Sonnet 5 does not cross the automated AI R&D risk threshold and remains less capable than Claude Mythos 5 across the board

That said, Anthropic notes Sonnet 5 shows somewhat higher rates of misaligned behavior than Opus 4.8 and Claude Mythos Preview — it's safer than its own predecessor, but still a step below Anthropic's most capable and most heavily safeguarded models.

Sonnet 5 vs. Opus 4.8: which should you use?

	Claude Sonnet 5	Claude Opus 4.8
Best for	Day-to-day agentic work, coding, high-volume automation	Highest-accuracy reasoning, hardest tasks
Input price	$2/Mtok (intro) → $3/Mtok	$4–5/Mtok
Output price	$10/Mtok (intro) → $15/Mtok	$25/Mtok
Agentic coding score	63.2%	69.2%
Knowledge work	Slightly ahead	Slightly behind
Cybersecurity capability	Low	Higher

Anthropic's own framing: "Opus 4.8 is still the model of choice for higher accuracy on these tasks, but Sonnet 5 provides developers with lower-priced options that are of much higher quality than what was previously available." For most day-to-day coding, automation, and knowledge-work tasks, Sonnet 5 is now the sensible default — reserve Opus 4.8 for the genuinely hard problems where accuracy matters more than cost.

How Sonnet 5 compares to competitors

Sonnet 5 also undercuts rival frontier models on price while staying competitive on capability. It's positioned as cheaper than both OpenAI's GPT-5.5 and Google's Gemini 3.1 Pro, though it remains pricier than lightweight models like Gemini 3.5 Flash. Combined with its agentic performance gains, that pricing puts real pressure on competitors targeting the same enterprise automation use case.

Conclusion

Claude Sonnet 5 isn't trying to be Anthropic's smartest model — it's trying to be the model most teams should actually be using. It narrows the gap with Opus 4.8 on the tasks that matter most for agentic workflows, ships with meaningfully better safety numbers than its predecessor, and costs a fraction of the flagship tier, especially during the introductory pricing window through August 31, 2026. For coding, automation, and day-to-day agent work, it's a strong new default. For the hardest reasoning tasks, Opus 4.8 still holds the edge.

Frequently Asked Questions