inference/research
verifiedhigh · 0.90updated 2026-10-04

Cerebras Inference API pricing - Developer Tier, per 1M tokens (recovered from the rendered DOM 2026-10-04). The Enterprise tier remains quote-only - that part of the old negative was true.

Cerebras cloud (US); no regional GPU SKU to price

Fields

id
supply/rent-cerebras-api-token-pricing
Kind
cloud
Region
Cerebras cloud (US); no regional GPU SKU to price
Channels
serverlessapinegative
Price (USD)
0.35
Price basis
USD per 1M TOKENS, LIST PRICE (ASKING), observed 2026-10-04 from the RENDERED DOM of the Cerebras Inference pricing page - this is a CORRECTION, see notes. Developer Tier, self-serve pay-as-you-go, obtainable with a free $5 promotional credit: OPENAI GPT OSS 120B, ~3000 tokens/s, $0.35/M tokens input and $0.75/M tokens output (price_usd = GPT OSS 120B input); QWEN Qwen 3.8 27B, ~1850 tokens/s, $0.99/M tokens input and $1.49/M tokens output. PUBLISHED THROUGHPUT ON THE SAME ROW, which is unusual: both models publish a tokens/s figure beside the price, so speed is a list attribute here rather than something you have to benchmark. NO GPU-HOUR RATE IS PUBLISHED, and Cerebras does not name serving silicon on this page beyond its own wafer-scale system. PERISHABLE-FIGURE STAMP (added 2026-10-04, uniform across data/supply/ so it is greppable). PRICE CLASS: PUBLISHED LIST PRICE (ASKING) - an asking rate a new customer is quoted, NOT a realised, discounted or negotiated transaction rate. OBSERVED: 2026-10-04. RE-VERIFY BY: 2027-01-04 - the repo's quarterly re-verification bar; after that date this figure is presumed drifted and must be re-checked against its source before it is quoted again. UNIT: USD per 1M tokens, Developer Tier, input (price_usd); output rates and a tokens/s figure share the row. READING RULE: THE UNIT AND THE TIER BOTH MATTER HERE, which is unusual. (a) TIER: do not read this as the rate anyone pays in production. The Developer tier is footnoted 'For development, evaluation, and experimentation; not intended for production use.' The Enterprise tier - the only one marked 'Production-ready capacity', with 'Best priority, higher rate limits, and enhanced latency options' and a 'Dedicated account team' - prices as 'Contact us for a quote' and publishes NO number at all. So Cerebras PRODUCTION pricing is genuinely not public, and this rate is specifically a non-production rate that doubles as a benchmark-benchmarking trap because of the $5 promotional credit (see notes). (b) UNIT: no GPU-hour rate is published and Cerebras does not name serving silicon on this page beyond its own wafer-scale system, so this token rate cannot be decomposed into hardware cost or sanity-checked against a GPU-hour the way a GPU-hour provider's can be. (c) This figure was RECOVERED from a rendered DOM on 2026-10-04 after an earlier static extraction wrongly recorded it as unpublished - see notes before citing the provenance.
Availability
in-stock
Lead time source
structural-na
Lead time basis
lead_time_weeks is null because NO NUMBER APPLIES, and the reason is specific rather than generic: this record prices a metered service term (tokens, egress, storage) rather than hardware capacity, so there is no order-to-delivery interval to report at all. Calling this 'no published lead time' would be wrong in the other direction - nothing was withheld.
Export control regime
not-applicable-service
Export control note
No accelerator crosses a border in a token-API transaction, so the ECCN 3A090 hardware licence regime does not attach. Export control instead appears as an end-user and terms question: the provider applies geographic and end-use restrictions to service access, and a customer can be refused under sanctions or end-use policy regardless of where the compute sits. For a restricted destination, read the provider's terms, not the EAR. Sources: spl-bis-2025-05-ai-diffusion-rescission, spl-bis-2026-05-d5-guidance.
notes
THIS RECORD WAS A NEGATIVE UNTIL 2026-10-04 AND THE NEGATIVE WAS WRONG. It previously carried no price_usd and asserted that Cerebras's per-model per-1M-token rates 'did not appear in the static HTML extraction'. That inference does not hold. Re-fetching the page with a rendered-DOM browser on 2026-10-04 returns the Developer Tier rate card in full (source pd-cerebras-pricing-dom-rendered): GPT OSS 120B $0.35 in / $0.75 out per 1M at ~3000 tok/s, Qwen 3.8 27B $0.99 in / $1.49 out per 1M at ~1850 tok/s. The rates were published all along; the table is rendered client-side, so a static extractor reads the section headings and the footnote prose and misses the table body entirely. The record has been corrected rather than merely annotated, because a wrong negative is worse than a missing record: it asserts that a vendor is opaque when the vendor is merely JS-rendered, and it invites a reader to believe no price exists. THE METHODOLOGICAL LESSON IS THE MOST PORTABLE THING HERE: 'I could not extract it' is a claim about the retrieval method, not about the vendor's disclosure. This repository already holds that principle in rent-pricing-not-public-negatives-2026-10 - but it was applied to Cerebras without the distinction being made, and one static probe was treated as proof of opacity. A negative claim should require two INDEPENDENT methods of failure before it is recorded. There is already a gotcha in this repo, rent-vendor-pricing-tables-are-js-but-their-apis-are-not, describing exactly this failure mode; this record is the proof that the gotcha was under-applied, and the 2026-10-04 pass is what caught it. WHAT SURVIVES AS A GENUINE NEGATIVE, NARROWED CORRECTLY: (1) THE ENTERPRISE TIER IS QUOTE-ONLY. It is the tier marked 'Production-ready capacity', 'Best priority, higher rate limits, and enhanced latency options' and 'Dedicated account team', and its price cell reads 'Contact us for a quote'. There is no number to extract because no number is published. So the correct negative is 'Cerebras PRODUCTION pricing is not public', NOT 'Cerebras pricing is not public'. (2) Enterprise-only capabilities are footnoted '**' = 'Available by agreement' - custom model weights, and fine-tuning/training services - with no published rate. (3) There is still no GPU-hour rate and no naming of serving silicon on the pricing page, so a Cerebras token rate cannot be sanity-checked against underlying hardware the way Fireworks' can. THE BENCHMARKING TRAP, unchanged and still true, now with the trap quantified: the Developer tier hands out a one-time $5 promotional credit on adding a payment method. It expires 30 days after activation, and API and Playground access PAUSE when it expires or is exhausted unless PayGo credits are purchased separately. The credit is non-transferable, has no cash value, and may be withheld or revoked for fraud, abuse or duplicate accounts. A benchmark run immediately after signup therefore measures a $0 effective price, not a real one - the first few thousand requests are free. Measure only after the credit is exhausted or on a paid tier, and treat the published $0.35 as a floor that no production customer necessarily pays. STRUCTURAL POINT UNCHANGED AND STILL WORTH MAKING: Cerebras sells inference on its own wafer-scale hardware at a scale where per-GPU-hour is not a purchasable unit - the accelerator is the system - which is precisely why its per-token rate cannot be cross-checked the way a GPU-hour provider's can. Cerebras is also routable through AWS Marketplace, OpenRouter, HuggingFace and Vercel, each of which adds its own layer and its own markup, so the direct rate is not necessarily the rate a buyer pays.

Notes

THIS RECORD WAS A NEGATIVE UNTIL 2026-10-04 AND THE NEGATIVE WAS WRONG. It previously carried no price_usd and asserted that Cerebras's per-model per-1M-token rates 'did not appear in the static HTML extraction'. That inference does not hold. Re-fetching the page with a rendered-DOM browser on 2026-10-04 returns the Developer Tier rate card in full (source pd-cerebras-pricing-dom-rendered): GPT OSS 120B $0.35 in / $0.75 out per 1M at ~3000 tok/s, Qwen 3.8 27B $0.99 in / $1.49 out per 1M at ~1850 tok/s. The rates were published all along; the table is rendered client-side, so a static extractor reads the section headings and the footnote prose and misses the table body entirely. The record has been corrected rather than merely annotated, because a wrong negative is worse than a missing record: it asserts that a vendor is opaque when the vendor is merely JS-rendered, and it invites a reader to believe no price exists. THE METHODOLOGICAL LESSON IS THE MOST PORTABLE THING HERE: 'I could not extract it' is a claim about the retrieval method, not about the vendor's disclosure. This repository already holds that principle in rent-pricing-not-public-negatives-2026-10 - but it was applied to Cerebras without the distinction being made, and one static probe was treated as proof of opacity. A negative claim should require two INDEPENDENT methods of failure before it is recorded. There is already a gotcha in this repo, rent-vendor-pricing-tables-are-js-but-their-apis-are-not, describing exactly this failure mode; this record is the proof that the gotcha was under-applied, and the 2026-10-04 pass is what caught it. WHAT SURVIVES AS A GENUINE NEGATIVE, NARROWED CORRECTLY: (1) THE ENTERPRISE TIER IS QUOTE-ONLY. It is the tier marked 'Production-ready capacity', 'Best priority, higher rate limits, and enhanced latency options' and 'Dedicated account team', and its price cell reads 'Contact us for a quote'. There is no number to extract because no number is published. So the correct negative is 'Cerebras PRODUCTION pricing is not public', NOT 'Cerebras pricing is not public'. (2) Enterprise-only capabilities are footnoted '**' = 'Available by agreement' - custom model weights, and fine-tuning/training services - with no published rate. (3) There is still no GPU-hour rate and no naming of serving silicon on the pricing page, so a Cerebras token rate cannot be sanity-checked against underlying hardware the way Fireworks' can. THE BENCHMARKING TRAP, unchanged and still true, now with the trap quantified: the Developer tier hands out a one-time $5 promotional credit on adding a payment method. It expires 30 days after activation, and API and Playground access PAUSE when it expires or is exhausted unless PayGo credits are purchased separately. The credit is non-transferable, has no cash value, and may be withheld or revoked for fraud, abuse or duplicate accounts. A benchmark run immediately after signup therefore measures a $0 effective price, not a real one - the first few thousand requests are free. Measure only after the credit is exhausted or on a paid tier, and treat the published $0.35 as a floor that no production customer necessarily pays. STRUCTURAL POINT UNCHANGED AND STILL WORTH MAKING: Cerebras sells inference on its own wafer-scale hardware at a scale where per-GPU-hour is not a purchasable unit - the accelerator is the system - which is precisely why its per-token rate cannot be cross-checked the way a GPU-hour provider's can. Cerebras is also routable through AWS Marketplace, OpenRouter, HuggingFace and Vercel, each of which adds its own layer and its own markup, so the direct rate is not necessarily the rate a buyer pays.

Cross-references

Accelerators (1)