Qwen3.5 397B A17B
Sold by 8 sellers, timed at 44.2 tok/s, scored 48.2 on the Coding Index, on 7 Design Arena boards, and with results on 1 openly licensed board.
retrieved sha256 · 0bb53a77713f
The same weights served by 3 providers, each at its own list price. OpenRouter's rate is its routed rate across whichever of them it uses, and it need not match any one of them.
What sellers charge
8 sellers · 2.08x from cheapest to dearestOne tick per seller, on a log scale of input price per million from $0.36 at DeepInfra to $0.75 at Venice. 2 sellers undercut the $0.45 listed; 5 are dearer. A faint tick declares no number format. Every seller is named on the endpoint ledger below.
The same sellers grouped by the number format each declares, on the same scale. The not declared group alone spans 89% of the range, so the cheap end is not cheap for its number format. A declaration is checked by nobody: this groups the rows and certifies none.
Measured speed
27,955 requests · 8 endpoints · 30-minute windowOpenRouter's own timing of its routed traffic, weighted by requests, over one 30-minute window read . Every route is on the endpoint ledger below.
Every endpoint
8 priced · 8 timed · cheapest first- DeepInfradeepinfra/fp8/flex$0.36$2.40$0.176fp882knot matchednonenonenot matched100%
- Alibabaalibaba$0.39$2.34not listednot declared66k41 tok/s281.11s5,130 req99.73%
- DeepInfradeepinfra/fp8 · listed$0.45$3.00$0.22fp882knot matchednonenonenot matched99.48%
- Parasailparasail/fp8$0.50$3.60$0.30fp8236k65 tok/s59481ms3,337 req96.27%
- Novitanovita$0.60$3.60not listednot declared66k19 tok/s121.66s967 req100%
- StreamLakestreamlake$0.60$3.60$0.12not declared64k99 tok/s711.68s4,460 req97.45%
- DeepInfradeepinfra/fp8/priority$0.675$4.50$0.33fp882knot matchednonenonenot matched98.08%
- Venicevenice$0.75$4.50not listednot declared33k52 tok/s101.28s14 req98.63%
- DeepInfranot matchednot matchednot matchednot matchedno offer38 tok/s291.16s7 reqno offer
- DeepInfranot matchednot matchednot matchednot matchedno offer24 tok/s16752ms13,541 reqno offer
- DeepInfranot matchednot matchednot matchednot matchedno offer46 tok/s25561ms499 reqno offer
Each track runs from a route's p90, the rate nine in ten of its requests met or beat, up to its median. Weighted by requests, the figure above reports where the traffic went rather than what one route does: 4 of the 8 timed routes are slower than it, and the middle route runs at 43.5 tok/s. The widest median-to-p90 gap is Venice's, 5.2x.
DeepInfra publishes 3 offers and had 3 routes timed. The two documents share no route id, so those offers and routes are printed apart rather than paired.
Uptime is OpenRouter's own figure for each seller over the trailing day, read once a day, and the column prints the latest reading, taken . 12 of the 12 sellers with a reading fell under 99% on at least one of the 23 days held, 23 Oct 2026 to 21 Nov 2026: DeepInfra (deepinfra/fp8/flex) at 13.79% on 17 Nov 2026, DigitalOcean at 45.00% on 26 Oct 2026, Parasail at 53.83% on 15 Nov 2026, DeepInfra (deepinfra/fp8/priority) at 60.96% on 16 Nov 2026, and 8 more. Source: OpenRouter's public endpoints API, one reading a day for each seller: 1,211 sellers carried one on the latest of the 23 days held.
Source: OpenRouter's public endpoints API. 1257 sellers across 359 models when the mirror ran, 186 of them served by more than one. List prices, not what anybody paid; a discount is the seller's own. Read . Speed is OpenRouter's own timing over one 30-minute window, read .
What each seller accepts
19 parameters · 8 sellers- include_reasoningevery seller
- max_tokensevery seller
- presence_penaltyevery seller
- reasoningevery seller
- response_formatevery seller
- stopevery seller
- temperatureevery seller
- top_pevery seller
- frequency_penalty7 of 8not StreamLake
- tool_choice7 of 8not StreamLake
- tools7 of 8not StreamLake
- top_k7 of 8not StreamLake
- seed6 of 8not StreamLake and Venice
- structured_outputs6 of 8not Alibaba and Novita
- logprobs5 of 8not DeepInfra (deepinfra/fp8), DeepInfra (deepinfra/fp8/flex), and DeepInfra (deepinfra/fp8/priority)
- repetition_penalty5 of 8not Alibaba, StreamLake, and Venice
- top_logprobs5 of 8not DeepInfra (deepinfra/fp8), DeepInfra (deepinfra/fp8/flex), and DeepInfra (deepinfra/fp8/priority)
- logit_bias4 of 8
- min_p3 of 8only DeepInfra (deepinfra/fp8), DeepInfra (deepinfra/fp8/flex), and DeepInfra (deepinfra/fp8/priority)
8 of these 19 are accepted by every one of the 8 sellers that publish a list; the other 11 are declared by some sellers and not others, so what a request sending one of them gets depends on which seller serves it. Each list is the seller's own declaration in OpenRouter's endpoints API, read , and nothing here tested one.
Price position
342 priced modelsBlended $1.088 per million tokens, 3 in to 1 out: cheaper than 139 of the 342 priced models in the catalog. Each bar is one stretch of the catalog on a logarithmic price scale; the mark is this listing.
Measured quality
145 models with an index · 132 on a design arena boardEach mark is one board this model stands on, placed at the share of that board it scores above, and led to its number in the list under the rule. The columns and the ledger below carry the figures.
- 3D scenes · Models1141 Elowin rate 56.7% · source rank 73the board, 113 ranked
- Code · Models1168 Elowin rate 52.7% · source rank 73the board, 120 ranked
- Data visualisation · Models1171 Elowin rate 53.2% · source rank 70the board, 110 ranked
- Game development · Models1129 Elowin rate 50.2% · source rank 86the board, 120 ranked
- SVG · Models1128 Elowin rate 56.1% · source rank 61the board, 83 ranked
- UI components · Models1152 Elowin rate 51.4% · source rank 84the board, 116 ranked
- Websites · Models1182 Elowin rate 52.6% · source rank 76the board, 126 ranked
Position 79 of the 258 models Epoch has placed, which is not the population of either board above, on a scale anchored at Claude 3.5 Sonnet 130 and GPT-5 150. Epoch states that absolute values are meaningless by themselves and that comparisons between models are what the index is for, that values are not linearly related to benchmark accuracy, and that a highly specialised model may score low while being capable in its own domain.
Benchmark results
1 of 6 boards · 7 results| Task | All runs | Effort | Score |
|---|---|---|---|
| FrontierMath-Tiers-1-3-v2-Private | 128 | none | 0.31 |
| unstated | 0.29 | ||
| GPQA diamond | 342 | none | 0.86 |
| unstated | 0.86 | ||
| Mystery Game Puzzles | 149 | none | 0.18 |
| OTIS Mock AIME 2024-2025 | 319 | none | 0.82 |
| unstated | 0.89 |
Mean accuracy, 0 to 1, on the task the row names, as Epoch AI runs it and averaged over repeated runs. All eighteen of their tasks, including the held-out FrontierMath tiers, named as they name them. The model id carries the effort setting the run used. Epoch AI Benchmarking Hub names these runs qwen3.5-397b-a17b_none and qwen3.5-397b-a17b, and the effort beside each score says which run it is, unstated where the publisher wrote none. A run missing from a task's rows has no published result there, which is not a score of zero. All runs is how many runs the publisher scored on the task, every model's. Read from epoch.ai ↗, which dates its own data . Epoch AI Benchmarking Hub data, by Epoch AI, under CC BY 4.0.
1 of the 6 openly licensed boards Vitna mirrors carries a result for Qwen3.5 397B A17B, each printed as its publisher published it and none added to another. MTEB, LiveBench, HELM, Vectara hallucination leaderboard and Aider polyglot have not published one under a name that matches its listing, and names are matched to the catalog exactly and never guessed. A task that is a link opens its column on /capability, with every run scored on it; an effort that is a link opens its run's row there, where the run is ranked on Vitna's scale. Every board, with its licence and what its publisher measures, is on /benchmarks.
Source: OpenRouter (openrouter.ai/models), as of 17 Nov 16:29 UTC. sha256 0bb53a77713f over 853,270 bytes.
Catalog: OpenRouter's public model list.
Intelligence, coding and agentic indexes: Artificial Analysis, as published by OpenRouter.
Elo, win rate and rank: Design Arena, as published by OpenRouter.
Source: each model's own Hugging Face repository (huggingface.co), as of 20 Nov 2026. 162 repositories read, 4 did not answer.
Licence, parameter count, model class and gating: each model's own Hugging Face repository, the one OpenRouter names for the listing.
A licence is the tag the uploader put on the repository, read from nothing else and verified by nobody.
A parameter count is the total the repository's safetensors files declare; for a mixture of experts that is every expert, not the share one token runs through.
Source: the Epoch AI Capabilities Index (epoch.ai/eci), as of 20 Nov 2026. 258 models placed over 2917 benchmark results.
Epoch AI Benchmarking Hub data, by Epoch AI, under CC BY 4.0.
Vitna ran none of these evaluations and adjusted none of these figures.
