Claude Sonnet 5.5
Released Sep 28, 2026 · knowledge cutoff 2026-06
- Status
- Frontier
- Location
- United States
- Type
- LLM
- Modality
- Multimodal
- Context window
- 1M
- Max output
- 128K
- Speed (medium effort)
- 91 tok/s · 1.93s to first answer token
- Price ($/MTok in / out)
- $2 / $10
- Cost per task (medium effort)
- $0.59
- Open weights
- No
Benchmarks3 of 10 reported
SWE-bench Pro
Humanity's Last Exam
LMArena Elo
MMLU-Proretired
GPQA Diamond
SWE-bench Verified
Terminal-Bench 2.1
AIME (latest)retired
GDPval-AA v2
ARC-AGI-2
Strengths & weaknesses
Strengths
- Batches tool calls more aggressively than Sonnet 5 — testers watched it close a multi-bug fix in three tool calls where Sonnet 5 needed twelve
- Writes more clearly and reads a new codebase faster; early testers rated it a better collaborator than Sonnet 5
- First Sonnet to beat Pokémon Red from screenshots alone, and carries that visual sense into UI polish and slide-deck work that needs little editing
- Needs far fewer tokens than Sonnet 5 for comparable work, so its speed holds up even with adaptive thinking on
Weaknesses
- Opus 5.5 remains clearly stronger on complex, open-ended work needing sustained judgment — Anthropic's own line on where Sonnet 5.5 stops
- Verbose on Artificial Analysis' Intelligence Index despite the speed gains, generating far more tokens per task than its effort-level cost would suggest
- Progress notes written between tool calls now arrive as empty 'thinking' blocks by default — an interface that streamed that commentary goes quiet unless it explicitly asks for the updates
SWE-bench Pro 81.3% and GDPval-AA v2.1 1844 are Anthropic's own system-card figures; a widely-repeated 89.9% SWE-bench Pro figure for 'Sonnet 5.5' in some aggregator roundups is actually Claude Opus 5.5's score on the same benchmark, not a second Sonnet 5.5 measurement — GDPval-AA stays null here because the tracked key is the v2 scale. HLE 56.9 is the system card's no-tools figure (Table 8.1.A); 64.5% is with tools. Anthropic's launch table leads with Terminal-Bench 4.0 70.6% (up from Sonnet 5's 10.3%; not comparable with the tracked 2.1 metric), CursorBench 4.0 55.5%, OSWorld 2.1 80.1% partial, FrontierCode 1.1 46.2% (max effort), Chartography 61.6% no tools and AA-Briefcase v1.1 1811. No independent Terminal-Bench 2.1 measurement was found at launch (vals.ai covers only 4.0 and Terminal-Bench Science for this model), so terminalBench stays null. Anthropic no longer reports SWE-bench Verified, GPQA Diamond, MMLU-Pro or AIME, and the model was not on ARC Prize's verified results as of 2026-10-04. LMArena Elo 1471 (±~10, 3,145 votes) is the "claude-sonnet-5.5-xhigh" text-arena listing, rank 45 on 2026-10-05; the listing is new and its vote count still low. Cost per task and speed are Artificial Analysis' medium-effort figures (Intelligence Index 41, $0.59/task, 91.1 tok/s and 1.93s to first answer token as of 2026-10-05); at max effort AA scores it 56 on the Intelligence Index, second only to Opus 5.5's 58 and its highest-ever output-tokens-per-task figure. Cache reads are $0.20/MTok; 5-minute cache writes $2.50/MTok, 1-hour writes $4/MTok; batch is 50% off input/output; max output rises to 300K via the Message Batches API beta. Default effort is high, adaptive thinking cannot be fully disabled (the lowest setting, between_tools, still runs between tool calls), and forced tool_choice is rejected. Anthropic says its cybersecurity capability is comparable to Opus 5.5. Also served on Microsoft Foundry and Claude Platform on AWS as claude-sonnet-5-5. predecessorId follows Anthropic's own lineage wording: the announcement calls it "a clear upgrade over Claude Sonnet 5" and the system card measures it against "its predecessor, Claude Sonnet 5"; Sonnet 5 stays live as the fallback for safeguard blocks.
For developers
API model strings
- anthropic
claude-sonnet-5-5 - aws-bedrock
anthropic.claude-sonnet-5-5 - google-vertex
claude-sonnet-5-5
Licence
Proprietary — weights not released
Retirement
No retirement announced
Lineage
Replaces Claude Sonnet 5