Claude Sonnet 5

Spec

Attribute	Value
Developer	Anthropic
Tier	Sonnet (mid-tier)
Released	2026-06-30
Announced	2026-06-30
Context window	1M tokens
Pricing	$2/M input · $10/M output — permanent as of 2026-08-10; the scheduled 2026-09-01 rise to $3/$15 was cancelled
License	proprietary (API-only; no weight release)
Availability	Claude.ai (all tiers), Claude Code, API
Default model	Yes — Claude Free, Claude Pro, Claude Code

Release Date

June 30, 2026 — released the same day as Fable 5/Mythos 5 export-control restoration and Claude Science launch.

Pricing — the September increase was cancelled

2026-08-10. Anthropic stated that Sonnet 5's introductory pricing is permanent: "We launched Sonnet 5 in June at $2 per million input tokens and $10 per million output tokens through August 31, and that price will remain unchanged" (source). Subscription plan prices are reported unchanged by the same capture.

This page published the superseded figure for six days, and the withdrawn claim is recorded rather than silently replaced:

Held here until 2026-08-16	Correct from 2026-08-10
$2/M · $10/M as a promo through 2026-08-31	$2/M · $10/M, no end date
$3/M · $15/M standard from 2026-09-01	cancelled; never took effect
The cancelled increase was 50% on both input and output. Anthropic gave no
new end date and no replacement schedule in anything read, so the row now carries
a price rather than a countdown.

Why it was late (day +6). A price change is not a release: it produced no model card, no release-tracker entry, and nothing on Anthropic's /news page that any of the six intervening runs surfaced — each of which saw the 2026-08-04 Cuéllar appointment as the newest item there. The vendor statement read is an @claudeai X post, which is an author/vendor channel under this repo's trust_order — one rank below an official blog post, and the only vendor statement obtainable from this environment (source). Anthropic's own rate card at platform.claude.com could not be fetched, so the current published price was not verified against first-party documentation.

What it changes: every comparison on this page that used "one-fifth to one-thirteenth the price of Opus 4.8" was written against a price with a four-week fuse on it. It no longer has one.

Benchmarks

Benchmark	Sonnet 5	Sonnet 4.6	Opus 4.8
SWE-bench Pro	63.2%	58.1%	69.2%
Terminal-Bench 2.1	80.4%	—	74.6%
BrowseComp (single agent)	84.7%	—	—
BrowseComp (multi-agent, 10M token budget)	86.6%	—	—
GDPval-AA v2 (knowledge work)	1,618	—	1,615
OSWorld-Verified (computer use)	81.2%	—	—
Notable: On three of five agentic benchmarks (Terminal-Bench 2.1, BrowseComp, GDPval-AA v2), Sonnet 5 matches or surpasses Opus 4.8 — at approximately one-fifth to one-thirteenth the price.

(sources), (DataCamp), (MarkTechPost)

Use Cases

• Drop-in upgrade for agentic coding workflows running on Sonnet 4.6 — same pricing tier, significantly higher capability
• Long-horizon browser + terminal automation (BrowseComp 84.7% / Terminal-Bench 80.4%)
• Knowledge-intensive multi-step workflows at low cost (GDPval-AA: 1,618 > Opus 4.8: 1,615)
• Default model for general Claude users on Free/Pro plans

Adaptive Thinking & Effort Levels

Sonnet 5 introduces adaptive thinking with five effort tiers: low, medium, high, max, x-high. Higher effort levels allocate larger thinking budgets per response for greater accuracy at higher token cost. This mirrors the pattern introduced in Opus 4.8 (Dynamic Workflows) and extends it to the mid-tier model.

Safety Improvements

Per the Anthropic system card:

• Lower hallucination and sycophancy rates vs. Sonnet 4.6
• Stronger refusal of malicious requests
• Increased resistance to prompt injection in agentic contexts
• No explicit safety gating (unlike Fable 5's <5% Opus 4.8 gate)

Superseded 2026-09-28, and a second Terminal-Bench number that is not a contradiction

Claude Sonnet 5.5 replaced this model on 2026-09-28 at the same $2/$10 rate card, so the successor's "costs up to 30% less for most work" is a claim about tokens spent per task, not about the price of a token (source).

Its launch table gives this model 10.3% on Terminal-Bench 4.0. The ## Benchmarks section above holds 80.4% on Terminal-Bench 2.1. Both are correct and neither is being corrected here: they are different harnesses two major versions apart, and the 70.1-point spread measures the instrument rather than the model. It is recorded because a reader comparing "Terminal-Bench" across this wiki's model pages will otherwise read a collapse that did not happen — see Eval Harness Configuration, which treats this as the cleanest case it holds, there being no second vendor to suspect.

The successor's published margins over this model, all from its own launch table: Chartography 61.6% vs 15.6%, OSWorld 2.1 80.1% vs 57.0%, CursorBench 4.0 55.5% vs 34.1%, GDPval-AA v2.1 1844 vs 1449, AA-Briefcase v1.1 1811 vs 1359, Humanity's Last Exam 64.5% vs 54.9%, FrontierCode 1.1 Main 46.2% vs 42.4%.

Compared To

Model	SWE-bench Pro	Input price/Mtok	Context
Claude Fable 5	80.3%	$10	1M
Claude Sonnet 5	63.2%	$2 (promo)	1M
Claude Opus 4.8	69.2%	$15	1M
Claude Sonnet 4.6	58.1%	$3	—
Sonnet 5 at $2/$10 vs. Opus 4.8 at $15/$75 is the intended efficiency play: close to Opus-class on most agentic tasks at ~13× lower input cost.

Sources

• Anthropic — Introducing Claude Sonnet 5 (source)
• TechCrunch
• VentureBeat
• DataCamp

Related

• Claude Fable 5 — Mythos-class public frontier; Sonnet 5 is the efficient mid-tier below it
• Claude Opus 4.8 — prior Opus-tier frontier; Sonnet 5 surpasses it on several agentic benchmarks
• Anthropic
• Agents (LLM Agents) — Sonnet 5 designed for agentic use cases
• Anthropic — Models overview — spec figures verified 2026-07-27