Add August 2026 model benchmarks

[atharvadeosthale]

Summary

• Adds five new models to the Arena benchmark: Gemini 3.7 Flash, Muse Spark 1.2, Grok 4.6, Kimi K3, and Claude Opus 5, with full results for both with-skills and without-skills modes.
• Adds --models <ids> / --list-models flags to the benchmark runner so a subset of models can be run instead of the whole roster (useful when a model gets stuck).
• Adds the Meta provider logo to ProviderLogo.

Results (overall)

Model	Without skills	With skills
Claude Opus 5	97.4%	97.4%
Kimi K3	95.0%	96.9%
Muse Spark 1.2	94.8%	97.8%
Grok 4.6	91.6%	97.8%
Gemini 3.7 Flash	93.6%	96.0%

Note: stealth/ox-alpha was also benchmarked but excluded — it does not support tool calling on OpenRouter, so it failed every MCQ (answered via tool call) and produced no usable with-skills results.

[greptile-apps]

Greptile Summary

The PR adds five August 2026 benchmark models and their result datasets, introduces model filtering and listing options for the benchmark runner, and adds Meta provider branding.

• Adds Gemini 3.7 Flash, Muse Spark 1.2, Grok 4.6, Kimi K3, and Claude Opus 5 configurations.
• Publishes with-skills and without-skills benchmark results for the new models.
• Adds --models and --list-models runner options, including validation and per-model timestamp preservation.
• Adds Meta monochrome and color provider icons.

Confidence Score: 4/5

The PR should not merge until filtered runs stop reporting already-complete selected models as newly evaluated.

The runner assigns a fresh timestamp before checking whether a selected model has any unfinished questions, causing unchanged results to receive a false public “Last updated” date.

Files Needing Attention: benchmark/src/index.ts

Important Files Changed

Filename	Overview
benchmark/src/index.ts	Adds model-selection CLI support and preserves existing timestamps, but still refreshes selected models that perform no benchmark work.
benchmark/src/config.ts	Adds configuration and provider metadata for five new benchmark models.
src/components/ProviderLogo.tsx	Adds Meta monochrome and color logo mappings.
src/data/results-with-skills.json	Adds with-skills results for the five new models and updates aggregate benchmark metadata.
src/data/results-without-skills.json	Adds without-skills results for the five new models and updates aggregate benchmark metadata.

Prompt To Fix All With AI

### Issue 1
benchmark/src/index.ts:459
**Skipped models receive fresh timestamps**

When a filtered run selects a model whose existing results are already complete, this assignment refreshes its `runDate` before `runBenchmark` skips every question, causing unchanged results to be displayed as newly evaluated on the model detail page.

---

For each issue above, determine whether it is valid and should be fixed. If so, fix it directly.

_{Reviews (2): Last reviewed commit: "Address review: preserve runDate for uns..." | Re-trigger Greptile}

[greptile-apps]

Skipped models receive fresh timestamps

When a filtered run selects a model whose existing results are already complete, this assignment refreshes its runDate before runBenchmark skips every question, causing unchanged results to be displayed as newly evaluated on the model detail page.

Knowledge Base Used:

• Benchmark orchestration
• Leaderboard results and ranking

Prompt To Fix With AI

This is a comment left during a code review.
Path: benchmark/src/index.ts
Line: 459

Comment:
**Skipped models receive fresh timestamps**

When a filtered run selects a model whose existing results are already complete, this assignment refreshes its `runDate` before `runBenchmark` skips every question, causing unchanged results to be displayed as newly evaluated on the model detail page.

**Knowledge Base Used:**
- [Benchmark orchestration](https://app.greptile.com/appwrite/-/custom-context/knowledge-base/appwrite/arena/-/docs/benchmark-orchestration.md)
- [Leaderboard results and ranking](https://app.greptile.com/appwrite/-/custom-context/knowledge-base/appwrite/arena/-/docs/leaderboard-results-and-ranking.md)

---

For each issue above, determine whether it is valid and should be fixed. If so, fix it directly.