Claude Opus 5.5 Tops the Coding Agent Index at 66 — and the Task Bill Rises 21%

The vendor cut Claude Opus 5.5's token prices by 20% on September 22, 2026. Two days later, Artificial Analysis published the coding-agent measurement that shows what the cut did to an agent's bill: cost per task went up 21%, to $13.04, from the $10.79 the same index charges for Claude Opus 5. The score went up too — 66 on the Coding Agent Index, which Artificial Analysis describes as the highest it has measured, six points clear of Claude Opus 5 and four clear of Claude Fable 5.1 in the same configuration. Both of those things are true at once, and the reason they are true at once is the whole story of this ranking. A cheaper token that a model spends more of is not a cheaper task, and at maximum reasoning effort on long agent runs, Claude Opus 5.5 spends a lot more of them.

What the Coding Agent Index actually scores

This is not a model leaderboard. Every row on it is a harness-and-model pair — a checkpoint running inside a specific coding agent, at a specific effort setting. The row everyone is quoting is Claude Opus 5.5 at max effort inside Claude Code, which is the harness Anthropic builds and tunes against. Claude Opus 5's 60 and Claude Fable 5.1's 62 come from the same harness and the same effort setting, which is what makes them the honest comparisons; a row for either model in a different coding agent is a different measurement and is not printed here as if it were the same one.

The index is versioned, and the version matters more than the brand name on it. The board currently published as v1.5 averages three evaluations, each weighted equally, each reported as pass@1 averaged over three attempts per task:

• Terminal-Bench 4.0 — agentic shell and command-line work: navigating a filesystem, using tools correctly, recovering from an intermediate failure, and finishing a multi-step workflow.

• DeepSWE v1.1 — software-engineering tasks, the repository-editing work.

• SWE-Atlas-QnA — repository-comprehension questions, where the answer is found by reading a codebase rather than changing it.

Three evaluations, one number, and a cost-per-task column printed next to it. That cost column is why this index is more useful than a score alone, and it is also why the headline number is harder to read than it looks.

The three components, and where the six points came from

Artificial Analysis published the component rows alongside the composite, and they do not move evenly:

• Terminal-Bench 4.0 — 63.1% for Claude Opus 5.5 against 54.5% for Claude Opus 5, a gain of 8.6 points. This is the largest single improvement in the set.

• DeepSWE v1.1 — 68.4% against 62.5%, up 5.9 points.

• SWE-Atlas-QnA — 66.4% against 62.1%, up 4.3 points.

Average those three and you get 66.0, which is the composite. The shape of the gain is the interesting part: the model improved least at reading a codebase and most at driving a shell. Terminal-Bench is the eval closest to what people actually mean by "coding agent" — a long-running loop where the model chooses commands, reads the output, and decides what to do next, with no human in the loop between steps. An 8.6-point jump there is a statement about sustained tool use, not about code generation.

Note what that implies about where to expect the improvement. If your workload is "explain this repository," Claude Opus 5.5 is a modest upgrade over Claude Opus 5 — four points. If your workload is "run until the test suite passes, fixing what breaks," it is the largest jump in the table.

The number printed next to the ranking: $13.04 a task

Here is the arithmetic that makes the price cut look different from the way it was announced. Anthropic's list price for Claude Opus 5.5 is $4.00 per million input tokens and $20.00 per million output, down from $5.00 and $25.00 for Claude Opus 5, with cache reads down 60% to $0.20 per million. Every one of those lines is cheaper. Artificial Analysis' estimate of what a task costs at max effort is higher anyway:

• Cost per task — $13.04 for Claude Opus 5.5 against $10.79 for Claude Opus 5, up 21% on the same index, same harness, same effort setting.

• Total tokens per task — roughly 15.6M against 11.4M, up about 37%.

• Output tokens per task — roughly 333,000 against 137,000, more than double. Output is the expensive direction, and it is the line that moved most.

• Cached input per task — roughly 14.6M against 10.9M, up about 34%. The 60% cache-read cut is doing real work here; it just is not enough to offset a task that reads a third more context.

Do the sum with the published rates and the shape falls out. Take the output tokens alone: 333,000 tokens at $20.00 per million is about $6.66 — roughly half of the entire $13.04 estimate, from a single line item that doubled. The cache-read line is enormous in volume and almost free in price: 14.6M tokens at $0.20 per million is under $3. The 20% cut is real and the cache cut is real; at max effort on an agent loop, they are simply outweighed by a model that thinks longer and writes more.

That has a direct consequence for how you budget. If your team runs 500 coding-agent tasks a day at max effort, the difference between $10.79 and $13.04 is about $1,125 a day — roughly $34,000 a month — for the same number of tasks. That is the number to put in front of whoever approves the model change, and it is not the number the price cut implies.

Why $5.98 and $13.04 are both true

Artificial Analysis published a different cost-per-task figure for the same model days earlier, and reading the two together without the labels makes them look like a contradiction. They are not — they are two different evaluations, and the difference is instructive.

On the Intelligence Index, the write-up on September 22 put Claude Opus 5.5's cost per task at $5.98, roughly level with Claude Opus 5's $5.86, despite about 119,000 output tokens per task against Opus 5's 73,000. There, the price cut and the extra tokens cancel out almost exactly. On the Coding Agent Index, at max effort, in Claude Code, the same model costs $13.04 against $10.79.

Both are honest measurements of different workloads. A question-answering eval is a short exchange; an agent task is a long loop of tool calls with a growing context, and the growth compounds in the output direction, where the price is highest. The practical lesson is that "does the price cut offset the extra tokens?" has no single answer — it depends on the task length, and the longer and more agentic the task, the worse the offset gets. If you are budgeting from a short-answer benchmark, you will underestimate an agent's bill.

Three caveats that belong on the row

The 66 is a Claude Code number, not a bare-model number. The index pairs models with harnesses deliberately, on the argument that tool routing, retry logic and context management are inseparable from an agent's measured performance. That cuts in Anthropic's favour here, because the harness it is scored in is its own. Artificial Analysis marks a harness-comparison view as coming soon; until it exists, no one can say how much of the six-point lead is the checkpoint and how much is the scaffolding around it.

Anthropic's own Terminal-Bench 4.0 figure is higher than this component, and was run by Anthropic. The launch material reports 66.4% at xhigh effort; the Coding Agent Index component is 63.1% at max, and Artificial Analysis' separate independent run measured 59.6% in its own harness on the same benchmark version. Three numbers, three configurations, three producers. The 63.1% on the row above is the index's own component and should only be compared to the other numbers on that index; the vendor's 66.4% is vendor-reported, unreproduced by a third party, and belongs to a different effort setting.

The index revision moves. This blog reported different Coding Agent Index rows in early September, on an earlier version of the same board, and those older figures are not comparable to v1.5 — the evaluation set itself changed when Terminal-Bench 4.0 replaced the earlier version. Third-party mirrors still carry stale snapshots with older version labels. If a number for Claude Opus 5.5 on this index is not from the current v1.5 row, do not put it next to a v1.5 number, and do not compute a delta across the two.

What to actually deploy

The ranking is a max-effort number, and max effort is the setting almost nobody should ship by default. Claude Opus 5.5 has five effort settings — low, medium, high, xhigh and max — and the model defaults to medium. Artificial Analysis has not published Coding Agent Index rows at the lower settings, so the cost saving there is unquantified on this index; on the Intelligence Index, the same model at medium scored 51 — matching Claude Opus 5's max — at $1.34 per task. That is the direction the savings live in, and it is a setting, not a different model.

The realistic pattern is a split: keep max effort for the long autonomous loops where the 8.6-point Terminal-Bench gain is the thing you are buying, and run the routine work at a lower effort where the model is still well ahead of where Claude Opus 5 finished. Because effort is a request parameter rather than a deployment, that split is a routing rule, and both models sit on the same catalogue — Anthropic's list prices passed through at 0% markup, so a vendor price cut is live on anthropic/claude-opus-5.5 the same day it is announced, and there is no second contract or code change involved in A/B-ing the two against your own tasks. For the max-effort path specifically, where a single task can be a long unattended run, automatic failover is what keeps a stalled provider from costing you the whole loop.

What is still unmeasured

Three things would change this picture and none of them exists yet. There is no matched-effort, matched-harness comparison of Claude Opus 5.5 against a rival checkpoint — every cross-vendor comparison available is two vendor tables arranged side by side. There is no published Coding Agent Index run at medium or high effort, which is where most production traffic would sit. And the harness-attribution question — how much of a 66 is the model and how much is Claude Code — has been acknowledged by the index and deferred.

What you can act on today is narrower and more useful than a ranking. Claude Opus 5.5 is genuinely better at sustained shell-driven agent work than Claude Opus 5 — that is the one component with a large, consistent, independently produced gain. It also costs more per task at the setting where that gain was measured, by about $2.25 a task, and the per-token price cut does not reach that number. Deploy it at max effort where the loop is long and the failure cost is high; keep the cheaper setting everywhere else; and re-check the index when the lower-effort rows appear, because that is the configuration most teams will actually pay for. If you are switching for the price cut alone, you are buying the wrong thing — the model is cheaper per token and more expensive per task, and only one of those two numbers shows up in your invoice.