Anthropic's Claude Opus 5.5 Tops Coding Benchmarks but Costs $13 per Task
Anthropic's newest flagship tops the Coding Agent Index at 66, but burns 37% more tokens per task than its predecessor to get there.
- Claude Opus 5.5 tops the Artificial Analysis Coding Agent Index with a score of 66, the highest ever measured.
- Gains across all three evals: Terminal-Bench 4.0 up 8.6 points to 63.1%, DeepSWE v1.1 to 68.4%, SWE-Atlas-QnA to 66.4%.
- API pricing cut to $4/$20 per million input/output tokens, with cache reads dropping 60% to $0.20 per million.
- Despite lower prices, Cost per Task rises 21% to $13.04 because the model uses 15.6M tokens vs 11.4M for Opus 5.
- Output tokens per task more than double, from 137k to 333k, reflecting deeper adaptive thinking at max effort.
- First model in the Claude 5.5 family; Sonnet 5.5 and Haiku 5.5 expected in the coming weeks.
Claude Opus 5.5 leads coding index at $13.04 per task
Anthropic’s Claude Opus 5.5 now leads the Coding Agent Index with a composite score of 66, the highest result the benchmark has recorded. Running in Claude Code at maximum reasoning effort, the model finishes six points ahead of Opus 5 and four ahead of Claude Fable 5.1. That configuration uses about 37% more tokens per task than Opus 5 and raises estimated API cost by 21%, despite lower unit prices.
Shell work drives the lead
The index combines three equally weighted evaluations covering shell use, software engineering, and repository comprehension. Each benchmark reports pass@1, which measures success in a single run, with results averaged across three attempts per task.
| Evaluation | Tasks | Opus 5.5 | Opus 5 | Change |
|---|---|---|---|---|
| Terminal-Bench 4.0 Agentic shell and command-line work |
66 | 63.1% | 54.5% | +8.6 points |
| DeepSWE v1.1 Software engineering tasks |
113 | 68.4% | 62.5% | +5.9 points |
| SWE-Atlas-QnA Repository-understanding questions |
124 | 66.4% | 62.1% | +4.3 points |
The 8.6-point Terminal-Bench gain accounts for the largest improvement. That evaluation requires an agent to navigate a shell, use tools correctly, recover from intermediate failures, and complete multi-step command-line workflows, making it the closest of the three to long-running coding-agent work.
Heavy token use lifts the bill
The published pricing details cut standard input and output rates by 20%. Cache writes, which store reusable context, also fall by 20%. Cache reads, which retrieve that context in later requests, drop by 60%.
| Token category | Opus 5.5 | Opus 5 | Change |
|---|---|---|---|
| Input | $4 per million | $5 per million | -20% |
| Output | $20 per million | $25 per million | -20% |
| Cache write | $5 per million | $6.25 per million | -20% |
| Cache read | $0.20 per million | $0.50 per million | -60% |
At maximum effort, Opus 5.5 consumes enough additional tokens to outweigh those rate cuts. Artificial Analysis estimates cost by applying API rates to the benchmark’s input, output, and cache usage.
| Per-task measure | Opus 5.5 | Opus 5 | Change |
|---|---|---|---|
| Estimated cost | $13.04 | $10.79 | +21% |
| Total tokens | About 15.6 million | About 11.4 million | +37% |
| Output tokens | About 333,000 | About 137,000 | +143% |
| Cached input | About 14.6 million | About 10.9 million | +34% |
Output-token use rises to roughly 2.4 times the Opus 5 level, while cached input grows by about one-third. On the benchmark’s score-versus-cost Pareto chart, which tracks the best score available at each spending level, Opus 5.5 extends the frontier at the expensive end. Fable 5.1 and Opus 5 remain cheaper options, with composite scores four and six points lower, respectively.
Maximum effort explains the gap
Anthropic estimates that Opus 5.5 matches Fable 5.1 on most work while costing 40% less to run than Opus 5 under default settings. The company attributes that estimate to lower token rates, reduced serving compute, and fewer tokens consumed per task.
Artificial Analysis configured Claude Code at maximum effort. Opus 5.5 keeps adaptive thinking enabled and uses an effort parameter to control how much reasoning it performs. Anthropic does not list a separate rate for reasoning tokens, so they are billed at the standard output price. Higher effort can therefore erase the savings from lower unit rates.
- Maximum effort: The index records the highest composite score alongside a $13.04 estimated task cost.
- Lower effort: Anthropic expects reduced token use and lower costs, but the index results provided here do not quantify the corresponding score or savings.
Faster and asynchronous paths
Opus 5.5 is the first model in the Claude 5.5 family, with Sonnet 5.5 and Haiku 5.5 expected in the following weeks. Claude Code and the Claude Platform also add two execution options for workloads that prioritize latency or throughput.
| Mode | Delivery | Pricing | Additional limit |
|---|---|---|---|
| Fast mode | Up to 2.5 times faster | $8 input and $40 output per million tokens | None announced |
| Batch processing | Asynchronous, without an immediate response guarantee | 50% off standard input and output rates | Up to 300,000 output tokens |
Anthropic-selected early testers reported completing large code migrations and audits in hours instead of days. They also said Opus 5.5 identified performance bottlenecks while changing less application behavior than Opus 5. These accounts provide workload examples, while the index supplies the controlled comparison.
Where each configuration fits
- Long-running, terminal-heavy agents: Opus 5.5 at maximum effort offers the highest measured score and an 8.6-point Terminal-Bench gain for an additional $2.25 per benchmark task.
- High-volume or shorter tasks: Lower Opus 5.5 effort settings, Fable 5.1, and Opus 5 warrant direct comparison when throughput and cost outweigh a four-to-six-point composite gap.
- Latency-sensitive tools: Fast mode trades doubled input and output rates for up to 2.5 times faster execution.
- Large asynchronous jobs: Batch processing halves standard input and output rates and supports up to 300,000 output tokens, fitting migrations and bulk generation that can tolerate delayed results.
Before a production switch, teams should run representative repositories and prompts at several effort levels, then record completion rate, retries, latency, input and output tokens, cache usage, and cost per successful task. The composite establishes Opus 5.5’s lead under maximum effort, while workload-specific testing determines whether that lead offsets its higher task cost.
Comments
