The Event
On June 30, 2026 at 18:02 UTC, Anthropic officially launched Claude Sonnet 5.
This is the first time a Sonnet model comes close to Opus 4.8 in agentic capability. Anthropic stated explicitly that across its four core dimensions—reasoning, tool use, coding, and knowledge work—Sonnet 5 strictly outperforms Sonnet 4.6, and on some tasks at high effort settings it can even match Opus 4.8.
Pricing comes in two tiers:
- Early-bird rate (before August 31, 2026): $2 per million input tokens, $10 per million output tokens
- Standard rate (after August 31): $3 per million input tokens, $15 per million output tokens
- SWE-bench Pro: 63.2%
- OSWorld-Verified: 81.2%
- Context window: native 1M tokens
- On BrowseComp (agentic search) and OSWorld-Verified (computer use), the cost-performance curves fully cover Sonnet 4.6, approaching or even surpassing Opus 4.8 in the mid-to-high effort range
- Official announcement: https://www.anthropic.com/news/claude-sonnet-5
- System card: https://www.anthropic.com/claude-sonnet-5-system-card
- Follow-up coverage: https://news.qq.com/rain/a/20260701A01X7000
- Sonnet 5 has a new tokenizer. The new tokenizer maps the same text to 1.0–1.35× the token count, affecting billing (while the early-bird rate is cost-neutral, users after August 31 will need to re-do their math).
- MML / cybersecurity capability is below Opus 4.8 and Mythos 5. Anthropic explicitly says the model was not specifically trained for cyber tasks, so exploit-related scenarios still require higher-tier models.
- The corrected chart is worth remembering—next time you see an AI company's launch event, wait two days before judging real performance.
- Anthropic's cybersecurity regulation is tightening; Sonnet 5's default safeguards are a form of "self-restraint," but the cost is that some legitimate research scenarios may need to go through enterprise unlock processes.
By comparison, Opus 4.8 costs $5 / $25. In other words, Sonnet 5's standard price is 60% of Opus 4.8, with a performance gap that has narrowed considerably.
Performance figures (from aihot's summary plus aimadetools' follow-up coverage):
Another early signal: Sonnet 5 will be part of Anthropic's Cyber Verification Program—real-time detection and blocking of dangerous cyber operations, enabled by default. This is the first time the Sonnet series gets this layer of protection.
Availability: default model on Free / Pro, fully available on Max / Team / Enterprise, live in Claude Code and Claude Platform simultaneously. The model API ID is claude-sonnet-5.
One caveat worth noting: Anthropic fixed its cost-performance chart after launch because the initial BrowseComp data used a simplified method that underestimated Sonnet 5. With the standard method, the performance curves shifted up overall. A detail worth remembering—AI companies' own launch charts can be miscalculated.
Reference links:
Deep Analysis
Sonnet 5 is not a fine-tune of Sonnet 4.6—it's a major repositioning of Anthropic's flagship workhorse.
For the past two years, the Sonnet series has played the role of "mid-range workhorse": less capable than Opus but price-friendly, suited for high-frequency batch work. Opus was "expensive but smart," used to crack the hardest problems.
Sonnet 5 pushes that line right up to Opus's feet. Anthropic put it bluntly in the official announcement: "just a few months ago, required larger and more expensive models"—things that months ago needed bigger, pricier models can now be done by Sonnet 5, at 40% less cost.
The structural impact of this product strategy:
For the developer ecosystem: SWE-bench Pro 63.2% + 1M context + tool use + native agentic capability means agent pipelines that previously required Opus can now run on Sonnet 5. Cybersecurity safeguards enabled by default reduce compliance friction for enterprise deployment.
For model competition: 90% of the performance at 60% of the price effectively eats into the "Opus tier" market. OpenAI's o3-pro / o4, DeepSeek's R2, and Google's Gemini 2.5 Pro all need to rethink pricing—Sonnet 5 has essentially planted a new benchmark in the agentic coding niche.
For Anthropic itself: the launch implies that Sonnet 5 will consume more tokens than 4.6 (maxed-out effort + long context), but Anthropic says pricing before August 31 is essentially cost-neutral. In other words, Anthropic is betting that "more usage by developers + lower unit price = higher total revenue."
Early customer profiles (officially cited): ClickHouse uses it for live data exploration; Pace runs insurance-claim FNOL workflows on it; Lovable values its "refuse unsafe request cleanly" behavior. These aren't fictional benchmark customers—they're real production environments.
Why This Matters
1. The cost structure of AI coding has been rewritten—the old default of "Opus for important tasks, Sonnet for daily work" now has its boundary pushed into territory Sonnet 5 can handle. 2. The Cyber Verification Program bundled with Sonnet 5—the first cybersecurity protection layer for the Sonnet series, meaning it can now be safely placed in code scenarios with compliance requirements. 3. Evaluation has shifted from single-point benchmarks to cost-performance curves—Anthropic's pitch is not just "higher scores" but "at different effort settings, you can choose your own cost-performance combination." This is a new product language.