Commit f680957
fix(adminapi): complex preset filters to actual reasoner variants
The capability flag "reasoning: true" is set by OpenRouter on any model
whose API accepts a reasoning parameter — which includes small 8B/14B
Qwens that don't actually perform at reasoner level. Listing them
under the complex Suggest button was misleading.
Require the model id to match a genuine-reasoner naming convention:
-thinking / :thinking suffix, /qwq, /deepseek-r<n>, -r1-, -reasoner.
Kept the ≤ $5/M prompt cap and multilingual/tools/non-free filters.
After the fix, complex Suggest returns 9 candidates sorted by prompt
price, starting at qwen/qwen3-30b-a3b-thinking-2507 ($0.08), through
qwen/qwen3-235b-a22b-thinking-2507 ($0.13), deepseek/deepseek-r1-0528
($0.5), moonshotai/kimi-k2-thinking ($0.6), up to qwen/qwen3-max-
thinking ($0.78). No more qwen3-8b.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>1 parent 572d176 commit f680957
1 file changed
Lines changed: 10 additions & 1 deletion
| Original file line number | Diff line number | Diff line change | |
|---|---|---|---|
| |||
43 | 43 | | |
44 | 44 | | |
45 | 45 | | |
| 46 | + | |
| 47 | + | |
| 48 | + | |
| 49 | + | |
| 50 | + | |
| 51 | + | |
| 52 | + | |
| 53 | + | |
46 | 54 | | |
47 | 55 | | |
48 | 56 | | |
| |||
94 | 102 | | |
95 | 103 | | |
96 | 104 | | |
97 | | - | |
| 105 | + | |
98 | 106 | | |
99 | 107 | | |
100 | 108 | | |
101 | 109 | | |
102 | 110 | | |
103 | 111 | | |
| 112 | + | |
104 | 113 | | |
105 | 114 | | |
106 | 115 | | |
| |||
0 commit comments