Opus 5.5 aced a test suite at 3.4x GPT-6 Sol’s cost

Claude Opus 5.5 and GPT-6 Sol both solved all 49 tasks of the same deterministic suite with zero errors, and Opus 5.5 cost 3.4 times as much per correct answer: $0.05422 against $0.015922 for Sol. The run comes from an Aipricing cost analysis of the two models, which both launched on September 22. Its authors argue that the comparison that counts is cost per accepted task, and that price per token is only the starting point.

The higher rate card was only part of why Opus lost this round. Opus 5.5 also used more tokens: 5,330 prompt and 1,645 completion tokens, against 3,426 and 907 for Sol. On these short machine-graded tasks, the authors write, Opus did not win back its rate premium through efficiency. They also say the run establishes nothing about repository-scale coding.

Twice the rate, same cache reads

For standard short-context API work, GPT-6 Sol costs $2 per 1M input tokens, $0.20 for cached input, $2.50 for cache writes and $10 for output. Claude Opus 5.5 costs $4 input, $0.20 for cache reads, $5 for five-minute cache writes and $20 output, plus $8 per 1M for one-hour cache writes. At equal token volumes that puts Opus at twice Sol’s rate for fresh input, cache writes and output, while cache reads cost the same. Context length changes the picture somewhat: Sol charges a higher rate once a request exceeds 272,000 input tokens, while Opus publishes one standard rate across its 1M context. DevX also lists a fast mode for Opus 5.5 at $8 input and $40 output per 1M tokens, with up to 2.5 times the speed.

At equal volume, Sol wins. For Opus to justify its premium, it has to cut the amount of work a task takes. Aipricing’s working formula for an agent bill adds model tokens, cache writes, tool fees, retries, failed runs and reviewer time. If Sol needs ten calls where Opus needs five, the 2x rate gap can disappear. If both finish in the same number of calls, Sol usually keeps its advantage. And if the cheaper run produces a patch that fails tests or needs an engineer to fix it, the token saving may not matter at all.

Where Anthropic’s 40% comes from

Anthropic prices Opus 5.5 at $4 input and $20 output, 20% below Opus 5’s $5 and $25. Cache reads fall 60%, from $0.50 to $0.20 per 1M, and Anthropic says cache reads make up the majority of agentic and coding costs. The company says the model needs less compute to serve than Opus 5 and also uses fewer tokens per task, so a typical workload at default settings comes out about 40% cheaper. That default is medium effort, according to the platform docs. Aipricing treats the 40% as a provider estimate. It is not a discount every invoice will show.

Aipricing’s own suite came close to that estimate. Opus 5.5 scored 49/49 for $0.05422, while Opus 5 scored 48/49 for $0.09366, which works out to 43% lower cost per correct answer. One early tester, Rogo, said it matched Opus 5’s output quality at a 60% lower token cost during testing, DevX reported.

Effort settings can move token counts a long way. In September, Artificial Analysis ran Opus 5.5 at max effort through its index, and the model generated 260M output tokens against a median of 92M for comparable models.

What does the launch table actually show?

In its launch table, Anthropic reports Opus 5.5 at 66.4% on Terminal-Bench 4.0 and 54.4% on FrontierCode, against 37.3% and 47.5% for GPT-6 Sol. On CursorBench the table gives 57.8% for Opus 5.5 and 41.7% for Sol. The same table puts Opus 5 at 52.3% on Terminal-Bench 4.0, 48.0% on FrontierCode v1.1 and 46.6% on CursorBench 4.0, and DevX notes that none of these numbers have been independently audited.

Aipricing calls the figures useful for building a shortlist. They are not a neutral head-to-head test: Anthropic publishes the table, effort settings differ from one result to the next, and each vendor uses its own harness and cost assumptions. The same caution applies to Anthropic’s claim that Opus 5.5 at default effort beats GPT-6 Astra at about one-fifth of the cost. That claim covers selected FrontierCode and GDPval comparisons. At equal token volumes, Opus 5.5’s standard input and output prices sit 60% below Astra’s short-context rates. Any saving beyond that has to come from the task trace: fewer tokens, calls, retries or failures.

Subscription routes

On launch day the OpenClaw build Aipricing checked did not list Opus 5.5 in its Anthropic model catalog. OpenClaw can run Claude through the official CLI with claude -p, and its documentation says that route draws on the signed-in subscription’s Agent SDK limits. Because the catalog entry is missing and the model changes how requests must be formed, Aipricing advises waiting for a tested compatibility path or validating a deliberate override.

Hermes has a separate experimental Claude Subscription DirectSDK plugin. Its September 22 catalog lists claude-opus-5-5[1m], and the plugin needs Hermes 0.21.4 or newer plus a logged-in Claude Code CLI. It routes Hermes turns through the subscription. In the plugin’s own test, usage was metered the same way as claude -p, at about 1.7 times the interactive Claude Code TUI meter. On identical coding tasks, Hermes sent about 0.6 times as many tokens as native Claude Code. The two effects can offset each other, and only a paired workload test settles the bill. This plugin is separate from Hermes’ direct Anthropic OAuth path, which Hermes documents with different plan and extra-usage behavior, so it is worth confirming which provider is selected before counting usage as included in a plan.

Breaking changes that cost loops

Opus 5.5 always uses adaptive thinking and rejects requests that try to turn it off. Thinking depth is set with the effort parameter. Forced tool choice is unsupported. Thinking blocks are tied to the model and the conversation, and Anthropic’s API and Google Cloud reject one older version of the computer-use tool. Progress text that used to appear between tool calls now moves into thinking blocks. Aipricing warns that a harness that mishandles any of these can retry, lose reasoning state, show no progress or fail a tool round, and every extra loop raises the effective cost per task.

Its suggested pilot keeps the task set, repository commit, tool list, permissions, cache state, effort and stop policy identical for both models. For each run it records successful tasks, total input, cache reads and writes, output, tool calls, wall time, retries, failures and minutes of human review.

Aipricing’s recommendation splits by workload. Sol fits where equal-volume price, predictable API economics and high daily throughput matter most. Opus 5.5 fits difficult coding or knowledge work that gets expensive mainly because weaker runs wander, retry or need heavy review, and only when your own traces show a better completion rate or fewer steps. The authors suggest keeping GPT-6 Luna or Claude Sonnet 5 in the pilot as cheaper controls: “A premium model should beat them on cost per accepted result, not merely on a launch benchmark.”