Qwen3.8 27B repriced across 5 hosts: Darkbloom input down 28%, Wafer output up 80%
Published 28 September 2026
Price move, per 1M tokens
- Wafer · input↓ 13%$0.083 → $0.072
- Wafer · output↑ 80%$2.45 → $4.40
- Wafer · cache read↓ 28%$0.0802 → $0.0574
- DekaLLM · inputunchanged$0.080 → $0.080
- DekaLLM · outputunchanged$2.50 → $2.50
- DekaLLM · cache read↓ 54%$0.087 → $0.040
- Darkbloom · input↓ 28%$0.069 → $0.050
- Darkbloom · outputunchanged$2.20 → $2.20
- Reka · input↓ 20%$0.090 → $0.072
- Reka · outputunchanged$4.40 → $4.40
- Reka · cache read↓ 32%$0.0850 → $0.0574
- Ionstream · input↓ 15%$0.105 → $0.089
- Ionstream · outputunchanged$2.55 → $2.55
- Ionstream · cache read↓ 15%$0.100 → $0.085
| host | rate | was | now | change |
|---|---|---|---|---|
| Wafer | input | $0.083 | $0.072 | ↓ 13% |
| Wafer | output | $2.45 | $4.40 | ↑ 80% |
| Wafer | cache read | $0.0802 | $0.0574 | ↓ 28% |
| DekaLLM | input | $0.080 | $0.080 | — |
| DekaLLM | output | $2.50 | $2.50 | — |
| DekaLLM | cache read | $0.087 | $0.040 | ↓ 54% |
| Darkbloom | input | $0.069 | $0.050 | ↓ 28% |
| Darkbloom | output | $2.20 | $2.20 | — |
| Reka | input | $0.090 | $0.072 | ↓ 20% |
| Reka | output | $4.40 | $4.40 | — |
| Reka | cache read | $0.0850 | $0.0574 | ↓ 32% |
| Ionstream | input | $0.105 | $0.089 | ↓ 15% |
| Ionstream | output | $2.55 | $2.55 | — |
| Ionstream | cache read | $0.100 | $0.085 | ↓ 15% |
Wafer input: ↓ 13%Wafer output: ↑ 80%DekaLLM cache read: ↓ 54%Darkbloom input: ↓ 28%Reka input: ↓ 20%Ionstream input and cache read: ↓ 15%
If you buy
On Wafer, output costs more while input and cache reads cost less, so this host suits prompt-heavy, short-answer work; re-check your output budget before your next top-up.
The model
Qwen3.8 27Bopen the model →
Open weights27.8B parameters262K context20 hosts
intelligence63rd of 168 writing93rd of 168 coding45th of 168 agents33rd of 55
Qwen3.8 27B is a downloadable model you can run yourself, and its measured strength is maths and reasoning. Agentic recovery and tool use sit near the bottom of the field, so it suits a solver rather than an agent that has to get itself back on track.
Source
Read on openrouter.ai · we published this on 28 September 2026.
read the original at openrouter.ai ↗