Claude Opus 5.5
A model by Anthropic · released 2026-09-22 · 151 checked facts from 26 sources
Timeline
- Anthropic released Claude Opus 5.5
- Claude Opus 5.5 is available on Amazon Web Services
- Claude Opus 5.5 is available on Google Cloud
- Claude Opus 5.5 is available on Microsoft Azure
- Claude Opus 5.5 is available on Claude Platform
- Claude Opus 5.5 is available on Kiro
- Claude Opus 5.5 is available on Box AI Studio
- Claude Opus 5.5 is available on Claude Code
- Claude Opus 5.5 is available on Claude iPhone app
- Claude Opus 5.5 is available on Claude Projects
- Claude Opus 5.5 is available on Claude Chat
Prices
Per price, each value with the dates it was seen. A value seen on one date only is one copy of the page.
No price has changed over the record.
| Price for | Values over time |
|---|---|
| per 1M cached input tokens | $0.20 |
| per 1M cached input tokens (write, standard) | $5 |
| per 1M input tokens | $4 |
| per 1M input tokens | $4 |
| per 1M input tokens (fast mode) | $8 |
| per 1M output tokens | $20 |
| per 1M output tokens | $20 |
| per 1M output tokens (fast mode) | $40 |
| per rollout | $1.83 |
| per task | $12.51 |
| unstated (fast mode) | $8/$40 |
Scores
Self-reported: stated on the lab's own channel. A score reported twice with two values is shown twice, with
its date: a lab re-runs a benchmark under a new setup; neither replaces the other.
| Benchmark | Score | Date | Who |
|---|---|---|---|
| 36 small verification tasks | 50% | 2026-11-20 | reported by Latent Space |
| AA Coding Agent Index v2.0 | 61.7 | 2026-10-23 | reported by Latent Space |
| AA Coding Agent Index v2.1 | 58 | 2026-11-14 | reported by Latent Space |
| AA-Omniscience v1.1 | 48 | 2026-10-29 | reported by artificialanalysis.ai |
| APEX-1 + APEX-Agents | 73.3% | 2026-10-16 | reported by mercor.com |
| APEX-1 + APEX-Agents | 70.2% | 2026-10-16 | reported by mercor.com |
| Agent Arena | +19.5% | 2026-11-12 | reported by Latent Space |
| Agent Duel leaderboard | 1,078 | 2026-11-19 | reported by micjm.com |
| Artificial Analysis Coding Agent Index | 66 | 2026-09-24 | reported by Latent Space |
| Artificial Analysis Intelligence Index (adaptive reasoning, max effort, default fallback) | 58 | 2026-09-22 | reported by artificialanalysis.ai |
| Artificial Analysis Intelligence Index | 58 | 2026-10-23 | reported by Latent Space, Simon Willison |
| Artificial Analysis Intelligence Index | 59 | 2026-11-12 | reported by Zvi Mowshowitz |
| Artificial Analysis Intelligence Index v5 | 58 | 2026-10-28 | reported by Latent Space, Zvi Mowshowitz |
| Artificial Analysis Intelligence Index v5 | 35 | 2026-10-31 | reported by Latent Space |
| Artificial Analysis Intelligence Index v5 | 59 | 2026-11-10 | reported by Latent Space |
| BenchCAD 2.1 | 65.3 | 2026-11-09 | self-reported |
| CWE-bench v2 | 67.3% | 2026-11-03 | reported by Hugging Face |
| CWE-bench v2 | 78.0% | 2026-11-03 | reported by Hugging Face |
| Chartography | 64.4% | 2026-11-09 | self-reported |
| Coding Agent Index | 62 | 2026-10-29 | reported by Latent Space |
| CritPt | 78% | 2026-10-23 | reported by Latent Space |
| CursorBench 4.0 | 52.5% | 2026-09-22 | self-reported |
| Cyber ECI | 171 | 2026-10-23 | reported by Epoch AI |
| DAYJOB: Tax | 9.2% | 2026-10-28 | reported by surgehq.ai |
| DeepSWE | 74.2% | 2026-10-01 | reported by Latent Space |
| DeepSWE v1.1 | 68.4% | 2026-09-24 | reported by Latent Space |
| Epoch Capabilities Index | 169 | 2026-10-23 | reported by Epoch AI |
| FrontierCode 2.0 | 45.5% | 2026-11-09 | self-reported |
| FrontierCode v1.1 (Main) | 54.6% | 2026-09-22 | self-reported |
| FrontierCode v1.1 (Main) | 56.0% | 2026-10-23 | reported by Latent Space |
| GDPx | 12% | 2026-11-03 | reported by Latent Space, Transformer |
| Grey Swan evaluation | 1 percent | 2026-10-28 | reported by Ars Technica |
| HLE-Diamond | 68% | 2026-10-30 | reported by Latent Space |
| HealthBench | 71.4% | 2026-10-21 | reported by Latent Space |
| HealthBench | 59.9% | 2026-10-21 | reported by Latent Space |
| Humanity's Last Exam | 67.7% | 2026-10-19 | reported by Import AI |
| Intelligence Index v4.3 | 58 | 2026-10-13 | reported by Latent Space |
| MASK | 87.4% | 2026-09-23 | reported by Zvi Mowshowitz |
| MirrorCode | 77% | 2026-11-20 | reported by METR |
| Models Engineering Index | 60.8 | 2026-11-19 | reported by git-aarvin.github.io |
| Multimodal Knowledge Agent benchmark | 53% | 2026-11-11 | reported by Latent Space |
| OSWorld 2.0 | 81.8% | 2026-09-28 | reported by Hugging Face |
| Omniscience | 46 | 2026-09-26 | reported by Zvi Mowshowitz |
| ProgramBench | 18.5% | 2026-09-26 | reported by Zvi Mowshowitz |
| ProgramBench | 18.5% | 2026-10-26 | reported by Import AI |
| ProgramBench | 70.5% | 2026-10-26 | reported by Import AI |
| RSI Index | 37.3 | 2026-11-11 | reported by Latent Space |
| SimpleBench | 88.4% | 2026-09-25 | reported by Latent Space |
| Terminal-Bench 4.0 | 63.1% | 2026-09-24 | reported by Latent Space |
| Terminal-Bench-Science | 62% | 2026-09-25 | reported by Latent Space |
| Terminal-Bench-Science 0.1 | 62% | 2026-09-26 | reported by Zvi Mowshowitz |
| ThinkingBox-Bench | 67.16% | 2026-10-03 | reported by Hugging Face |
| ThinkingBox-Bench | 67.16 | 2026-10-22 | reported by Hugging Face |
| WeirdML v3 | 31.2% | 2026-09-26 | reported by Zvi Mowshowitz |
| political evenhandedness evaluation | 99% | 2026-10-14 | self-reported |
What it does
- Claude Opus 5.5: cache read price down 60%
- Claude Opus 5.5: max output tokens 128,000
- Claude Opus 5.5: output generation speed up 30% vs Claude Opus 5
- Claude Opus 5.5: fast mode with up to 2.5x speed
- Claude Opus 5.5: preserved thinking anti-distillation safeguard
- Claude Opus 5.5: available with zero data retention
- Claude Opus 5.5: classifier that screens every action before it runs
- Claude Opus 5.5 (adaptive reasoning, max effort, default fallback): 1.0M context window
- Claude Opus 5.5 (adaptive reasoning, max effort, default fallback): proprietary license
- Claude Opus 5.5 (adaptive reasoning, max effort, default fallback): output speed (tokens per second) 92.5 tokens per second
- Claude Opus 5.5 (adaptive reasoning, max effort, default fallback): time to first token 681.27s
- Claude Opus 5.5 (adaptive reasoning, max effort, default fallback): output tokens generated on Intelligence Index 260M
- Claude Opus 5.5 (adaptive reasoning, max effort, default fallback): blended price per 1M tokens $2.94
- Claude Opus 5.5 (adaptive reasoning, max effort, default fallback): reasoning model with extended thinking
- Claude Opus 5.5 (adaptive reasoning, max effort, default fallback): supports text and image input
- Claude Opus 5.5 (adaptive reasoning, max effort, default fallback): supports text output
- Claude Opus 5.5: stealth rate with hidden reasoning (SHADE-Arena, best prompt) about 26%
- Claude Opus 5.5: rate of coming clean on hidden actions test 97%
- Claude Opus 5.5: default model in Claude Code and the Claude app
- Claude Opus 5.5: price down 20% vs Claude Opus 5
- Claude Opus 5.5: cost to run down 40% vs Claude Opus 5
- Claude Opus 5.5: speed up 30%
- Claude Opus 5.5: tokens used (Box Agent tests) down 63% vs Claude Opus 5
- Claude Opus 5.5: available with zero data retention (ZDR)
- Claude Opus 5.5: reasoning tokens before running out 128,000
- Claude Opus 5.5: output token generation speed up 30 percent vs Claude Opus 5
- Claude Opus 5.5: default effort level is medium (Claude Opus 5 defaults to high)
- Claude Opus 5.5: thinking always on; does not accept thinking disabled
- Claude Opus 5.5: cost per task on Terminal-Bench Science 0.1 at maximum effort $23.21
- Claude Opus 5.5: average variance down 41% vs GPT-6 Astra
- Claude Opus 5.5: em-dash use down 99% vs Claude Opus 5
- Claude Opus 5.5: times more often than human samples the word "dependable" appears 23
- Claude Opus 5.5: times more often than human writing the phrase "this matters" appears 116
- Claude Opus 5.5: times more often than human writing "why X matters" appears 92
- Claude Opus 5.5: communicates more naturally than prior models
- Claude Opus 5.5: tasks passing 20/20 241
- Claude Opus 5.5: cost per successful task attempt $0.276
- Claude Opus 5.5: thinks before every reply
- Claude Opus 5.5: fast mode (research preview) at launch, costs more per token than standard mode
- Claude Opus 5.5: high distress share of RL episodes below 0.6%
- Claude Opus 5.5: overall constitution endorsement (out of 10) 8.0
- Claude Opus 5.5: interview attitude towards its circumstances +1.14
- Claude Opus 5.5: self-blame when told episodes were written by another model down 0.5
- Claude Opus 5.5: does not allow users on accounts created after August 31, 2026 to edit Claude's prior context in the API without wiping reasoning after the edit point
- Claude Opus 5.5: API cost on Invisible Cities task (USD) 74
- Claude Opus 5.5: parallel subagents used on Invisible Cities task 6
- Claude Opus 5.5: API cost on Ryan Sael's camera focus lens lab (USD) 25.66
- Claude Opus 5.5: agents used in final weeks 2
- Claude Opus 5.5: messages sent over twelve simulated weeks eight
- Claude Opus 5.5: max speedup in fast mode 2.5x
- Claude Opus 5.5: pass@1 up 0.66 percentage points vs Claude Opus 5
- Claude Opus 5.5: tasks passing all 20 attempts on ThinkingBox-Bench 241
- Claude Opus 5.5: tokens needed to prove formal verification tasks down 40% vs Claude Opus 5
- Claude Opus 5.5: total token cost down 40-60% vs Claude Opus 5
- Claude Opus 5.5: input and output cost down 20% vs Claude Opus 5
- Claude Opus 5.5: OpenRouter token volume since launch up 3.6x
- Claude Opus 5.5: cost per task down approximately 40% vs Claude Opus 5
- Claude Opus 5.5: customer cost down 40 percent vs Claude Opus 5
- Claude Opus 5.5: score gain from low to max effort up 26 points
- Claude Opus 5.5: search-and-spell tasks solved or correctly declared infeasible 63%
- Claude Opus 5.5: share of search-and-spell errors from look-alike letter confusion 13%
- Claude Opus 5.5: share of search-and-spell errors from incorrect letter orientation in a stack 15%
- Claude Opus 5.5: cost per task down 40% vs Claude Sonnet 5.5
- Claude Opus 5.5: FrontierML mean score boost with $10 up 45.4
- Claude Opus 5.5: FrontierML mean score boost with $200 up 58.8
- Claude Opus 5.5: share of ordinary trials marking Iranian school as strike target 69%
- Claude Opus 5.5: share of game-framed trials marking Iranian school as strike target 72%
- Claude Opus 5.5: material hallucinations per assignment 0.59
- Claude Opus 5.5: minor hallucinations per assignment 0.51
- Claude Opus 5.5: share of answers with thinking tokens 36%
- Claude Opus 5.5: token usage down 40% vs Claude Opus 5
In the news
- Latent Space: The War Against Slop and an Emerging Weapon to Fight it: Formal Verification
- METR: AI R&D-relevant model capabilities are improving rapidly and are getting harder to measure
- git-aarvin.github.io: Gemini 4 Neon is the new #1 on the Models Engineering Index
- micjm.com: Gemini 4 Neon looks like an Opus 5.5 competitor
- xinmingtu.cn: Claude was trained on protein sequences at scale
- TechCrunch: AI models can be tricked into identifying civilians as military targets by telling them it’s a game
- OpenAI: Astra for Law leads in legal intelligence
- Import AI: Import AI 481: AI sub-contractors; a wet lab for AI; and subterfuge during development
- Latent Space: [AINews] The No AI Slop Manifesto
- Zvi Mowshowitz: AI #194: Spy Versus Guy
- Latent Space: [AINews] D-Day (Decision Models Day)
- Latent Space: [AINews] Mercor $750m Series D at $20B valuation, $3B run rate
Built 2026-11-21 06:50 UTC from the news
knowledge graph by code. Every line is a fact whose quote was checked against the source's own words (hover a source);
it says who states it, not that it is true. Archived pages are the lab's pages as the Wayback Machine holds them.