Qwen 3.8 Flash API: $0.15/$0.47, and the Reasoning Tokens That Bill You

Qwen 3.8 Flash bills $0.15 per million input tokens and $0.47 output — the cheapest output rate in the Flash tier. It is also a reasoning model, which is the part that catches people: send a short max_tokens and you get an empty string back, having paid for tokens you never saw.

Model ID:      bailian/qwen3.8-flash   (alias: qwen3.8-flash)
Price:         $0.15 in / $0.47 out per 1M
Context:       1,131,072 tokens
Max output:    131,072 tokens
Modality:      text + image → text
Endpoints:     /v1/chat/completions, /v1/responses
Parameters:    temperature, top_p, max_tokens, stop, tools,
               tool_choice, response_format, reasoning

Read from the live Ofox /v1/models endpoint on 6 September 2026. Behaviour below was measured against that endpoint the same day.

The Empty Response Everyone Hits First

We sent max_tokens: 30 and got back an empty string. Not an error — a successful response, billed:

{"content": "",
 "usage": {"prompt_tokens": 66, "completion_tokens": 30,
           "completion_tokens_details": {"reasoning_tokens": 30}}}

Thirty completion tokens, all of them reasoning tokens, nothing visible. The model spent the entire budget thinking before emitting a single character of answer.

Two fixes, both measured on the same prompt:

Raising the ceiling works because the model only needed 27 tokens total; the 30-token cap cut it off mid-thought.

Turning reasoning off works better. For a prompt this trivial, effort: none produced the same answer in 1 token instead of 27 — a 27x reduction. On short-output work at scale, that is the difference between $0.47 per million and a small fraction of it.

The practical rule on any reasoning model: max_tokens has to cover reasoning plus the answer, and reasoning bills at the output rate whether you see it or not. Budgeting max_tokens as if it were answer length is how a working model looks broken.

Where $0.47 Sits in the Flash Tier

All rates from the live Ofox catalog, 6 September 2026:

Qwen 3.8 Flash and GLM-5.3 Flash are effectively tied — identical input rate, three cents apart on output. Between those two, price is not the deciding factor and you should test both on your actual task.

Against the rest of the tier the gap is large. Qwen 3.8 Flash is 2.8x cheaper on output than DeepSeek V4 Flash and 16x cheaper than Gemini 3.8 Flash.

Two things that gap does not tell you:

• DeepSeek V4 Flash has an independent score and Qwen 3.8 Flash does not. DeepSeek V4 Flash 0731 carries 88.80% on vals.ai SWE-Bench Verified — higher than Claude Opus 4.8 — while Qwen 3.8 Flash is not on that leaderboard at all. You are paying less for a model with less public evidence behind it, which may be fine and is worth knowing.
• DeepSeek V4 Flash allows a 384,000-token completion against Qwen’s 131,072. If you generate very long single outputs, that ceiling matters more than the rate.

What It Costs in Practice

At $0.15 / $0.47, with the reasoning caveat above folded in:

The first two rows are the same job. The difference is whether reasoning is on — roughly 4.3x on this shape of work. That is the number to internalise before you pick this model for a high-volume pipeline.

Calling It

curl -X POST https://api.ofox.io/v1/chat/completions \
  -H "Authorization: Bearer $OFOX_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "bailian/qwen3.8-flash",
    "messages": [{"role": "user", "content": "Classify the sentiment: ..."}],
    "reasoning": {"effort": "none"},
    "max_tokens": 50
  }'

Read usage.completion_tokens_details.reasoning_tokens on your first few calls. If it is a large fraction of completion_tokens on work that does not need deliberation, effort: none is free money.

The model takes image input as well as text, so it also covers cheap vision work — captioning, OCR-adjacent extraction, screenshot classification — at a rate well below the vision-capable Flash models above.

When Not to Use It

• You need an independent benchmark to justify the choice. Qwen 3.8 Flash is not on vals.ai SWE-Bench or the Artificial Analysis index. DeepSeek V4 Flash at $0.44 / $1.32 has a published 88.80%.
• You generate very long single responses. The 131,072-token completion ceiling is a third of DeepSeek V4 Flash’s.
• The work is genuinely hard. This is a Flash-tier model. If accuracy on complex reasoning decides your outcome, the frontier comparison covers what the extra money buys.
• You are already on GLM-5.3 Flash. Three cents per million output is not a migration reason.

Related

• Qwen 3.8 Max pricing and access — the flagship in the same family at $2.00 / $6.00.
• Qwen 3.8 Max vs DeepSeek V4 Flash — the same cheap-versus-flagship question one tier up.
• Qwen 3.8 27B run locally — if the reason you want a cheap Qwen is that you would rather self-host.

Sources

• https://ofox.io/models/bailian/qwen3.8-flash
• https://ofox.io/models/deepseek/deepseek-v4-flash-0731
• https://ofox.io/models/z-ai/glm-5.3-flash

Prices, context lengths, completion ceilings, supported parameters and endpoints were read from the live Ofox /v1/models endpoint on 6 September 2026. The reasoning-token behaviour is from three live /v1/chat/completions calls against bailian/qwen3.8-flash the same day, identical prompt, differing only in max_tokens and reasoning.effort. That is a small sample on one route on one day rather than a specification — measure it on your own workload before sizing a pipeline around it. The SWE-Bench figure for DeepSeek V4 Flash is from vals.ai, leaderboard updated 2026-08-26.