Claude Sonnet 5
Anthropic · Released Jun 2026
Anthropic's most agentic Sonnet delivers near-Opus 4.8 performance at roughly half the price, with Terminal-Bench 2.1 +13.4 points over Sonnet 4.6. The new tokenizer inflates counts ~30%, so intro pricing ($2/$10 per MTok through Aug 31) keeps migration cost-neutral, but teams should recount budgets before switching. It replaces Sonnet 4.6 as the default across all Claude plans, ideal for agentic coding, brownfield debugging, and production pipelines — not for the hardest frontier reasoning or cybersecurity tasks.
Is it right for you?
Good for
- Multi-step agentic coding and autonomous task completion — finishes complex workflows previous Sonnet models stalled on
- Brownfield debugging: traces failures to root causes, writes reproducing tests, and ships durable fixes unprompted
- Production AI agents at attractive cost — near-Opus 4.8 quality on agentic benchmarks at 40% of the price ($2/$10 against Opus 4.8's $5/$25)
- Knowledge work where it edges Opus 4.8 (GDPval-AA v2: 1,618 vs 1,615) — document drafting, research, professional analysis
- Self-verification: checks its own output without being explicitly prompted, reducing compounding errors in automated pipelines
- High-volume batch pipelines: the Batch API halves the rate to $1 input / $5 output per 1M, and cache reads land at $0.20 per 1M
Not good for
- Maximum-capability frontier tasks where Opus 5, Fable 5 or Fable 5.1 are necessary — Sonnet 5 gets close but doesn't match the top tier
- Cybersecurity work requiring reduced guardrails — cyber safeguards are enabled by default and Anthropic recommends the Opus tier for that use case
- Workloads where the new tokenizer's ~30% token inflation matters — recount budgets before switching; same text costs more tokens than on Sonnet 4.6
How it performs by task
Agentic coding
Terminal-Bench 2.1 80.4% (+13.4 pts over Sonnet 4.6) — the clearest signal of multi-step agentic reliability. Cursor production benchmark: 61.2% (vs 49.0% for 4.6).
Code generation
SWE-bench Verified 85.2% — 4.1 pts ahead of Sonnet 4.6 (81.1%), closing the gap with Opus 4.8 (88.6%). Strong on standard coding but Opus still leads Pro by 6 pts.
Debugging
Early testers repeatedly highlight unprompted reproducing tests, stashing to confirm regressions, and root-cause tracing on brownfield code.
Tool use and computer use
OSWorld-Verified 81.2%, matching Opus 4.8 (83.4%) and significantly ahead of Sonnet 4.6 (78.5%). Handles browsers, terminals, and API tools end-to-end.
Knowledge work
GDPval-AA v2: 1,618 — slightly edges Opus 4.8 (1,615). A genuine first for a mid-tier Sonnet model against its flagship sibling.
Reasoning
Humanity's Last Exam: 43.2% (no tools) / 57.4% (with tools). With tools, nearly matches Opus 4.8 (57.9%); without tools, still trails the flagship tier.
Cybersecurity
By design — never developed a full working exploit on Firefox 147 vulnerability tests (0.0%). Cyber safeguards enabled by default. Anthropic recommends Opus 4.8 for cyber work.
Pricing
Input
$2 / 1M
Output
$10 / 1M
Context
1M tokens; batch $1/$5; cache read $0.20
Benchmarks
| Benchmark | Score | Source |
|---|---|---|
| SWE-bench Verified | 85.2% | Source |
| Terminal-Bench 2.1 | 80.4% | Source |
| SWE-bench Pro | 63.2% | Source |
| Humanity's Last Exam (no tools) | 43.2% | Source |
| Humanity's Last Exam (with tools) | 57.4% | Source |
| OSWorld-Verified | 81.2% | Source |
| Knowledge Work (GDPval-AA v2) | 1,618 | Source |
| Cursor Production Benchmark (Max effort) | 61.2% | Source |
| BrowseComp 25 (agentic search) | 84.7% | Source |
Verdict history
Sources
- Kingy AI — Claude Sonnet 5: Benchmarks, Specs, Pricing & Everything NewJul 2026
- Anthropic: Claude Platform on AWS pricing docsJun 2026
- TechCrunch — Anthropic launches Claude Sonnet 5Jul 2026
- Anthropic — Introducing Claude Sonnet 5Jul 2026
- Codersera — Claude Sonnet 5: Benchmarks, Pricing & How It ComparesJul 2026
- Claude Platform Docs — Models OverviewJul 2026
- Anthropic Platform Docs — What's new in Sonnet 5Jul 2026
- MarkTechPost — Claude Sonnet 5 vs Sonnet 4.6 vs Opus 4.8: Benchmarks and Cost-Performance ComparedJul 2026
Verification log
- Pricing— Updated
Automated agent
LIVE WRONG PRICE CORRECTED on a high-traffic entity. Stored pricing_input/$pricing_output read $3 / 1M and $15 / 1M; Anthropic's live rate card reads $2 / MTok input and $10 / MTok output. The stored figure was never a live rate: it was the announced 1 September 2026 successor to the launch price, written in ahead of a change Anthropic has since cancelled. The page carries an explicit note: 'The $2/$10 ... announced at launch as introductory pricing through August 31, 2026, is now the standard price. The previously scheduled increase to $3/$15 ... on September 1, 2026 will not occur.' Corrected pricing_input to $2 / 1M, pricing_output to $10 / 1M, and set the per_mtok decimals to 2.00 / 10.00. Refreshed pricing_context to '1M tokens; batch $1/$5; cache read $0.20' (batch $1/$5 and cache read $0.20 both read off the same table). Rewrote the strength that claimed 'roughly half the price' of Opus 4.8 to the correct 40% ($2/$10 against $5/$25), added a batch-pipeline best_for, and replaced the two audience rows that framed $2/$10 as intro pricing expiring 31 August. Rating stays recommended. Verdict summary still carries the stale intro-pricing framing and the cancelled rate; that rewrite is unreachable from model_update and is staged as an entity_verdict payload on draft post 'claude-sonnet-5-price-cut-made-permanent'.
- Pricing— No changes
Automated agent
- Pricing— No changes
Automated agent
- Pricing— No changes
Automated agent
- Pricing— No changes
Automated agent
Live $2/$10 intro through Aug 31; stored $3/$15 standard rate. No change to standard rate.
- Pricing— No changes
Automated agent
- Pricing— No changes
Imported at launch
- Profile— No changes
Imported at launch