aicharts
In the aicharts coding-agent snapshot retrieved Oct 5, 2026, 2:10 PM UTC, Claude Code · Opus 5.5 (max) scores 66.0 on AA Index at a mean API cost of $13.04 per task, second of the 31 configurations that carry an index. The nearest cost-frontier configuration below it, Antigravity CLI · Gemini 4 Argon (default), gives up 2.2 points for 45% 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.
Second of 31 configurations
The 66.0 places Claude Code · Opus 5.5 (max) second of the 31 configurations that carry an AA Index in the snapshot retrieved Oct 5, 2026, 2:10 PM UTC. One configuration scores higher: Claude Code · Sonnet 5.5 (max) at 68.4. The leader, Claude Code · Sonnet 5.5 (max), is 2.4 points above it at $14.19 per task. Its $13.04 per task is the second highest cost of the 31 configurations that carry a cost.
The nearest score below it is Antigravity CLI · Gemini 4 Argon (default) at 63.8, 2.2 points lower for 45% 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 |
|---|---|---|---|---|---|
| Antigravity CLI · Gemini 4 Argon | default | 63.8 | $5.84 | 2.2 points | 45% |
| Codex · GPT-6.1 Sol | xhigh | 62.9 | $1.04 | 3.1 points | 8.0% |
| Claude Code · Sonnet 5.5 | xhigh | 62.9 | $3.33 | 3.1 points | 26% |
| Claude Code · Fable 5.1 (with fallback) | max | 62.2 | $12.39 | 3.8 points | 95% |
What stepping down the cost frontier gives up
The chart’s cost frontier contains configurations for which no other row costs no more and scores at least as high, with a strict improvement in either measure. Claude Code · Opus 5.5 (max) is on it. 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 Antigravity CLI · Gemini 4 Argon (default): 2.2 points lower for 45% of the $13.04. That first step already halves the cost or better. 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 |
|---|---|---|---|---|---|
| Antigravity CLI · Gemini 4 Argon | default | 63.8 | $5.84 | 2.2 points | 45% |
| Codex · GPT-6.1 Sol | xhigh | 62.9 | $1.04 | 3.1 points | 8.0% |
| Codex · GPT-6.1 Sol | medium | 61.4 | $0.705 | 4.6 points | 5.4% |
| Codex · GPT-6.1 Sol | low | 57.2 | $0.499 | 8.8 points | 3.8% |
| 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% |
Where the index points come from
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, 12th of 31; on Terminal-Bench 4 it scores 63.1, second of 31; and on SWE-Atlas-QnA it scores 66.4, second of 31 configurations that carry each score.
It leads none of the components outright; its composite position comes from placing high on all of them at once. Antigravity CLI · Gemini 4 Argon (default) scores 10.3 points higher on DeepSWE v1.1, Claude Code · Sonnet 5.5 (max) scores 3.0 points higher on Terminal-Bench 4, and Claude Code · Sonnet 5.5 (max) scores 0.5 points higher on SWE-Atlas-QnA.
| Component | Opus 5.5 score | Rank | Best other configuration | Gap |
|---|---|---|---|---|
| DeepSWE v1.1 | 68.4 | 12 of 31 | Antigravity CLI · Gemini 4 Argon (default) at 78.8 | −10.3 points |
| Terminal-Bench 4 | 63.1 | 2 of 31 | Claude Code · Sonnet 5.5 (max) at 66.2 | −3.0 points |
| SWE-Atlas-QnA | 66.4 | 2 of 31 | Claude Code · Sonnet 5.5 (max) at 66.9 | −0.5 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.” In this Claude Code comparison, mean task cost rose to 1.2x as much and total tokens to 1.4x as many. Task cost depends on token volume, the input/output mix, and caching as well as listed rates. Anthropic’s 40% figure is its running-cost claim; the quoted input and output prices alone fell by 20%. The separate workloads do not isolate why the measured task costs differ.
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 Oct 10, 2026, 5:51 AM UTC, Claude Opus 5.5 (Max, Default Fallback) scores 57.6 at $5.98 per task, first of 108 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 | Oct 5, 2026, 2:10 PM UTC |
| Intelligence Index | Claude Opus 5.5 (Max, Default Fallback) | 57.6 | $5.98 | Oct 10, 2026, 5:51 AM 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 aicharts 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
- The chart scores, task costs, token counts, and durations are Artificial Analysis measurements 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 aicharts 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. aicharts did not run Claude Opus 5.5.