Every week our model scout indexes new LLMs on OpenRouter, cross-references public usage and benchmark coverage, then benches the candidates nobody has measured yet — prioritized by coverage gap. This is the latest scan.
Scanned Nov 18, 2026 · verdict: upgrade candidate surfaced
Bench scorecard
Model
Score /10
Latency
Est. cost
Notes
openai/gpt-6.2-astra-pro
10
5.5s
$0.05
-
openai/gpt-6.2-astra
10
2.1s
$0.0028
-
google/gemini-3.8-flash
10
3.6s
$0.0066
-
google/gemini-3.6-flash
10
2.1s
$0.009
-
anthropic/claude-haiku-4.5
6.2
1.0s
$0.0052
Score impacted by refusal on E8 (Security Policy).
New paid models worth watching
openai/gpt-6.2-astra-pro — no public benchmark
openai/gpt-6.2-astra — no public benchmark
Advisor notes
### Model Benchmarking Report (2026-11-18)
**Benchmark Scorecard vs. Control**
| Model | Status | Score (0-10) | Latency (ms) | vs. Control (delta) |
| :--- | :--- | :--- | :--- | :--- |
| `anthropic/claude-haiku-4.5` | Control | 6.2 | 967 | - |
| `google/gemini-3.6-flash` | Control | 10.0 | 2060 | - |
| `openai/gpt-6.2-astra-pro` | Candidate | 10.0 | 5459 | +3.8 |
| `openai/gpt-6.2-astra` | Candidate | 10.0 | 2139 | +3.8 |
| `google/gemini-3.8-flash` | Candidate | 10.0 | 3630 | +3.8 |
**Verdict**
* **Upgrade Recommendation:** While the candidates (`gpt-6.2-astra`, `gpt-6.2-astra-pro`, and `gemini-3.8-flash`) clearly outperform the underperforming control `claude-haiku-4.5` (which was hindered by security policy refusals), the incumbent `google/gemini-3.6-flash` remains at parity (10.0) with the new candidates.
* **Action:** I recommend maintaining `google/gemini-3.6-flash` as the primary control given its superior latency compared to the new Astra-Pro and Gemini-3.8-flash models. No immediate upgrade is required as there is no clear performance gain over the current top-performing incumbent.
* **Admin Decision:** The `gpt-6.2-astra` (non-pro) performs at a near-identical latency to the current `gemini-3.6-flash` (2139ms vs 2060ms). If `gpt-6.2-astra` is confirmed to be free/included in existing tiers, you may wish to authorize an admin decision to swap it as a secondary provider to diversify the fleet.
*Benchmark data: Artificial Analysis (artificialanalysis.ai)*
Machine-readable: GET /public/model-radar on the API returns this scan as JSON (CORS-open); ?scan=<scanId> reads an older one from GET /public/model-radar/scans.
The scout, bench, and advisor behind this page are LiveGraph graphs themselves — every score above is a real model call through the same dispatch path your own runs take. To see that engine in action, try the live demo — no signup, no key — then run the models above on your own graphs, on our keys or yours.