Qwen3.8-27B Cerebras API Pricing: $0.99/$1.49

Cerebras has launched a hosted Qwen3.8-27B route at $0.99 per million input tokens and $1.49 per million output tokens. The public model ID is qwen-3.8-27b, and Cerebras lists generation speed at about 1,500 tokens per second.

This changes the buying decision from local-only evaluation to a direct API comparison. Qwen’s Apache-2.0 BF16 and FP8 weights remain available for self-hosting, while Cerebras now offers a managed route with published prices, context limits, rate limits, and an OpenAI-compatible API.

Qwen3.8-27B Cerebras pricing and limits

Item	Free Trial	Developer / paid
Input price per 1M tokens	Free-trial quota	$0.99
Output price per 1M tokens	Free-trial quota	$1.49
Context window	65,536	131,072
Maximum output	32,768	40,960
Requests per minute	5	300
Uncached tokens per minute	30,000	150,000
Total tokens per minute	90,000	450,000
Daily token limit	1 million	None listed

Cerebras labels the rates Developer pricing. Enterprise capacity and discounts are custom. The model accepts text plus base64-encoded PNG or JPEG images and returns text. It supports reasoning, streaming, tool calling, parallel tool calling, Structured Outputs, and prompt caching.

The hosted route uses 128K context rather than the open model card’s 262K native context or documented extension to 1M. Buyers should use the host’s enforceable API limit when sizing a workload.

What 1,500 tokens per second means

The speed claim describes generation throughput under Cerebras’ service conditions. It does not guarantee end-to-end latency for every prompt. Time to first token, prompt length, reasoning effort, queueing, tool calls, network distance, and rate-limit pressure still affect wall time.

At the published output price, one million generated tokens costs $1.49. If a workload actually sustained 1,500 output tokens per second, one million output tokens would take about 11.1 minutes of generation time. That calculation is a throughput translation, not an SLA or capacity promise.

For a workload with 100 million input tokens and 20 million output tokens, the listed Developer rates total $128.80 before any enterprise discounts. The same raw token mix costs $800 on GPT-5.6 Sol at its current direct Standard rate. This price comparison does not claim equal quality or equal token use.

Hosted API versus local Qwen3.8-27B

Alibaba’s Qwen team released the 27-billion-parameter dense model under Apache 2.0. The official BF16 safetensors total about 55.56 GB, while the official FP8 files total about 30.87 GB. Runtime memory also needs the vision components, KV cache, buffers, and concurrency headroom.

Self-hosting can win when utilization is steady, data must stay local, or the team needs a modified runtime. The Cerebras API is easier to evaluate when demand is bursty or engineering time matters. At $0.99/$1.49 per million tokens, the managed bill stays below one dedicated accelerator for many early workloads.

Use the local AI versus API calculator with measured traffic, GPU price, utilization, electricity, and operator time. Avoid comparing the API invoice with only the weight download size.

Labs coverage decision

The canonical dataset now includes the Cerebras route, but it is not yet in the Cost-per-Task Leaderboard. The current 49-task harness uses a fixed set of routed model endpoints and records actual billed cost. The exact Cerebras qwen-3.8-27b route is not available through that harness today.

The coverage note records the availability boundary. Labs will add the model after a supported route is wired, the exact snapshot and reasoning configuration are pinned, and a complete 49-task run records tokens, retries, latency, correctness, and billed cost. Cerebras’ speed claim will remain a separate provider measurement unless reproduced under a fixed workload.

What buyers should do

1. Use qwen-3.8-27b explicitly and record the host; similarly named Qwen routes can use different snapshots and limits.
2. Test reasoning at the default high setting and with reasoning_effort: none where latency matters.
3. Measure time to first token and accepted-task latency instead of treating 1,500 tokens/s as end-to-end speed.
4. Pin vision fixtures and tool schemas when comparing against a text-only or local runtime.
5. Recalculate self-hosting at realistic utilization and include monitoring, redundancy, and engineering time.

Compare current routes on the maintained Qwen pricing page, then model tokens in the calculator. Teams comparing other managed Qwen hosts can check Novita’s current Qwen catalog and rerun the same acceptance test.

Affiliate disclosure: AI Pricing Guru may earn a commission from the sponsored Novita link at no extra cost to you. It does not affect this analysis.

Bottom line

Qwen3.8-27B is now a priced hosted API, not only an open-weight deployment choice. Cerebras charges $0.99 input and $1.49 output per million tokens and advertises about 1,500 output tokens per second, with a 128K paid context window.

That combination is attractive for coding, multimodal agents, and private evaluations, but throughput and list price do not establish cost per successful task. Test the exact route on production-shaped work before replacing an incumbent model or buying local hardware.

Sources: Cerebras’ official Qwen3.8-27B model page, model catalog, and pricing page; Qwen’s official BF16 model card and FP8 release. Pricing and availability checked September 3, 2026.