Claude Sonnet 5.5 Beats Opus 5.5 on Terminal-Bench. Should You Pay for Opus?

On Terminal-Bench 4.0, Anthropic reports 70.6% for Claude Sonnet 5.5 and 66.4% for Claude Opus 5.5 at Xhigh effort, a 4.2-point lead for the cheaper model. An independent run by Artificial Analysis agrees on the direction: 63.6% for Sonnet 5.5 and 59.6% for Opus 5.5, a 4.0-point gap. For terminal-style agentic work, the data does not support paying double per token for Opus. For other workloads it is less clear.

Does Sonnet 5.5 actually beat Opus 5.5 on Terminal-Bench 4.0?

Yes, in both published data sources, by a similar margin. The table below keeps lab-reported and independent numbers apart, because they come from different harnesses and should not be averaged. All figures are Terminal-Bench 4.0, captured as of 13 October 2026.

Two things stand out. The ordering is stable across sources. The absolute level is not: the independent harness scores both models about 7 points below Anthropic's figures, which is consistent with harness sensitivity.

What does Terminal-Bench 4.0 measure, and how is it scored?

Agents work through tasks in a terminal, and per Artificial Analysis, a task passes only if every test passes. Artificial Analysis runs all 66 tasks with the mini-swe-agent harness and reports pass@1 averaged over three repeats per task. It shows no confidence intervals.

The Terminal-Bench 4.0 release post says the version removed eight tasks (two each for saturation, refusals, public solutions and quality or platform-compatibility issues) and fixed 19. It also set a flat 8-hour agent timeout on all tasks. The maintainers call it 4.0 rather than 3.1 because the resource and task-set changes require re-running trials.

Who runs it, and are there conflicts of interest?

The benchmark is hosted by Stanford, Harbor and the Laude Institute. The release post says leaderboard experiments are supported by grants from OpenAI, Anthropic, Z.ai, SpaceX AI and the Laude Institute. That does not make the results unreliable, but a grant from a vendor whose model is on the board is worth knowing.

The two score sources differ in independence. Anthropic's figures are lab-reported and the launch page gives no agent, harness or timeout detail for Terminal-Bench, pointing readers to its system card for evaluation setup. Artificial Analysis states that its evaluations are conducted independently.

Where does Opus 5.5 still win?

On every one of the other launch-table rows listed below. All of these are lab-reported, on Anthropic's launch page, and effort levels are mostly unstated.

The largest Opus advantage is FrontierCode, at 8.2 points against Sonnet's Max-effort score. Anthropic's footnote says Sonnet 5.5 scores lower at Max than at Xhigh there, because Claude Code's code-review skill split work across many subagents and, in two cases Cognition examined, caused a timeout or out-of-scope edits. That quirk is reported for FrontierCode, not Terminal-Bench. On CursorBench, Anthropic says Sonnet's best score is within about two points of Opus 5.5. GDPval-AA is effectively a tie.

How much does each model cost?

Per token, Sonnet 5.5 is half the price of Opus 5.5. Prices below are as of 13 October 2026.

Source: Anthropic and the Opus 5.5 docs, which also list a 1M-token context window, 128K max output and a 50% batch discount. Cost per task is the number that matters for agents, and it depends on tokens spent, not just the rate card. Anthropic presents accuracy-versus-cost curves for Terminal-Bench 4.0 across Low to Max effort as charts, and says Sonnet 5.5 at Medium effort far exceeds Sonnet 5's best score for under a tenth of the cost per task. The comparison between Sonnet and Opus per task is not given as a table, so a precise ratio cannot be stated from the published text.

The release post for the benchmark offers a reminder that token counts vary widely: on its leaderboard run, Sonnet 5 used 21.6B tokens against 6.5B for Opus 5, and hit timeouts and output-token errors. That is a different model pair, but it shows why the per-token price is only half the cost picture.

Should you pay for Opus 5.5?

For terminal and shell-heavy agent work, start with Sonnet 5.5 and measure. It leads on Terminal-Bench 4.0 in both sources and costs half as much per token. Move to Opus 5.5 for workloads that resemble the rows where it leads, such as FrontierCode-style merge-ready code changes, where Anthropic's own table shows an 8.2-point gap against Sonnet's Max-effort score.

A practical test is to replay 50 to 100 of your own tasks on both models at the effort level you would actually deploy, and record pass rate and tokens per task. Defaults differ: Anthropic says Medium is the default in the Claude apps and Claude Code, and High on the Claude Platform.

What's still unclear?

Several points limit how far the headline can be pushed.

• Sonnet's effort setting. Anthropic states Opus's Terminal-Bench result is at Xhigh but does not label the Sonnet 5.5 figure, so the 70.6% cannot be tied to a specific effort or cost point.
• No confidence intervals. Neither source publishes them. A 4-point gap on 66 tasks is plausible signal, but a single task is about 1.5 points of the score, so rankings within a few points can flip.
• Harness gap. The independent scores sit well below the lab-reported ones, and Anthropic's harness is not described on the launch page.
• Self-reported ceiling. Anthropic describes the Opus figure as the model's highest score, which means the comparison is Sonnet at an unspecified setting against Opus at its best.
• Weakest point of the argument. Terminal-Bench is one benchmark. Opus 5.5 leads Sonnet 5.5 on all seven other launch-table rows listed above, if the near-tie on GDPval-AA is counted as a lead.

This post analyses lab-reported and third-party figures only. We did not run the benchmark ourselves.