Claude Sonnet 5.5: Faster Everyday Coding Beside Opus 5.5

Anthropic released Claude Sonnet 5.5 on September 28, 2026 as the second model in the Claude 5.5 family, after Opus 5.5. The positioning is a complement, not a crown: faster and cheaper for well-scoped everyday work, while Opus keeps the complex open-ended lane.

That buyer story matters more than any single bench row. Sonnet 5.5 is strongest on everyday coding, bug fixes, and polished docs, slides, and spreadsheets (Anthropic highlights a strong design eye). It is a clear upgrade over Sonnet 5 on coding and knowledge work, with the same $2 / $10 list price and a claim of fewer tokens per task.

Headline self-reported numbers from the announcement table: 70.6% on Terminal-Bench 4.0, 55.5% on CursorBench 4.0, 52.1% on FrontierCode 1.1 Main at Xhigh, 64.5% on HLE with tools. List price is $2 / $10 with $0.20 cache reads. API id: claude-sonnet-5-5. None of the launch scores are LLM Stats verified yet.

At a Glance

• Catalog / API id: claude-sonnet-5-5 only
• Organization: Anthropic
• Release: September 28, 2026. Second Claude 5.5 family model after Opus 5.5. Haiku 5.5 still ahead.
• Pricing: $2 input / $0.20 cache read / $2.50 (5m) or $4 (1h) cache write / $10 output per 1M. Batch 50% off. Same list as Sonnet 5.
• Context: 1M input / 128K max output
• Modalities: Text + image in, text out
• Thinking: Adaptive, on by default. Default effort high on Platform / API; medium in Claude apps / Claude Code
• Knowledge cutoff: June 2026
• Safeguards: First Sonnet with cyber safeguards comparable to Opus-class; biology safeguards same as Sonnet 5
• Secondary card depth: SWE-bench Pro 81.3%, DeepSWE v1.1 71.0%; OSWorld 2.1 strict 43.5% vs partial 80.1%

What's New

Sonnet 5.5 is Anthropic's mid-tier follow-through after Opus 5.5. The product claim is speed and cost for the work most teams actually run every day: scoped coding tasks, bug fixes, and polished office artifacts. Anthropic reports outputs generate 30%+ faster than Sonnet 5, with typically fewer tokens and up to 30% less cost per task at the same list price.

Capability-wise, the announcement table shows large lifts over Sonnet 5 on agentic coding and knowledge-work Elo, and CursorBench within about two points of Opus 5.5. That is the "close enough for most jobs" story, not a claim that Sonnet replaces Opus.

On safeguards, Anthropic says this is the first Sonnet with cyber safeguards comparable to Opus-class models. Cyber capabilities jumped; biology safeguards stay at the Sonnet 5 level. This post does not reprint safety, refusal, or jailbreak tables.

Benchmarks

Prefer Anthropic's announcement table for the hero read. Scores below are self-reported. Not LLM Stats verified. Interpret by job, and keep the Terminal-Bench honesty flag in view before you rank Sonnet above Opus.

Coding and agents

Terminal-Bench 4.0 at 70.6%is the loudest coding headline versus Sonnet 5's 10.3% and Opus 5.5's 66.4% (Opus at Xhigh). Treat the Sonnet-vs-Opus margin carefully: third-party and Anthropic notes flag a possible Max-versus-Xhigh mismatch. At matched Xhigh, Anthropic's cost framing still places Sonnet lower. Do not claim Sonnet is overall stronger than Opus. Anthropic says Opus remains clearly stronger at complex open-ended work.

FrontierCode 1.1 Main at 52.1% (Xhigh) is the primary row in this post, versus Sonnet 5 at 42.4% and Opus 5.5 at 54.4%. Max on Sonnet is 46.2%, lower than Xhigh because code-review subagents hit timeouts and out-of-scope edits. CursorBench 4.0 at 55.5%lands within about two points of Opus 5.5's 57.8%, and well above Sonnet 5's 34.1%.

The System Card adds software-engineering depth behind that agent table: SWE-bench Pro 81.3%, Multilingual 90.3%, Multimodal 54.3%, DeepSWE v1.1 71.0%, FrontierSWE V2 61.9% (Proximal harness, mean across trials), FrontierCode Extended 64.4% at Xhigh, and ProgramBench 79.7%. Terminal-Bench-Science 0.1 is 59.9% on the card. Those are absolute card rows, not announcement vs-prior deltas.

Knowledge work and computer use

GDPval-AA v2.1 at 1844 Elo and AA-Briefcase v1.1 at 1811 Elo are the announcement knowledge-work rows (bare titles only). Both sit within a few Elo of Opus 5.5 (1846 / 1822) and far above Sonnet 5 (1449 / 1359). Elo is not a percent bench, so it stays out of the launch bar chart.

HLE with tools at 64.5%improves on Sonnet 5's 54.9% and trails Opus 5.5's 67.7%. The card also reports HLE without tools at 56.9%. Computer use needs both OSWorld 2.1 numbers: the announcement partial score is 80.1%; the System Card strict score is 43.5% on the same suite. That gap is the honest read. Chartography without tools is 61.6% on the launch table; with tools the card is 90.2%.

System card domains

Below are absolute System Card §8 capability rows, grouped by job. Safety, refusal, and jailbreak rows are skipped. No invented peer deltas. Self-reported; not LLM Stats verified.

Vision and CAD: BenchCAD is 74.7% without tools and 96.3% with a Python tool. Knowledge and office: OfficeQA 76.9% / OfficeQA Pro 65.6%; Legal Agent Benchmark all-pass 11.7% (card mean criterion-pass 92.1%; prefer all-pass as the standard metric); Toolathlon Verified Pass@1 77.8%. AutomationBench is 44.7% (Zapier, max effort, API default fallbacks enabled). Health: prefer length-adjusted HealthBench Professional 69.2% from Table 8.1.A (raw 77.1%); PhysicianBench is 63.2% (100 physician EHR tasks, max effort); HealthBench is stored as raw 69.4% with an LA band of about 64.7-65.4% across efforts, so there is no single max LA cite. Math: ArXivMath 86.8% without tools / 95.2% with tools. Multilingual: GMMLU 92.1% (42 languages), MILU 91.6% (11 languages). Bio capability (not refusal): BioMysteryBench Human Solvable 89.2% / Human Difficult 44.7%; LatchBio SpatialBench Verified 72.5%, SingleCellBench 59.1%; plus System Card §8.17 life-sciences rows (morphology-to-molecule 25.0%, medicinal chemistry ADME 65.3%, protein design sequence 51.0% / library ranking 54.8%, de novo binders 82.3%, biomedical imaging 72.2%, Protocols troubleshooting 67.3% / understanding V2 66.6%).

How to read the table

Interpret by job, not by a single rank. Everyday coding agents: Terminal-Bench, FrontierCode (prefer Xhigh), CursorBench, plus SWE / DeepSWE card depth. Computer use: always pair OSWorld 2.1 partial with strict. Knowledge work: GDPval-AA and AA-Briefcase Elo plus office and AutomationBench rows. Visual / CAD: Chartography and BenchCAD with and without tools. Multidisciplinary reasoning: HLE with and without tools. And keep Anthropic's own split: Sonnet for scoped speed and cost, Opus for hard open-ended judgment.

Pricing & Efficiency

First-party Anthropic list, USD per million tokens. Input and output match Sonnet 5 at $2 / $10, half of Opus 5.5's $4 / $20. Cache reads match Opus at $0.20. The efficiency story is not a sticker cut versus Sonnet 5. It is fewer tokens and faster generation at the same list.

Batch discounts are 50% on input and output. Public callers use claude-sonnet-5-5 only.

Effort / Thinking

Adaptive thinking is on by default. Default effort depends on the surface: high on Claude Platform and the API, medium in Claude apps and Claude Code. Effort levels run low → medium → high → xhigh → max. If you migrate from Sonnet 5, re-measure tokens-per-task at the default for your surface before you assume you need max.

When to Use / Migrate

• Good fit: Everyday coding, bug fixes, scoped agent loops, and polished docs / slides / spreadsheets where Sonnet 5 was already the default and you want a clear upgrade without Opus pricing.
• Migrate from Sonnet 5when coding and knowledge-work quality dominate. Same list price; expect fewer tokens and faster outputs on Anthropic's account. Watch default effort (high on Platform, medium in apps).
• Stay on / prefer Opus 5.5for complex open-ended judgment, long-horizon ambiguous agent work, and cases where Anthropic's own guidance says Opus remains clearly stronger. Terminal-Bench alone is not a reason to demote Opus.
• Cyber / biology: first Sonnet with Opus-class cyber safeguards; biology safeguards remain at the Sonnet 5 level. This post does not reprint safety scoreboards.

Caveats

• All launch benches above are self-reported by Anthropic, not LLM Stats verified.
• Terminal-Bench 4.0 Sonnet 70.6% vs Opus 66.4% may mix Max and Xhigh effort. Do not read it as overall Sonnet superiority. Anthropic says Opus remains clearly stronger at complex open-ended work.
• FrontierCode Main primary is Xhigh 52.1%. Max is 46.2% because code-review subagents timed out or made out-of-scope edits.
• OSWorld 2.1 partial (80.1%, announcement) and strict (43.5%, system card) are the same suite; do not cite partial alone. Suite is 2.1, not 2.0.
• Prefer Legal all-pass 11.7% over mean criterion-pass 92.1%. Prefer HealthBench Professional length-adjusted 69.2% (Table 8.1.A; raw 77.1%); PhysicianBench is 63.2%. Plain HealthBench is raw 69.4% with an LA band only. Bio §8.17 rows are capability/knowledge only.
• System Card §8 software-engineering and domain rows are absolute self-reports with no peer deltas in this post. Safety, refusal, and jailbreak tables are intentionally omitted.
• A third-party Elo run for GDPval-AA / AA-Briefcase used a pre-release Platform deployment with a later-fixed structured-outputs bug that may slightly understate scores.
• Knowledge cutoff is June 2026.

Outlook

Anthropic says Claude Haiku 5.5 will follow in the coming weeks. This Sonnet post ages when LLM Stats verifies Terminal-Bench and CursorBench, and when Haiku 5.5 lands.

Through-line for now: second Claude 5.5 model, faster everyday coding and polished knowledge work beside Opus, same Sonnet sticker with fewer tokens and a 30%+ speed claim. For the primary sources, see Anthropic's Sonnet 5.5 announcement, the platform overview, and the system card.