OpenRouter dates the post October 30, 2026. The structured time is 00:00:00 UTC that day, midnight on the hour, so it is a poor event clock. The calendar date itself falls after the backfill line. The post says the prices are list prices from October 21, and that the benchmark rows were pulled on October 21 from runs dated October 17. These figures were not measured on October 30.
[1]
The three models are one generation. The post says Opus 5.5 was released on September 22, 2026, Sonnet 5.5 on September 28, and Haiku 5.5 on October 7. All three have a 1-million-token context window and accept text, images, and PDFs. They differ mostly in price, and Haiku also scores lower on benchmarks. List prices per million tokens are 4.00 dollars of input for Opus, 2.00 for Sonnet, and 0.10 for Haiku, with output at 20.00, 10.00, and 0.50 dollars.
[1]The recommendation in the post is to use Sonnet 5.5 as the default, Haiku 5.5 for high-volume work and sub-agents, and Opus 5.5 when a task is long, ambiguous, or expensive to get wrong. That is a judgment on the tasks in this post, not a general measurement.
[1]Each model received the same 11 coding tasks with hidden tests, and the same bug-fix task in Claude Code. The post says every model passed everything, so cost is what separated them. Per task, Haiku was 19 to 30 times cheaper than Opus, and Sonnet cost between half and two-thirds as much as Opus.
[1]On the benchmarks OpenRouter runs, Sonnet does not sit below Opus. On this run of GPQA Diamond, Sonnet scored 92.4 percent, Opus 90.4 percent, and Haiku 86.6 percent. On this run of the airline slice of τ-bench, Sonnet scored 78.7 percent, Opus 75.5 percent, and Haiku 76.0 percent. The Artificial Analysis index is 57.6 for Opus, 56.0 for Sonnet, and 43.4 for Haiku, so Opus leads by 1.6 points. Cost per GPQA question was 0.025 dollars for Sonnet and 0.054 dollars for Opus. Cost per τ-bench task was 0.16 dollars for Sonnet and 0.34 dollars for Opus. The post says these runs had 100 to 200 tasks per model, so gaps of a few points may be noise. Haiku trails Sonnet by 5.8 points on GPQA and by 12.6 points on the index. The index weights agentic tasks and general knowledge at 30 percent each, and coding and scientific reasoning at 20 percent each.
[1]The coding test has two parts. Of the 11 Python tasks, eight were everyday tasks and three were harder. Each everyday task was attempted twice and each hard task three times, 25 attempts in all. All three tiers passed 25 of 25. Cost per everyday task was 0.0211 dollars for Opus, 0.0112 for Sonnet, and 0.0007 for Haiku. Cost per hard task was 0.0739, 0.0380, and 0.0027 dollars, with times of 32.1, 25.7, and 22.6 seconds. Haiku used 3,202 reasoning tokens on average against 1,252 for Opus, and still finished the hard tasks fastest, at about a twenty-seventh of Opus's cost.
[1]The bug-fix gave Claude Code a small invoicing library with three bugs and five tests, three runs per model. All three tiers fixed all three bugs on 3 of 3 runs. Median session cost was 0.098 dollars for Opus, 0.062 for Sonnet, and 0.0052 for Haiku. Claude Code's own cost field did not recognize the Haiku model name, and the figure it reported was 30 to 60 times the 0.0052 dollars computed from list prices.
[1]The author says the tasks are small and well specified, and do not cover long-running agentic coding. A perfect score on the short tasks does not show that the long tasks have no gap. In the week of October 12 to 18, Opus carried about two and a half times Sonnet's token volume, and Opus is what Claude Code uses on most plans when no model is set.
[1]