Claude Opus 5.5
Anthropic's September 2026 Opus release: the highest-scoring model on the Artificial Analysis Intelligence Index v4.3, at a lower list price than Claude Opus 5.
Model Story
Claude Opus 5.5 was released by Anthropic on September 22, 2026. It sits in a crowded Anthropic lineup: Claude Fable 5.1, released on September 1, 2026 at $10 input / $50 output per 1M tokens, is the premium flagship; Claude Sonnet 5.5 posts an Anthropic-reported 81.3% on SWE-bench Pro; and Claude Mythos 5 and 5.1 remain restricted-access models.
On independent measures, Opus 5.5 is the model to beat. Artificial Analysis lists it at 58 on its Intelligence Index v4.3.x at max effort, the top score of any model, ahead of GPT-6 Astra, Claude Fable 5.1 and Gemini 4 Argon at 53. It scores 56 at xhigh effort and 42 at low effort, so the effort setting is a real lever on both quality and cost.
It is also a rare price cut at the top end. Opus 5.5 lists at $4 input / $20 output per 1M tokens, 20% below Claude Opus 5 ($5 / $25), and Anthropic says typical workloads cost around 40% less than on Opus 5 because the model uses fewer tokens. The broader pattern still holds, though: Anthropic's own flagship, Fable 5.1, costs two and a half times as much per token, and per-token bills grow with usage.
For Faraday Machines customers, Claude Opus 5.5 is typically a hybrid component: sensitive data stays on-premises with open-weight models, while specific high-complexity tasks are routed to Claude via API.
Key Specifications
| Developer | Anthropic |
|---|---|
| Release Date | September 22, 2026 |
| Architecture | Undisclosed (proprietary) |
| License | Proprietary (API / enterprise license) |
| Effort Levels | Selectable; Artificial Analysis publishes low, xhigh, and max results |
| Fast Mode | $8 input / $40 output per 1M tokens |
| Batch Processing | Half price |
| Premium Sibling | Claude Fable 5.1 ($10 / $50 per 1M tokens) |
| Other Siblings | Claude Sonnet 5.5; Claude Mythos 5 / 5.1 (restricted access) |
| Predecessor | Claude Opus 5 ($5 / $25 per 1M tokens) |
API Pricing
Claude Opus 5.5 is priced 20% below Claude Opus 5. Batch processing is half price, and fast mode costs $8 input / $40 output per 1M tokens. At max effort, Artificial Analysis measures $5.98 per Intelligence Index task, falling to $3.46 at xhigh and $0.55 at low effort. Cheaper list prices help, but cloud prices move in both directions without warning, and per-token bills still scale with every query your team runs.
Benchmarks
The 89.9% SWE-bench Pro figure is Anthropic's own. Vendor-reported SWE-bench Pro scores are not directly comparable: OpenAI's July 2026 audit estimated that around 30% of the public tasks are broken, and the top score on Scale's standardized public set is around 61.5%. The Artificial Analysis scores use the v4.3 index introduced in September 2026, which is not comparable with v4.1 scores on older profiles.
On-Premises Considerations
Claude Opus 5.5 is a proprietary model with no open-weight release, so it cannot be self-hosted on Faraday Machines hardware. Faraday can architect hybrid deployments where your data pipeline runs locally on open-weight models, with selective API calls to Claude for tasks that need its specific reasoning strengths.
For work that has to stay on-premises, the open-weight models Faraday ships cover each tier: Qwen3.8-27B on Starter; GLM-5.3-Flash and MiMo-V2.6-Flash on Growth; and MiniMax M3 and DeepSeek V4.1 Flash on Scale. GLM-5.3-Flash scores 42 on the Artificial Analysis Intelligence Index v4.3 at $0.25 per task, the same score as Opus 5.5 at low effort ($0.55 per task). On larger clusters, MiMo-V2.6-Pro leads open-weight models at 46, still 12 points behind Opus 5.5 at max effort. The gap to the very top is real; so is the cost difference.
Access is the other consideration. On June 12, 2026, a US Commerce Department order led Anthropic to disable Claude Fable 5 and Mythos 5 for everyone. The order was lifted on June 30 and access was restored the next day, and Fable 5.1 is now generally available on the Claude API, AWS, Google Cloud and Azure. It is still a "Covered Model", so organizations under zero-data-retention agreements need express authorization from Anthropic. The lesson stands: access to a cloud model can be switched off overnight. Models running on hardware you own cannot.
Frequently Asked Questions
Can I run Claude Opus 5.5 on-premises?
No. Claude Opus 5.5 is a proprietary Anthropic model with no open-weight release, so it cannot be self-hosted on Faraday Machines hardware. Faraday Machines can build hybrid deployments in which your data pipeline runs locally on open-weight models such as GLM-5.3-Flash or DeepSeek V4.1 Flash, with selective API calls to Claude for tasks that need its specific reasoning strengths.
How much does Claude Opus 5.5 cost?
Claude Opus 5.5 lists at $4 per 1M input tokens and $20 per 1M output tokens, 20% below Claude Opus 5 ($5 / $25). Cache reads cost $0.20 per 1M tokens, batch processing is half price, and fast mode costs $8 / $40. Artificial Analysis measures $5.98 per Intelligence Index task at max effort, $3.46 at xhigh and $0.55 at low effort.
Is Claude Opus 5.5 the highest-scoring AI model?
Yes, on the Artificial Analysis Intelligence Index v4.3.x: Claude Opus 5.5 at max effort scores 58, the top score of any model, ahead of GPT-6 Astra, Claude Fable 5.1 and Gemini 4 Argon at 53. It scores 56 at xhigh effort and 42 at low effort. Anthropic separately reports 89.9% on SWE-bench Pro, a vendor-reported figure that is not directly comparable with standardized runs.
How does Claude Opus 5.5 compare to open-weight models?
Claude Opus 5.5 leads the best open-weight model by 12 points on the Artificial Analysis Intelligence Index v4.3: 58 at max effort versus 46 for MiMo-V2.6-Pro. At low effort the gap disappears: GLM-5.3-Flash, which Faraday Machines runs on its two-unit Growth tier, scores 42 at $0.25 per task, the same score as Opus 5.5 at low effort ($0.55 per task).
Explore Other Models
See how every model compares on price, benchmarks and hardware fit in the AI Model Comparison.