AI Charts
In the AI Charts coding-agent snapshot retrieved Sep 25, 2026, 2:39 PM UTC, Claude Code · Opus 5.5 (max) scores 66.0 on AA Index at a mean API cost of $13.04 per task, first of the 20 configurations that carry an index and the highest cost per task on the chart. The nearest cost-frontier configuration below it, Claude Code · Fable 5.1 (with fallback) (max), gives up 3.8 points for 95% of the cost. Anthropic released Claude Opus 5.5 on September 22, 2026, and Artificial Analysis added the Claude Code row to its coding-agents comparison afterwards. The snapshot’s update log records the row on September 25, 2026.
One model inside one harness, scored on three benchmarks
The coding-agent chart is a daily snapshot of the public Artificial Analysis coding-agents comparison. Each row is one configuration: a model, the agent harness that ran it, and an effort setting. Artificial Analysis runs the harness on tasks from three coding benchmarks, DeepSWE v1.1, Terminal-Bench 4, and SWE-Atlas-QnA, and AA Index is the mean of the three scores. The cost is the mean API bill for one task in that harness at list prices, including every tool call and every repeated read of the repository the harness sends.
Claude Opus 5.5’s row is the model inside Claude Code, Anthropic’s own coding agent, at the max setting. One task in that configuration used 15.6 million tokens and took 64 minutes of harness time on average. The same model in another harness, or at another setting, would be another row; the snapshot stores one configuration running Claude Opus 5.5.
First of 20 configurations, at the chart’s highest cost
The 66.0 places Claude Code · Opus 5.5 (max) first of the 20 configurations that carry an AA Index in the snapshot retrieved Sep 25, 2026, 2:39 PM UTC. No configuration scores higher. Its $13.04 per task is also the highest cost of the 20 configurations that carry a cost, so the row sits at the top right of the chart as both the highest-scoring configuration and the most expensive to run one task through.
The nearest score below it is Claude Code · Fable 5.1 (with fallback) (max) at 62.2, 3.8 points lower for 95% of the Opus 5.5 row’s cost. The table lists the four highest-scoring configurations after it, with each cost as a share of its $13.04.
| Configuration | Setting | AA Index | Cost per task | Points below Opus 5.5 | Share of Opus 5.5’s cost |
|---|---|---|---|---|---|
| Claude Code · Fable 5.1 (with fallback) | max | 62.2 | $12.39 | 3.8 points | 95% |
| Devin Fusion CLI · Claude Fable 5.1 XHigh + SWE-2 Medium | default | 61.7 | $7.90 | 4.3 points | 61% |
| Codex · GPT-6 Astra | max | 61.6 | $7.47 | 4.3 points | 57% |
| Claude Code · Opus 5 | max | 59.7 | $10.79 | 6.3 points | 83% |
What stepping down the cost frontier gives up
The chart’s cost frontier is the set of configurations that no other configuration beats on both axes: nothing scores at least as high for the same or less money. Claude Code · Opus 5.5 (max) is on it, and as the highest-scoring row it is the frontier’s top vertex by definition: any row with the top score and a cost is on the frontier whatever it costs. The question the frontier answers is what a reader gives up by stepping down from the top score to a cheaper row that nothing dominates.
The first step down is Claude Code · Fable 5.1 (with fallback) (max): 3.8 points lower for 95% of the $13.04. The first vertex at half the cost or less is Devin Fusion CLI · GPT-6 Astra XHigh + SWE-2 Medium (default), which gives up 7.1 points for 35% of the cost. The frontier ends at Codex · DeepSeek V4 Flash 0731 (max): 27.2 points lower for 0.7% of the cost.
| Configuration | Setting | AA Index | Cost per task | Points below Opus 5.5 | Share of Opus 5.5’s cost |
|---|---|---|---|---|---|
| Claude Code · Fable 5.1 (with fallback) | max | 62.2 | $12.39 | 3.8 points | 95% |
| Devin Fusion CLI · Claude Fable 5.1 XHigh + SWE-2 Medium | default | 61.7 | $7.90 | 4.3 points | 61% |
| Codex · GPT-6 Astra | max | 61.6 | $7.47 | 4.3 points | 57% |
| Devin Fusion CLI · GPT-6 Astra XHigh + SWE-2 Medium | default | 58.9 | $4.54 | 7.1 points | 35% |
| Codex · GPT-6 Sol | max | 56.7 | $2.99 | 9.3 points | 23% |
| Codex · GPT-5.6 Luna | max | 43.2 | $0.438 | 22.8 points | 3.4% |
| Codex · DeepSeek V4 Pro 0813 | max | 43.1 | $0.238 | 22.9 points | 1.8% |
| Codex · GPT-6 Luna | max | 41.1 | $0.176 | 24.9 points | 1.3% |
| Codex · DeepSeek V4 Flash 0731 | max | 38.7 | $0.085 | 27.2 points | 0.7% |
Terminal-Bench 4 and SWE-Atlas-QnA carry the lead
AA Index is the mean of three component benchmarks, and the Opus 5.5 row does not lead all three. On DeepSWE v1.1 it scores 68.4, sixth of 20; on Terminal-Bench 4 it scores 63.1, first of 20; and on SWE-Atlas-QnA it scores 66.4, first of 20 configurations that carry each score.
It leads Terminal-Bench 4 by 5.6 points over Claude Code · Fable 5.1 (with fallback) (max) and SWE-Atlas-QnA by 0.3 points over Kimi Code CLI · Kimi K3 (default). Muse Code · Muse Spark 1.3 (xhigh) scores 4.7 points higher on DeepSWE v1.1. The composite lead is a Terminal-Bench 4 and SWE-Atlas-QnA lead that the DeepSWE v1.1 gap does not cancel.
| Component | Opus 5.5 score | Rank | Best other configuration | Gap |
|---|---|---|---|---|
| DeepSWE v1.1 | 68.4 | 6 of 20 | Muse Code · Muse Spark 1.3 (xhigh) at 73.2 | −4.7 points |
| Terminal-Bench 4 | 63.1 | 1 of 20 | Claude Code · Fable 5.1 (with fallback) (max) at 57.6 | +5.6 points |
| SWE-Atlas-QnA | 66.4 | 1 of 20 | Kimi Code CLI · Kimi K3 (default) at 66.1 | +0.3 points |
Every configuration with a lower AA Index also scores lower on Terminal-Bench 4, so a reader who cares only about terminal work orders the top of this chart the same way the composite does.
Opus 5 to Opus 5.5 at the same setting
The snapshot stores the previous Opus generation in the same harness at the same setting: Claude Code · Opus 5 (max) at 59.7 for $10.79 per task. Opus 5.5 adds 6.3 points at 1.2x the mean cost per task, 1.4x the total tokens per task, and 1.5x the mean time per task.
| Measure | Opus 5 | Opus 5.5 | Change |
|---|---|---|---|
| AA Index | 59.7 | 66.0 | +6.3 points |
| DeepSWE v1.1 | 62.5 | 68.4 | +5.9 points |
| Terminal-Bench 4 | 54.5 | 63.1 | +8.6 points |
| SWE-Atlas-QnA | 62.1 | 66.4 | +4.3 points |
| Mean API cost per task | $10.79 | $13.04 | 1.2x |
| Total tokens per task | 11.4 million | 15.6 million | 1.4x |
| Mean time per task | 42 minutes | 64 minutes | 1.5x |
Anthropic’s announcement prices Claude Opus 5.5 at $4 per million input tokens and $20 per million output tokens, against $5 and $25 for Opus 5, and opens with the claim that “It performs at the level of Claude Fable 5.1 on most work and costs 40% less to run than Opus 5.” The Claude Code row moved the other way: a task cost 1.2x as much because the run used 1.4x the tokens, and the volume increase outran the lower price per token. Anthropic’s figure describes the price of a token, and the harness figure describes the tokens one task consumed; both can be true at once.
The Intelligence Index row is a different measurement
The Artificial Analysis Intelligence Index runs the model through its API under one harness that is the same for every model, across 10 evaluations at version 4.3.2, and its cost per task is the average bill for one of those evaluation tasks. In the snapshot retrieved Sep 23, 2026, 1:38 PM UTC, Claude Opus 5.5 (Adaptive Reasoning, Max Effort, Default Fallback) scores 57.6 at $5.98 per task, first of 97 configurations with a measured cost. The Intelligence Index note walks that chart’s frontier and the model’s effort levels.
| Chart | Configuration | Score | Cost per task | Snapshot retrieved |
|---|---|---|---|---|
| Coding agents (AA Index) | Claude Code · Opus 5.5 (max) | 66.0 | $13.04 | Sep 25, 2026, 2:39 PM UTC |
| Intelligence Index | Claude Opus 5.5 (Adaptive Reasoning, Max Effort, Default Fallback) | 57.6 | $5.98 | Sep 23, 2026, 1:38 PM UTC |
The 66.0 is a mean of three coding benchmarks run inside Claude Code, and the 57.6 is a weighted average of 10 evaluations run through the API. The $13.04 includes every tool call and every repeated read of the repository that the harness sends, 15.6 million tokens per task in this snapshot; the $5.98 is the average bill for one evaluation task under Artificial Analysis’s standardized harness, with 119,166 output tokens per task. Artificial Analysis’s model page, captured September 25, 2026 UTC, lists the same list prices behind both figures, $4.00 per million input tokens and $20.00 per million output tokens with a 95% cache discount.
Anthropic’s own Terminal-Bench figure
Anthropic’s announcement reports its own Terminal-Bench 4.0 run at 66.4% at xhigh effort, alongside FrontierCode v1.1 (Main), CursorBench 4.0, GDPval-AA v2.1, AutomationBench, Humanity’s Last Exam, Terminal-Bench-Science 0.1, OSWorld 2.0, and Chartography, none of which appears on an AI Charts chart. The 63.1 that Artificial Analysis measured for Claude Code · Opus 5.5 (max) comes from a different run, a different effort setting, and Artificial Analysis’s protocol inside Claude Code, so the two figures describe two evaluations of the model rather than one result reported twice. Anthropic also writes that “at these levels of capability we’ve found that benchmark margins have become a less reliable guide to real-world differences.”
Limits
- Every score, cost, token count, and duration above is an Artificial Analysis measurement of the named configuration on the retrieval date, under DeepSWE v1.1, Terminal-Bench 4, and SWE-Atlas-QnA for the coding-agent chart and Intelligence Index version 4.3.2 for the capability chart. None of them establishes a result on other repositories, tasks, or harnesses.
- The rank, cost rank, frontier steps, component gaps, and Opus 5 multiples are AI Charts derivations from the snapshots named in each caption. A configuration added, removed, or rescored by Artificial Analysis moves them, and the coding-agent snapshot advances daily.
- Claude Opus 5.5 in Cursor, Devin, or any harness other than Claude Code, or at a setting other than max, is a configuration this snapshot does not store, so this note says nothing about it.
- The Opus 5 comparison holds the harness and setting fixed, but the snapshot records outcomes, not run dates or benchmark versions at run time; Artificial Analysis may have measured the two generations weeks apart.
- The prices, the cost claim against Opus 5, and the vendor-run benchmark figures belong to Anthropic. AI Charts did not run Claude Opus 5.5.