Claude Sonnet 5 launched June 30, 2026, and it's the most agentic Sonnet Anthropic has ever shipped. It lands near Opus 4.8 quality at roughly 40% of the price, with a 1M-token context window and always-on adaptive thinking. For developers, researchers, and AI practitioners, it's arguably the best value-to-performance ratio in the Claude lineup right now.
Here's the full breakdown.
Headline Specs
- Model ID:
claude-sonnet-5 - Context window: 1M tokens
- Max output: 128K tokens (up to 300K via batch API beta)
- Knowledge cutoff: January 2026
- Adaptive thinking: Always on, defaults to high effort
- Available on: Claude API, claude.ai (Free and Pro default), Claude Code, Cursor, VS Code, GitHub Copilot
What's New Versus Sonnet 4.6
The gains aren't uniform across capabilities. Reasoning, tool use, and long-context coding saw the biggest jumps:
| Benchmark | Sonnet 5 | Sonnet 4.6 | Improvement |
|---|---|---|---|
| Terminal-Bench 2.1 (terminal/tool use) | 80.4% | 67.0% | +13.4 pts |
| Humanity's Last Exam (no tools) | 43.2% | 34.6% | +8.6 pts |
| Humanity's Last Exam (with tools) | 57.4% | 46.8% | +10.6 pts |
| SWE-bench Pro (agentic coding) | 63.2% | 58.1% | +5.1 pts |
| OSWorld-Verified (computer use) | 81.2% | 78.5% | +2.7 pts |
| GDPval-AA v2 (knowledge work) | 1618 | 1395 | +223 |
The standout number is Terminal-Bench 2.1: a 13.4-point jump means Sonnet 5 is dramatically better at operating inside a shell, running commands, reading output, and recovering from errors. If you run agents that live in a terminal, this is the most tangible upgrade.
Another notable result: on GDPval-AA v2 (professional knowledge work), Sonnet 5 scores 1618, edging out Opus 4.8's 1615. That's the first time a mid-tier Sonnet has beaten the flagship Opus on any headline benchmark.
How Close Is It to Opus 4.8?
On four of five benchmarks, Sonnet 5 lands between Sonnet 4.6 and Opus 4.8, closer to Opus than to its own predecessor. On the fifth (knowledge work), it narrowly beats Opus. The practical takeaway: Sonnet 5 reaches near-flagship quality for coding, reasoning, and computer use at a fraction of the cost.
The one area where Opus 4.8 maintains a clear lead is raw agentic coding on the harder SWE-bench Pro (69.2% vs 63.2%). If your workload is exclusively complex multi-file refactors, Opus remains worth the premium. For everything else, the gap is small enough that Sonnet 5's pricing makes it the rational default.
Pricing
| Tier | Input (per M tokens) | Output (per M tokens) |
|---|---|---|
| Sonnet 5 intro (through Aug 31, 2026) | $2 | $10 |
| Sonnet 5 standard (from Sep 1, 2026) | $3 | $15 |
| Sonnet 4.6 | $3 | $15 |
| Opus 4.8 | $5 | $25 |
The introductory rate runs through August 31, 2026. After that, Sonnet 5 settles at the same price as Sonnet 4.6 while delivering substantially better performance across the board.
The tokenizer catch: Sonnet 5 uses an updated tokenizer. The same input text can generate 1.0 to 1.35x as many tokens compared to Sonnet 4.6. Anthropic set the introductory pricing to make this transition roughly cost-neutral, but it's worth budgeting for the token count change once standard pricing kicks in.
Behavioral Change: It Finishes
The most-cited difference from early users isn't a benchmark number. It's completion. Sonnet 5 drives long agentic chains to completion more reliably than Sonnet 4.6. It stalls less, asks for permission less, and declares victory prematurely less often. This aligns with the Terminal-Bench improvement and the always-on adaptive thinking at high effort. If you've been frustrated by agents that start strong but lose the thread, Sonnet 5 is a meaningful upgrade.
What About Fable 5 and Mythos 5?
Anthropic's two most powerful models, Fable 5 and Mythos 5, have been offline since June 12, 2026, due to a US Commerce Department export-control directive. Anthropic could not technically separate foreign users from domestic users fast enough, so it disabled both models for all customers worldwide.
Until that changes, Sonnet 5 is not just an upgrade. For nearly everyone, it's the best Claude model you can actually use. Opus 4.8 remains available and is the recommended choice for security research and maximum coding capability, but Sonnet 5 covers most use cases at less than half the cost.
Migration Notes
If you're moving from Sonnet 4.6 to Sonnet 5, there are two breaking changes to plan for:
- Manual thinking controls are gone. The extended-thinking budget API from Sonnet 4.6 returns an error on Sonnet 5. Adaptive thinking replaces it entirely. If your integration sets thinking parameters, you'll need to update those calls.
- Priority Tier is not available on Sonnet 5. Anthropic has raised rate limits across Chat, Cowork, Claude Code, and the Platform to compensate.
The Verdict
Claude Sonnet 5 is the model that makes the strongest case for itself in the current Claude lineup. It delivers near-flagship performance at mid-tier pricing. It's the default for Free and Pro users. And with Fable 5 and Mythos 5 suspended indefinitely, it's the practical ceiling for what most developers can access today.
If you're choosing one Claude model to build on in July 2026, choose Sonnet 5. The only exception is if you need maximum coding capability for complex refactors (use Opus 4.8) or if your integration depends on manual thinking controls (stay on Sonnet 4.6 until you can migrate).
Claude Sonnet 5 is available now through the Anthropic API, claude.ai, Claude Code, and major IDE integrations. Introductory pricing of $2/$10 per million tokens runs through August 31, 2026.