Anthropic Launches Claude Opus 5.5 at $4/$20 Per 1M Tokens, Takes #1 on AA Intelligence Index
Anthropic released Claude Opus 5.5 on September 22, taking the top position on the Artificial Analysis Intelligence Index. LiveBench September 2026 puts it at 83.2 overall — second to Claude Fable 5.1 (83.4) and 34% cheaper per successful task.
Pricing is $4 per million input tokens and $20 per million output tokens. Context window is 1M tokens.
What Differentiates It From Opus 5
Anthropic describes a model that completes comparable tasks in fewer steps and fewer tokens than its predecessor, with clearer natural-language reporting on progress, findings, and blockers. Thinking is always adaptive: effort level is the primary dial for trading depth against latency and cost. Lower-effort settings remain effective for latency-sensitive workloads.
Benchmark Position
On LiveBench September 2026, Opus 5.5 Thinking (Max Effort) scores:
| Category | Score |
|---|---|
| Overall | 83.2 |
| Reasoning | 92.2 |
| Coding | 89.3 |
| Agentic Coding | 71.7 |
| Mathematics | 97.1 |
The 71.7 Agentic Coding score is the highest among closed frontier models on the September run. DeepSeek V4.1 Flash scores higher at 77.3 but is an open-weight model priced at $0.029 per successful task versus Opus 5.5’s $0.799.
On Artificial Analysis, Opus 5.5 (max with fallback) and Opus 5.5 (xhigh with fallback) rank as the top two intelligence models as of September 22. Claude Fable 5.1 follows in fourth.
Cost per Task Comparison
| Model | LiveBench Overall | Cost/Task* |
|---|---|---|
| Claude Fable 5.1 Max Effort | 83.4 | $1.212 |
| Claude Opus 5.5 Max Effort | 83.2 | $0.799 |
| GPT-6 Astra Max Effort | 82.2 | $0.736 |
| Muse Spark 1.3 xHigh | 81.6 | $0.219 |
| DeepSeek V4.1 Flash | 81.1 | $0.029 |
*Cost per successful task as reported on LiveBench September 2026 leaderboard.
At $0.799 per successful task, Opus 5.5 is 34% cheaper than Fable 5.1 while matching it on overall score and outrunning it by 5.6 points on Agentic Coding.
Pricing in Context
| Model | Input | Output |
|---|---|---|
| Claude Opus 5.5 | $4/M | $20/M |
| Claude Opus 4.7 | $5/M | $25/M |
| Claude Opus 4.8 | $5/M | $25/M |
Opus 5.5 is the cheapest Opus-class API model Anthropic has released. Claude Opus 5’s standard API rate was $5/$25 per million tokens — the same as Opus 4.7 and 4.8. Opus 5.5 at $4/$20 undercuts all three.
Deprecations
As of September 1, the standalone Claude Opus 5 Fast dedicated model is deprecated. Fast mode now runs through service_tier: "fast" on the standard Opus 5 endpoint; existing integrations are unaffected. Claude Opus 4.8 Fast has been similarly folded into the unified fast tier on Claude Opus 4.8.
SWE-bench Verified and tau-bench results for Opus 5.5 are not yet available. LiveBench Agentic Coding (71.7) is an auxiliary signal until primary benchmarks land.