Claude Sonnet 5

Anthropic‘s mid-tier model (claude-sonnet-5), released 2026-06-30 and pitched as “the most agentic Sonnet model yet” — built to make plans, drive tools like browsers and terminals, and run autonomously rather than answer one turn at a time. It succeeds Sonnet 4.6.

The headline claim

Anthropic positions Sonnet 5’s performance as close to Opus 4.8 while keeping the cheaper Sonnet price — a substantial step over Sonnet 4.6 on reasoning, tool use, coding, and knowledge work. That framing matters to this spoke’s central question (see synthesis): if the second-tier model lands near the flagship, the frontier premium is being squeezed from inside the same lab, not only from below by open weights.

Worth a caveat, though: the system card is blunter than the marketing — it says Sonnet 5 “trails our Opus and Mythos-class models” in almost all cases. So “close to Opus” is a price-adjusted claim, not a parity claim. The card also reveals a class above Opus: Mythos, which it treats as the current capability frontier and benchmarks Sonnet 5 below throughout.

Benchmarks (vendor-reported)

From the authoritative table in the system card (Sonnet 5 figures):

SWE-bench Verified: 85.2% (Pro 63.2, Multilingual 78.3, Multimodal 28.1).
Humanity’s Last Exam: 43.2% without tools, 57.4% with tools.
OSWorld-Verified: 81.2% — the computer-use / agentic-desktop benchmark, consistent with the “most agentic Sonnet” pitch.
FrontierCode v1: 38.8 vs GPT-5.5’s 25.5 — its widest coding margin.

These are Anthropic’s own numbers; treat them as a dated snapshot, not an independent verdict. The neutral cross-check would be artificial-analysis once it re-runs its index.

Correction (2026-06-30): the figures first taken from the launch announcement (HLE 34.6 / 46.8, OSWorld 78.5) turned out to be Sonnet 4.6’s numbers in the system-card table — a misattribution in the earlier lossy fetch, now fixed above. See claude-sonnet-5-system-card.

Pricing

A dated snapshot (llm-api-pricing):

Introductory (through 2026-08-31): $2 / 1M input, $10 / 1M output.
Standard (after): $3 / 1M input, $15 / 1M output.

The standard rate is the familiar Sonnet tier; the introductory cut is a launch discount. Either way it sits far under the $30/$180 frontier-reasoning ceiling tracked in llm-api-pricing — the point of the “close to Opus at Sonnet cost” pitch.

Context window: 1M tokens standard, extended to a 10M-token limit via context compaction (triggered at 200k) for long-horizon agentic evals (system card).

API behavior & migration

From the official prompting guide. Several changes are hard 400 errors, not soft shifts: temperature/top_p/top_k are rejected (steer tone via the system prompt), and manual extended thinking (budget_tokens) is removed. Adaptive thinking is on by default (a no-thinking request now thinks, unlike 4.6). Intelligence is tuned with the effort parameter (low→max, default high, xhigh for the hardest coding/agentic work); notably Sonnet 5 at medium ≈ Sonnet 4.6 at high — more capability per tier. A new tokenizer emits ~30% more tokens for the same text, which both truncates 4.6-tuned max_tokens and raises effective cost (see synthesis). The model is also more agentic by default and follows instructions more literally.

Availability

Default model for Free and Pro plans, and available to Max, Team, and Enterprise users. Reachable through Claude Code, the Claude Platform, and the Claude API.

Safety

Anthropic reports an overall lower rate of undesirable behaviors than Sonnet 4.6, and substantially poorer performance than the Opus models on cybersecurity exploit development, with safeguards on by default.

Why it matters

The release sharpens the spoke’s open question — does the frontier premium survive? Here the pressure comes from within: a mid-tier model claimed near-flagship at a fraction of the price. It also feeds the agentic-tooling read that “Sonnet + harness beats raw Opus” on cost — Sonnet 5 is explicitly tuned for the harness use-case (plans, tools, autonomy).