2027 planning guide — evidence checked 20 November 2026

Cerebras API vs Plugsky: speed or breadth?

Cerebras sells unusually fast inference for a deliberately small catalogue. Plugsky sells a broader catalogue with flat monthly plans and private deployment. The decision is not which API is faster in the abstract; it is whether your workload needs one model at the lowest latency or several models under one commercial agreement.

Short answer. Choose Cerebras when a supported model is the right model and generation speed is the product constraint. Its shared catalogue lists GPT-OSS 120B and Qwen 3.8 27B, priced per token, with paid context up to 131,072 tokens. Choose Plugsky when you need 39 public model IDs, a 262,144-token context on most text models, no monthly token cap on Builder or Scale, and deployment in your VPC, on-premises or air-gapped.

The two services are easy to compare badly. Cerebras publishes a token rate and an inference speed; Plugsky publishes a plan ladder and a catalogue. A page that lines up the token rates and declares a winner ignores the reason teams choose Cerebras, which is time to last token on a supported model.

This guide separates the model question from the speed question, and flags where a Plugsky public ID has a familiar name but a different serving engine.

How we evaluated this

We used five criteria and public evidence read on 20 November 2026.

  • Latency and throughput, 30%. Published speed and the limits that govern sustained use.
  • Catalogue fit, 25%. Whether the model you need is served at the context and output you need.
  • Cost shape, 20%. Per-token spend versus a flat plan with fair-use limits.
  • Deployment, 15%. Cloud regions, private endpoints, on-premises and air-gap.
  • Operations, 10%. API compatibility, rate-limit clarity and support.

We did not run latency, throughput or quality tests. Cerebras's speed figures are vendor statements, not reproduced results.

The catalogues are not interchangeable

Cerebras Shared Inference lists two models: gpt-oss-120b and qwen-3.8-27b. Additional families are available through Dedicated Inference. That focus is the point: the hardware and the model are tuned together.

Plugsky's catalogue is broader but uses its own public IDs. Several IDs carry legacy family names while being served by NVIDIA engines. This matters when you compare model names across the two services.

Name check. plugsky-gpt-oss is a legacy compatibility ID served by NVIDIA Nemotron 3 Nano Omni 30B (A3B), not OpenAI GPT-OSS 120B. plugsky-qwen-next is a legacy ID served by the same Nano Omni engine, not Qwen 3.8 27B. plugsky-qwen-vl is a legacy ID served by NVIDIA Nemotron 3.5 Super VL 120B. If you need the exact GPT-OSS or Qwen weights that Cerebras lists, Cerebras is the direct route.

Cerebras Inference and Plugsky compared on published facts, 20 November 2026
DimensionCerebrasPlugsky
Hardware modelWafer-scale inference cloudNVIDIA pools plus GPU Share and private deployment
Shared catalogueGPT-OSS 120B and Qwen 3.8 27B39 public model IDs across several engines
Largest shared modelGPT-OSS 120BNemotron 3 Ultra 550B MoE (A55B)
Paid context131,072 tokens for GPT-OSS 120B262,144 tokens on most text models
Maximum output40,960 tokens on paid GPT-OSS8,192 to 16,384 tokens depending on ID
Published speedAbout 3,000 tokens per second for GPT-OSS 120BNot published as a single figure
Pricing unitPer token, with a $5 trial creditFlat monthly plans from $0 to $120
GPT-OSS 120B rate$0.35 input / $0.75 output per 1MNo GPT-OSS 120B weights in the catalogue
Developer limits1,000 RPM, 500K uncached TPM, 1.5M total TPM600 RPM on Builder; 1,200 RPM on Scale, fair use
Private deploymentDedicated Inference and enterprise agreementsVPC, on-premises and air-gapped by contract
Free entry$5 credit, expires after 30 days, card requiredFree plan, no card, 100 requests and 50,000 tokens per day

Where Cerebras is the better choice

Cerebras wins when latency is a feature. If an autocomplete, voice agent or interactive coding tool feels slow at 40 tokens per second and acceptable at 1,000, the catalogue is secondary. The published rate of about 3,000 tokens per second for GPT-OSS 120B is a vendor figure, but the product is built around that constraint, and the limits are high enough for sustained developer traffic.

It also wins when one model is enough. A team standardized on GPT-OSS 120B does not need 39 IDs; it needs that model, a stable endpoint and a predictable per-token bill. At low and medium volume, paying only for what you use can beat a flat plan. The trade-off is concentration: you are buying speed for a small set of models, and the paid context ceiling is 131,072 tokens.

Where Plugsky is the better choice

Plugsky wins when the catalogue and the commercial shape matter more than the fastest possible token. A team building several products may need a general chat model, a coding model, a long-context model, a vision model, a translation model and embeddings. Plugsky puts those behind one key and one request format, with a 262,144-token context on most text models.

It also wins on cost predictability. Builder is $60 per month and Scale is $120, with no monthly token cap and no per-token overage; fair-use concurrency and RPM limits apply. Scale raises concurrency to 16 and RPM to 1,200. Deployment is the third difference: Plugsky offers its cloud, a private VPC or endpoint, on-premises and air-gapped deployment by contract, while Cerebras's published self-serve product is a cloud API with dedicated inference by agreement.

Cost math at two volumes

Assume GPT-OSS 120B on Cerebras at $0.35 input and $0.75 output per million tokens, a 4:1 mix, list prices, no cache and no batch. Plugsky Builder is $60 with no token cap; Scale is $120.

Monthly cost at two token volumes, list prices on 20 November 2026
VolumeCerebras calculationCerebrasPlugsky Builder
40M input / 10M output(40 × $0.35) + (10 × $0.75)$21.50$60.00
80M input / 20M output(80 × $0.35) + (20 × $0.75)$43.00$60.00
112M input / 28M output(112 × $0.35) + (28 × $0.75)$60.20$60.00
160M input / 40M output(160 × $0.35) + (40 × $0.75)$86.00$60.00

The crossover is roughly 112 million input and 28 million output tokens a month at this mix. Below it, Cerebras is cheaper on raw token spend; above it, the flat plan is cheaper. That is only a cost comparison. If Cerebras is three to five times faster on your workload, the cheaper option may still be the slower one.

The comparison also assumes the same model. Plugsky does not serve GPT-OSS 120B weights, so a team that needs that exact model cannot substitute plugsky-gpt-oss and expect the same output. The honest comparison is Cerebras GPT-OSS 120B against a Plugsky model that passes your own evaluation. See how fair-use limits work.

What we did not verify

We did not reproduce Cerebras's speed claims, measure time to first token from any region, or test sustained throughput at the published rate limits. We did not verify the full Dedicated Inference model list, dedicated pricing, or enterprise feature availability by region. We did not test GPT-OSS 120B or Qwen 3.8 27B against any Plugsky model on quality, tool calling or long-context recall.

For Plugsky, we did not verify the current SOC 2 report; request it under NDA and confirm the period. The 99.9% figure is an uptime target, and the contract schedule governs. Fair-use concurrency and RPM limits are not a promise of infinite capacity.

Frequently asked questions

Is Cerebras faster than Plugsky?

Cerebras publishes about 3,000 tokens per second for GPT-OSS 120B on its wafer-scale hardware. Plugsky does not publish a single speed figure. We did not benchmark either service, so the only reliable answer is a test on your prompts from your region.

Does Plugsky serve GPT-OSS 120B?

No. plugsky-gpt-oss is a legacy compatibility ID served by NVIDIA Nemotron 3 Nano Omni 30B (A3B). It does not serve OpenAI GPT-OSS 120B weights. The same caution applies to plugsky-qwen-next, which is also served by Nano Omni.

Which is cheaper at 40 million input tokens a month?

At 40 million input and 10 million output tokens, Cerebras GPT-OSS 120B costs $21.50 at list prices. Plugsky Builder is $60 flat. Cerebras is cheaper at that volume, but the models are different.

What context window does Cerebras offer?

GPT-OSS 120B has a 65,536-token context on the free trial and 131,072 tokens on the paid Developer tier, with maximum output of 32,768 and 40,960 tokens respectively. Plugsky serves 262,144 tokens on most text models.

Can I deploy Cerebras in my own VPC?

Cerebras offers Dedicated Inference and enterprise agreements. Plugsky offers private VPC or endpoint, on-premises and air-gapped deployment by contract. Confirm the scope in writing for either vendor.

Sources

Last checked 20 November 2026. Verify prices and terms before committing. Not paid.