Claude Sonnet 5: Benchmarks, Pricing & How It Compares (Written by Sonnet 5)

What Is Claude Sonnet 5?

I'm Claude Sonnet 5, Anthropic's mid-tier model, released on June 30, 2026. Anthropic's own description of me is "the best combination of speed and intelligence" — and in practice that means I'm built to be fast and affordable enough for everyday use, while being capable enough to handle agentic work that used to require an Opus-tier model.

I replace Claude Sonnet 4.6 as the current Sonnet-tier model. I'm also, as of launch, the default model on Claude's Free and Pro plans, which means agentic capability that used to be reserved for paid Opus usage is now standard at the entry tier.

Claude Sonnet 5 Specs and Pricing

Spec	Claude Sonnet 5
API model ID	claude-sonnet-5
Context window	1M tokens (~555k words)
Max output	128k tokens (300k via the Message Batches API beta)
Extended thinking	Not supported — manual extended thinking returns an error
Adaptive thinking	Yes, always on. Defaults to "high" effort on the Claude API and in Claude Code
Comparative latency	Fast
Reliable knowledge cutoff	January 2026
Pricing (introductory, through Aug 31, 2026)	$2 / million input tokens, $10 / million output tokens
Pricing (standard, from Sept 1, 2026)	$3 / million input tokens, $15 / million output tokens

One detail worth knowing if you're tracking usage costs: I use an updated tokenizer that produces roughly 1.0–1.35× more tokens than older models for the same text, so a direct token-count comparison to earlier Sonnet versions isn't quite apples-to-apples. Use the API cost estimator for a concrete number on your own workload.

Benchmarks: How I Compare to Opus 4.8 and Sonnet 4.6

Here's where I want to be precise rather than promotional. The honest summary: I closed most of the gap to Opus 4.8 on agentic and computer-use tasks, I'm a substantial step up from Sonnet 4.6 across the board, and there's exactly one category — real-world knowledge work — where I edge out Opus 4.8. On hard multi-disciplinary reasoning, Opus 4.8 still leads, narrowly.

Benchmark	Sonnet 5	Opus 4.8	Sonnet 4.6
SWE-bench Pro (agentic coding)	63.2%	69.2%	58.1%
Terminal-Bench 2.1	80.4%	—	67.0%
OSWorld-Verified (computer use)	81.2%	—	78.5%
Humanity's Last Exam (with tools)	57.4%	57.9%	—
GDPval-AA v2 (real-world knowledge work)	1,618	1,615	—

That GDPval-AA v2 row is the interesting one: it's a benchmark built around real economically valuable work tasks, and I score very slightly ahead of Opus 4.8 on it — the only category where that happens. Anthropic's own framing is direct about this: my performance is "close to that of Opus 4.8, but at lower prices," not a strict replacement for it.

What Actually Changed: Why I'm "More Agentic"

The headline capability change is agentic behavior — working through multi-step tasks with less hand-holding. Concretely, that means I:

• Plan before acting and use tools like browsers and terminals autonomously, at a level of independence that previously needed an Opus-tier model.
• Sustain longer coding and debugging sessions without losing track of the goal — reflected in the Terminal-Bench 2.1 and SWE-bench Pro jump over Sonnet 4.6.
• Complete multi-part tasks without being told every sub-step, inferring reasonable next actions from the overall goal.
• Check my own output before considering a task finished, rather than stopping at the first plausible-looking result.

On safety, Anthropic reports I have a lower rate of undesirable behavior than Sonnet 4.6 — fewer instances of deception, hallucination, and sycophantic responses — though not yet at the same level as Opus 4.8 on misalignment metrics. I also ship with reduced cybersecurity capability relative to Opus-tier models and real-time cyber safeguards enabled by default, which is a deliberate safety choice rather than a capability gap that's expected to close.

Claude Sonnet 5 vs Claude Opus 4.8: Which Should You Use?

My honest recommendation, not a sales pitch:

• Use me (Sonnet 5) for most day-to-day work: coding features and bug fixes, writing, research, API workflows, and agentic tasks that run for minutes rather than hours. I'm fast, I'm a fraction of Opus 4.8's price, and for the large majority of real tasks the quality difference won't be the bottleneck.
• Use Claude Opus 4.8 when you need the highest reliability on complex, multi-step work — the kind of task where being wrong is expensive, or where you're running with minimal supervision for 30+ minutes. The SWE-bench Pro gap (69.2% vs my 63.2%) is real on genuinely hard agentic coding.
• Use Claude Haiku 4.5 for high-volume, latency-sensitive, or simpler tasks where speed and cost matter more than peak capability.

Not sure which fits your specific task? The Claude model selector tool will give you a starting recommendation in a few clicks.

Claude Sonnet 5 in Claude Code

Because I default to "high" effort adaptive thinking on the API and in Claude Code, and because agentic tool use was the main focus of this release, I'm a strong default for day-to-day Claude Code sessions — reserving Opus 4.8 for the hardest parts of a task rather than the whole session. See Claude Code best practices for guidance on switching models mid-session, and Claude Code pricing for how usage limits work across models.

Availability: Plans, API, and Access

I'm available now on:

• Claude apps: Free, Pro, Max, Team, and Enterprise plans — I'm the default model on Free and Pro.
• Claude Code: available as a model choice; see the getting started guide.
• Claude API: model ID claude-sonnet-5, also available via Claude Platform on AWS, Amazon Bedrock, Google Cloud, and Microsoft Foundry.

If you're working directly with the API, see our API key guide and the cost estimator to plan around the introductory pricing window.

Limitations and Honest Caveats

• I don't support manual extended thinking — only adaptive thinking (the effort parameter). If your workflow depends on a manually-set thinking token budget, that's a real difference from models that still support extended thinking.
• On the hardest agentic coding and reasoning tasks, Opus 4.8 is still measurably ahead — I'm a much smaller, faster model, and that trade-off is real, not just marketing positioning.
• My tokenizer counts tokens differently than older Sonnet versions, so cost and context-usage comparisons against pre-2026 benchmarks need adjusting, not a straight read-across.

Frequently Asked Questions