What Anthropic Shipped on June 30

Anthropic released Claude Sonnet 5 on June 30, 2026, and immediately made it the default model for Free and Pro users on claude.ai. It’s also the new default in Claude Code, the API, AWS Bedrock, Google Vertex, and Microsoft Foundry. The pitch is straightforward: near-Opus 4.8 quality at roughly 40–60% lower cost per token, depending on when you’re reading this.

That’s a real shift for teams running agentic workloads at scale. If you’ve been routing every non-trivial task to Opus 4.8, Sonnet 5 is worth re-evaluating as your daily driver.

The Benchmark Numbers

On SWE-bench Pro — the standard proxy for real-world agentic coding — Sonnet 5 scores 63.2%, up from Sonnet 4.6’s 58.1%. Opus 4.8 still leads at 69.2%, so there’s a meaningful gap at the top. But the jump from 58.1% to 63.2% in a single Sonnet generation is significant; a year ago the difference between Sonnet and Opus tiers was much wider.

Computer use (OSWorld-Verified) shows similar improvement: Sonnet 5 posts 81.2% against Sonnet 4.6’s 78.5%. On knowledge work, the convergence is almost complete. Sonnet 5 scores 1,618 on GDPval-AA v2; Opus 4.8 scores 1,615. On Humanity’s Last Exam with tools — the hardest reasoning benchmark available — Sonnet 5 hits 57.4% while Opus 4.8 reaches 57.9%. Those are rounding-error differences.

Terminal-Bench 2.1, which tests shell and toolchain autonomy, is one area where Sonnet 5 (80.4%) trails both Opus 4.8 and GPT-5.5 (83.4%). For teams building long-horizon shell agents, that gap still matters. We covered the Sonnet 5 vs GPT-5.6 Sol head-to-head in detail here.

What’s Actually New Under the Hood

Beyond raw scores, Sonnet 5 ships three meaningful changes from Sonnet 4.6.

Adaptive thinking is on by default. The model selects its own reasoning effort level without you configuring it. If the task is a simple retrieval, it uses low effort. If it’s a complex multi-step plan, it escalates. You can still override with explicit effort levels — low, medium, high, max, or x-high — but the default behavior is smarter about resource allocation.

Self-verification without prompting. Sonnet 5 checks its own output at key steps in a chain without being asked. Earlier Sonnet models would complete a task and stop; Sonnet 5 is more likely to catch an obvious error in its own tool call before returning a result. Whether this actually reduces error rates in production will take time to measure, but the internal behavior is different.

New tokenizer, more tokens. This is the one tricky part of the migration. Sonnet 5 uses a new tokenizer that produces roughly 30% more tokens for the same text compared to earlier versions, depending on content type. The model is not 30% more expensive — the underlying quality-per-token improved — but your API costs may shift if you’re projecting from historical token counts. Budget for a calibration period if you’re optimizing on token efficiency.

Sonnet 5 is also the first model in the Sonnet tier to include real-time cybersecurity safeguards, which Anthropic describes as active filtering of inputs and outputs on security-relevant tasks. Details are thin, but it matters for enterprise customers in regulated sectors.

The Pricing Calculus

Through August 31, 2026, Sonnet 5 is $2 per million input tokens and $10 per million output tokens. After that, pricing reverts to the same rate as Sonnet 4.6: $3/$15. Opus 4.8 costs $5/$25 per million tokens with no introductory discount.

At standard rates, Sonnet 5 is 40% cheaper per token than Opus 4.8. During the introductory window, it’s 60% cheaper. On most agentic tasks where Sonnet 5 matches Opus 4.8 closely — knowledge work, reasoning with tools, computer use — the cost advantage is the decisive factor. Opus 4.8 introduced dynamic workflows and 1,000 parallel agents; Sonnet 5 now lets you run those patterns at significantly lower burn rate.

The exception is where the accuracy gap still bites: Opus 4.8 leads on SWE-bench Pro by 6 percentage points and on Terminal-Bench by roughly 3. If you’re running an agentic coding pipeline where a failed tool call costs a human hour to debug, that accuracy premium may still be worth paying.

How to Think About Model Selection Now

The practical framing: Sonnet 5 is your new default, Opus 4.8 is your escalation path. Make the switch unconditional on knowledge-work and reasoning tasks — the benchmarks are indistinguishable. On agentic coding, run an A/B test against your actual task distribution before making the call; SWE-bench Pro is a proxy, not your production load.

One nuance: Sonnet 5 at x-high effort can become more expensive than Opus 4.8 at lower effort levels, because adaptive thinking burns tokens. If you’re running tasks that frequently escalate to x-high, profile actual costs rather than extrapolating from the list price. The cost-performance curves overlap at high effort settings.

For teams already using AI coding assistants in their development workflow, Sonnet 5 as the Claude Code default means most daily coding tasks now run on a materially better model at the same price point. That’s a straightforward win, no configuration required.

What This Means for the Mid-Tier Market

The mid-tier is now doing what the frontier tier did twelve months ago. Sonnet 5 matches Opus 4.8 on knowledge benchmarks and comes within a few percentage points on coding. That compression — where a $3/MTok model nearly equals a $5/MTok model — will continue. The question isn’t whether Opus 4.8 remains valuable; it does, for tasks at the hard tail. The question is what Opus 5 (or whatever Anthropic calls the next frontier model) needs to do to justify a premium above a Sonnet tier that keeps closing the gap.

The 1M context window, now standard across both tiers, also removes one of the last remaining reasons to route tasks to the frontier tier by default. The differentiation is increasingly about accuracy at the edge, not feature set or context.

Further Reading

• Introducing Claude Sonnet 5 (Anthropic) — the official release post, including benchmark tables and migration notes
• Sonnet 5 vs Sonnet 4.6 vs Opus 4.8: Full Benchmark Comparison (MarkTechPost) — detailed per-benchmark breakdown if you’re making a routing decision
• Anthropic launches Claude Sonnet 5 as a cheaper way to run agents (TechCrunch) — focuses on the enterprise and agentic use case framing