These are dated, source-specific candidates, not current leaderboard ranks or universal winners. Check each benchmark's capture date, harness and effort, then verify the linked current model profile before choosing.
Last verified:Snapshot of catalog vendor-page checks: oldest October 1, 2026, newest October 2, 2026.
Coding-agent benchmark candidate
Candidate: Claude Opus 5.5
The public CursorBench 4.0 table retrieved 2026-10-01 lists Opus 5.5 Max at 57.8%, Extra High/High at 56.0%, and Medium at 52.5%; Anthropic's model overview recommends starting with Opus 5.5 for many workloads. These are a harness-specific coding snapshot and a vendor recommendation, not a general coding crown.
Caveat: CursorBench tests Cursor agent configurations, and model effort changes both score and cost. Its current table does not establish a best coding model across other harnesses; evaluate current model IDs on your repository. AA component data are not fully public, so no current AA-index crown is asserted.
OpenAI's current API profile lists GPT-6 Luna at $0.10/$0.50 per million input/output tokens with a 1.05M context. This identifies a budget-tier option; no benchmark-based claim that it is the strongest or cheapest across all workloads is made.
Caveat: Compare current alternatives using exact model IDs, prompt mix, caching, effort and task success. API rates do not establish ChatGPT plan availability or cost per accepted task.
DeepSeek's current pricing page lists V4.1-Flash at $0.30/$1.20 per million input/output tokens at peak and $0.15/$0.60 off-peak; its model card identifies the published weights as MIT-licensed. This is a rate/availability candidate, not a quality-per-dollar benchmark result.
Caveat: Peak schedules affect API bills, while self-hosting adds hardware and operations. Compare quality, modalities, host, cache use and current pricing on your own workload; no cross-board rank is asserted.
Google lists Gemini 3.5 Flash-Lite as an available Gemini API model and describes it for low-latency, high-throughput workloads including agentic search and document processing. This supports a workload-fit candidate, not a cheapest-model or leaderboard crown.
Caveat: Check the current rate card, model endpoint, limits and modality support before choosing it. Google lists other Flash/Flash-Lite rows, so this entry does not claim it is the cheapest Google model or best multimodal model.
Xiaomi's published MiMo-V2.6-Pro-RL checkpoint is MIT-licensed and has a 1M context in the model card. This makes it an open-weight evaluation candidate; its vendor evaluation results are not independent reproductions.
Caveat: A checkpoint's open licence does not make its API equivalent or local deployment inexpensive. Confirm memory, runtime support, model limits and hosted rates for the exact target; no current open-model leaderboard rank is asserted.
The best headline score can still lose on total cost.
Compare successful outputs, retries, latency, and human review alongside token price. A lower-priced model that needs twice the supervision is not automatically the better value.
