Claude Sonnet 5.5 matches Claude Opus 5.5 on Vals AI's index at two thirds of the cost per task and fixes more planted bugs than any model tested
The Aggregate Digest — Sunday, October 4, 2026
Claude Sonnet 5.5 matches Claude Opus 5.5 on Vals AI's index at two thirds of the cost per task and fixes more planted bugs than any model tested. Claude Sonnet 5.5 has been on the table since Tuesday and is #10 at 1967, sixty points below Claude Opus 5.5 at #4. The Vals Index averages Vals AI's agent benchmarks by sector and weights each sector by its share of US GDP. Sonnet 5.5 scores 67.04 percent and Opus 5.5 scores 66.97. They are second and third of 43 behind Gemini 4 Argon at 68.9. Vals AI's cost table puts one index task at $21.34 with Sonnet 5.5 and $32.14 with Opus 5.5, with Gemini 4 Argon at $15.68.
Vals AI Vals Index cost table, October 4 snapshot. The eight highest Vals Index scores of 43 models, in score order; the bar is the dollar cost of one index task. The index averages Vals AI's finance, coding and legal agent benchmarks weighted by sector share of US GDP.
Bug repair and code generation go to the cheaper model. Bug Hunt Bench plants 105 defects in a VS Code extension and a learning-management system and counts the verified fixes. Sonnet 5.5 at max effort fixes 51.3 of the 105 and is first of 101 entries. GPT-6 Astra is second with 45.0 and Opus 5.5 eighth with 41.7, both at max effort. LiveBench's code generation task asks for a complete program written against real-world libraries. Sonnet 5.5 at max effort scores 95.78 percent there and is first of 66. Opus 5.5 scores 91.55. LLM Stats' listing of Terminal-Bench 4.0 puts Sonnet 5.5 first at 70.6 to Opus 5.5's 66.4. Vals AI's run of the same benchmark puts Opus 5.5 first at 65.15 to Sonnet 5.5's 64.14. The benchmark's own board has tested neither model.
Closed-book knowledge is where Opus 5.5 keeps its distance. Opus 5.5 is ahead on 105 of the 129 boards that score both models, and the gap is widest on recall. HLE-Diamond is 1,000 cleaned Humanity's Last Exam questions answered without tools. Sonnet 5.5 at max effort scores 43.3 percent there and is ninth of 24. Opus 5.5 scores 60.9. SimpleQA Verified asks a thousand short factual questions from memory and the two score 46.5 and 72.2. Vals AI's CyberBench runs agentic security tasks and goes 59.58 to 74.58.
October 4 results from LiveBench, LLM Stats, Vals AI, Mercor, lastexam.ai and Epoch AI. Each row compares the same task and metric across the two models at the same effort where the board records one. Different tasks, not a combined score; the two Terminal-Bench 4.0 rows are separate evaluators' runs of one benchmark.
Gemini 4 Argon sits one point above Claude Opus 5.5 on the table at 2028 to 2027. Its standard error is 18 points on 134 results. Claude Opus 5.5 keeps the best available model crown at 2027 because a crown moves only after a five-point lead holds for three run dates. This week Argon went first of 413 on Chatbot Arena's text board. Claude Opus 5.5 at high effort is tied for third there. Argon is also first of 50 on Mercor APEX, which runs long-horizon professional-services agent tasks. It scores 82.2 percent to Sonnet 5.5's 75.5 and Opus 5.5's 73.5.
Elsewhere. ClockBench has a model read analog clocks, 720 questions from telling the time to time-zone conversion. Human rows joined the board this week: an average human scores 90.7 percent and the top human 95.0, while the best model, Claude Opus 5.5 at max effort, reads 77.2. KaliBench is new. From an analyst's plain-language request the model writes the exact Kali Linux command, and its 5,000 queries are scored on an exact match. GPT-5.6 Sol leads 24 entries at 61.68 percent. GPT-5.5 under Codex at xhigh effort has 51.68 and Claude Opus 5 has 44.02. Neither Opus 5.5 nor GPT-6 Astra was tested, and neither was Gemini 4 Argon.
