Claude Sonnet 5

Claude Sonnet 5 is Anthropic’s June 30, 2026 successor to Sonnet 4.6 — and the company’s clearest statement yet that agentic capability is moving down-market. The pitch: planning, tool use and autonomous multi-step work “at a level that, just a few months ago, required larger and more expensive models,” priced at $2 / $10 per 1M tokens through August 31 (then $3 / $15, exactly what Sonnet 4.6 charged).

The launch framing is cost, not a new capability ceiling. At the introductory rate Sonnet 5 undercuts GPT-5.5 ($5/$30), Gemini 3.1 Pro (~$2/$12) and Anthropic’s own Opus 4.8 ($5/$25) — while landing close to Opus on quality. It even beats Opus 4.8 on Anthropic’s knowledge-work benchmark. There is one important catch, flagged by independent testing: Sonnet 5 works much harder per task, so per-task cost at standard pricing can actually exceed Opus 4.8. More on that below.

Benchmarks

Anthropic’s launch table compares Sonnet 5 against Sonnet 4.6 and Opus 4.8. Numbers are Anthropic-reported as of June 30, 2026; treat them as a dated snapshot.

Three results stand out:

• Terminal-Bench 2.1 jumps 13.4 points over Sonnet 4.6 — the clearest signal of better real-world agent behavior in terminals and CLIs.
• Humanity’s Last Exam (with tools) lands within half a point of Opus 4.8 (57.4% vs 57.9%), up from a 11-point gap for Sonnet 4.6.
• GDPval-AA v2 actually edges Opus 4.8 (1,618 vs 1,615). Note the v2: Anthropic updated the grader for this run, so these Elo figures are not comparable to the v1 numbers (e.g. Opus 4.8’s 1,890) quoted elsewhere on this site.

The one place Opus 4.8 keeps a clear lead is SWE-bench Pro (69.2% vs 63.2%). For the hardest, longest-horizon coding, Opus remains the stronger pick.

The independent read: Artificial Analysis

Anthropic’s numbers are its own, so the most useful outside signal comes from Artificial Analysis, which evaluated Sonnet 5 ahead of release:

• Intelligence Index: 53 at max effort — the #5 model overall, a 6-point gain over Sonnet 4.6, level with GPT-5.5 at high reasoning, and 2–3 points behind GPT-5.5 (xhigh) and Opus 4.8 (max).
• Cost per task: $2.29 on the Index at standard $3/$15 pricing — roughly double Sonnet 4.6 and ~15% more than Opus 4.8. The driver is token usage, not price: at max effort Sonnet 5 used ~40% more output tokens per task than Sonnet 4.6 and about 3× the agentic turns on knowledge-work benchmarks.
• Effort scales sharply: max effort used ~6× more turns than low effort on GDPval-AA. The effort dial, not the rate card, is what controls your bill.
• Where it punches above its tier: on AA-Briefcase and GDPval-AA (agentic knowledge work), Sonnet 5 sits just ahead of Opus 4.8 and trails only Claude Fable 5.
• Where it doesn’t: on the CritPt frontier-physics benchmark it scores 17% — 14 points above Sonnet 4.6, but behind GLM-5.2, Opus, Fable and GPT-5.5 at xhigh.

So the honest cost summary: cheaper than Opus holds during the promo window and at lower effort settings. At standard pricing and max effort, Sonnet 5 can cost more per task than Opus 4.8 while delivering slightly less intelligence. Price it per task, not per token.

Pricing

Two cost caveats worth planning around:

• The tokenizer changed. Sonnet 5’s updated tokenizer maps the same text to roughly 1.0–1.35× more tokens (code and non-English text at the high end) — similar to the change Opus 4.7 introduced. Anthropic says intro pricing was set so the migration is roughly cost-neutral; re-baseline on representative prompts before assuming the per-token discount reaches your bill.
• Token usage is the real variable. As the Artificial Analysis data shows, high effort settings multiply turns and output tokens. For high-volume deployments, set effort explicitly per workload.

Thinking and effort

Adaptive thinking is always on — the model decides how much to reason based on task difficulty. Effort defaults to high on the API and in Claude Code. Sonnet 5 adds an xhigh setting Sonnet 4.6 lacked, giving it the same five levels as Opus 4.8: max, xhigh, high, medium, low.

This is the same reasoning approach as the Opus tier, and it’s a big part of why Sonnet 5 closes so much of the agentic gap: it can plan a multi-step job, call a tool, read the result and correct course without a human nudging each step. It’s also why costs scale so steeply with effort — max effort spends several times more agentic turns than low on the same task.

Built for agents

The launch’s clearest theme: Sonnet 5 is designed to run agents, not just answer questions. The big gains (Terminal-Bench, OSWorld) are agent benchmarks, and partner feedback centered on follow-through:

• Zapier reported that multi-step automations that “used to stall halfway” now complete end to end — “for day-to-day automation, it’s a no-brainer.”
• Lovable emphasized refusal quality: “a model that knows when to say no is just as important as one that knows how to build.”
• ClickHouse highlighted tighter reasoning steps and faster time-to-insight on live-data agents; Eve (legal) called it Pareto-optimal on plaintiff-law tasks; Pace runs insurance back-office workflows on it.

For teams already routing to Opus 4.8 for agent work, the practical pattern is: run the fleet on Sonnet 5, escalate only the hardest coding steps to Opus.

Safety

Anthropic reports Sonnet 5 is safer than Sonnet 4.6 across its pre-deployment suite: lower rates of misaligned behavior, less cooperation with misuse, less deception, fewer hallucinations, and better refusal of malicious requests — plus stronger resistance to prompt-injection hijacks in agentic settings (though its automated-audit misalignment rate remains above Opus 4.8’s).

One deliberate design choice stands out: Sonnet 5 was not trained for offensive cybersecurity. On exploit-development evaluations (e.g. building working Firefox exploits) it scored 0% full success — far below Opus 4.8 and Mythos 5. Anthropic frames this as a safety property for a model deployed at massive scale, and ships it with cyber safeguards enabled by default (the same tier as Opus 4.7/4.8, less strict than Fable 5’s). If you do defensive security work that needs fewer guardrails, Anthropic points you to Opus 4.8 and its Cyber Verification Program.

Availability

Shipped broadly on day one (June 30, 2026):

Making it the Free-tier default puts an agent-capable model in front of Anthropic’s entire consumer base immediately — and Anthropic raised rate limits across Chat, Cowork, Claude Code and the Platform to absorb the higher token usage of high effort levels.

How Claude Sonnet 5 compares

• vs GPT-5.5: Sonnet 5 takes SWE-bench Pro; GPT-5.5 takes Terminal-Bench (on its own harness) and accepts audio/video. Sonnet 5 is far cheaper per token.
• vs Gemini 3.1 Pro: near-identical pricing at Sonnet 5’s intro rate; Gemini is natively omnimodal, Sonnet 5 leads the agentic coding benchmarks.
• vs Opus 4.8: the six-point SWE-bench Pro gap is now the main thing the Opus premium buys. On knowledge work they’re tied.

No flagship wins every category — what separates Sonnet 5 is Opus-adjacent quality at the lowest sticker price in the Claude ecosystem.

The Claude lineup (July 2026)

Sonnet 5 replaces Sonnet 4.6 outright — better on every published benchmark, at the same standard price. The question it forces is no longer “can a mid-tier model run agents?” but “which steps still justify Opus 5 or Fable pricing?”

Best for

• High-volume agent deployments where Opus pricing doesn't scale
• Day-to-day coding, terminal and browser automation
• Agentic knowledge work — document analysis, research, professional outputs

Pros & cons

Frequently asked questions

Sources

• Anthropic — Introducing Claude Sonnet 5
• Anthropic — Claude Sonnet 5 System Card
• Artificial Analysis — Claude Sonnet 5 evaluation (X thread)
• Fello AI — Claude Sonnet 5 Just Launched