Cerebras

Executive Summary

What it is: Cerebras is an inference-only provider that runs open-weight models from OpenAI OSS, Google DeepMind (Gemma), Zhipu/Z.ai, Moonshot AI, Alibaba Qwen, Meta, Mistral, DeepSeek, and others on its custom Wafer-Scale Engine (WSE-3) chips at speeds up to roughly 3,000 tokens/sec. As of June 30, 2026 the shared public API serves three models: GPT OSS 120B (production), plus Gemma 4 31B and GLM 4.7 (both preview). A flat-rate subscription coding product called Cerebras Code (Pro $50/mo, Max $200/mo) remains sold out. Enterprise customers get dedicated endpoints for 30+ models including Kimi K2.6. Cerebras went public on NASDAQ (CBRS) at $185/share on May 14, 2026 and reported its first quarterly results as a public company on June 23, 2026. Source: https://www.cerebras.ai/pricing

What to watch out for: Two stories dominate June. First, Cerebras delivered its first earnings report since the IPO: Q1 2026 revenue was $193.4 million, up 92% year-over-year, but the company guided its core gross margin down from 46.5% in Q1 to 36-38% in Q2 and 38-41% for the full year, citing a plan to rent back equipment from a large customer. The stock fell roughly 20% over the two days around the report (closing at $226.72 on June 23, down 28% from its post-IPO opening print of $350). Second, customer concentration is extreme: G42 and the UAE's MBZUAI together accounted for 86% of Cerebras' 2025 revenue, and June reporting confirmed G42 will deploy a 64-system Cerebras supercomputer in India. For developers, both Cerebras Code subscription tiers (Pro and Max) remain sold out for the second consecutive month with no reopening date. Source: https://www.cnbc.com/2026/06/23/cerebras-cbrs-q1-earnings-report-2026.html

Bottom line: June added a third public API model (Gemma 4 31B, the first multimodal/vision model on the platform and the first Google DeepMind model) while exposing the company's financial and concentration risks through its first earnings. For coding buyers, the practical state is unchanged: speed leadership is intact, the shared API still has only three models, prompt-caching pricing is still undisclosed, and the flat-rate coding product is still inaccessible. Teams should continue testing the free API tier and treat Cerebras as a speed layer, not a Claude/Opus replacement, for complex coding quality. Source: https://inference-docs.cerebras.ai/models/overview

Key Terms

  • Wafer-Scale Engine (WSE-3) -- Cerebras's custom AI processor, roughly 58x larger than a standard GPU die, built for ultra-fast inference by keeping model weights on-chip and avoiding external-memory loads. Source: Cerebras – Chip
  • Dedicated Endpoint -- A private, provisioned inference instance reserved for one organization, offering guaranteed throughput, custom model weights, and fine-tuning. Not available on the shared API; the 30+ model catalog lives here. Source: Inference-Docs – Overview
  • Cerebras Code -- A flat-rate subscription product giving access to top open-weight models for coding via API, usable with third-party IDEs and CLI tools (not a standalone IDE). Both tiers are sold out. Source: Cerebras – Pricing
  • Prompt Caching -- Stores and reuses previously processed prompt tokens to cut latency and cost on repeated queries. Cerebras lists prompt caching as supported on Gemma 4 31B and other models, but does not publish a price discount for cached tokens. Source: Inference-Docs – Gemma 4 31B
  • Multi-LoRA -- Multi-adapter support for Low-Rank Adaptation, allowing per-request LoRA adapter switching on a single base model. Launched in private preview in May 2026 for dedicated endpoints. Source: Cerebras – Introducing Multi Lora On Cerebras Inference
  • Multimodal / image input -- A model that accepts image inputs (screenshots, documents, charts) in addition to text. Gemma 4 31B is the first model on the Cerebras public API to support this. Source: Inference-Docs – Gemma 4 31B
  • Artificial Analysis Intelligence Index -- A third-party composite score across ten benchmarks used to compare model quality. Cerebras cites it for speed-vs-quality positioning against closed models. Source: Cerebras – Gemma 4 On Cerebras The Fastest Inference Is Now Multimodal

Latest Changes

Changes since the 2026-05 report.

  • New public model (coding-adjacent): Gemma 4 31B launched in public preview on June 29, 2026, the first Google DeepMind model and the first multimodal (image+text) model on the Cerebras public API. Artificial Analysis measured it at 1,851 output tokens/sec, described as 35x a typical GPU endpoint and 18x faster than Claude Haiku 4.5 (comparable quality: AA Intelligence Index 29 vs Haiku 4.5 at 30). Pricing $0.99 input / $1.49 output per MTok. Source: Cerebras – Gemma 4 On Cerebras The Fastest Inference Is Now Multimodal
  • First earnings as a public company: On June 23, 2026 Cerebras reported Q1 2026 revenue of $193.4 million (+92% YoY from $99.5M), net loss narrowed to $14M (from $23.9M / 46 cents per share a year earlier). Source: Cnbc – Cerebras Cbrs Q1 Earnings Report 2026
  • Margin guidance (financial risk): Core gross margin was 46.5% in Q1 but guided to 36-38% in Q2 and 38-41% for full-year 2026, partly due to a plan to rent back equipment from a large customer. CEO Andrew Feldman called the reaction "misunderstood." Stock fell ~10% after-hours June 23 and was down ~20% over the two-day window (closed $226.72 on June 23, down 28% from the $350 opening print). Source: Cnbc – Cerebras Cbrs Stock Earnings
  • Sovereign AI deal (India/UAE): A May 15 agreement (reported June 1) will deploy a 64-system Cerebras AI supercomputer in India, operated by G42's Core42 unit with data under Indian governance. This is the first country on G42's "Intelligence Grid" sovereign-AI network. Source: Restofworld – India Uae G42 Cerebras Ai Sovereignty
  • Customer concentration disclosed: G42 and MBZUAI together accounted for 86% of Cerebras' 2025 revenue per the SEC filing, flagged by Rest of World in June. Source: Restofworld – India Uae G42 Cerebras Ai Sovereignty
  • Competitive comparison blog: On June 5 Cerebras published a head-to-head of Kimi K2.6 on Cerebras vs Google's newly launched Gemini 3.5 Flash. Cerebras claims 981 tok/s vs 181 tok/s (5.4x), end-to-end 5.6s vs 17.5s, and 452ms time-to-first-token vs 960ms for Flash, the first frontier model below the 500ms real-time-voice bar. Source: Cerebras – Which Is Faster Gemini 3 5 Flash Or Kimi K2 6 On Cerebras
  • Cybersecurity partner blog: On June 15 Cerebras highlighted security use cases, naming Armis (AI for Security) and Operant AI (Security for AI) as partners using Cerebras inference. Source: Cerebras – Ai Inference Cybersecurity
  • Developer-guidance blogs: "The Economics of AI Reasoning" (June 17) argued that reasoning adds ~6x tokens and 7-11x latency for ~10-20% quality gain, recommending disabling reasoning for simple agentic steps; "Never Loop Without Verifiers" (June 24) demonstrated Gemma 4 visual loops producing STEP files in ~1.2s at ~1,500 tok/s. Sources: Cerebras – The Economics Of Ai Reasoning , Cerebras – Never Loop Without Verifiers
  • Plans still sold out (unchanged): Both Cerebras Code tiers (Pro $50/mo, Max $200/mo) remain listed as "sold out" as of June 30, 2026, with no reopening date announced. This is the second consecutive month of sold-out status. Source: Cerebras – Pricing
  • Lock-up expiry: Per the IPO prospectus, roughly 28 million Class A shares held by directors, officers, and non-employee shareholders became tradeable on the second trading day after the June 23 earnings, under a staggered (not single-date) lock-up schedule. Source: Cnbc – Cerebras Cbrs Stock Earnings

Plans

Inference API Access

Plan Price Rate Limits Key Inclusions
Free $0 5 RPM, 30K input tokens/min, 1M tokens/day (GPT OSS 120B / Gemma 4 / GLM 4.7) Access to all three shared public models, community support via Discord
Developer Self-serve, starting at $10 (pay-per-token) 10x higher limits than Free (per-model: see API Pricing table); no daily token cap disclosed Higher priority processing
Enterprise Undisclosed (contact sales) Highest limits (undisclosed), dedicated queue priority Custom model weights, fine-tuning/training services, dedicated support team with response-time guarantees, Multi-LoRA, dedicated endpoints (30+ models)

Terms explained:

  • RPM -- Requests per minute, the standard API rate-limit unit.
  • Self-serve starting at $10 -- Cerebras states the Developer tier starts at $10 but does not disclose what that buys in token credits or billing increments. Actual cost is usage-based per token. Source: Cerebras – Pricing

Cerebras Code (Subscription)

Plan Price Token Allowance Status
Pro $50/month Up to 24M tokens/day (~$48/day value) Sold out
Max $200/month Up to 120M tokens/day (~$240/day value) Sold out

Both plans remain unavailable for purchase (second consecutive month). Source: Cerebras – Pricing

API Pricing

Per-token pricing on the shared public endpoint. All prices are per million tokens.

Model Model ID Status Input ($/MTok) Output ($/MTok) Speed (tok/s) Context (paid) Max Output (paid) Modality
OpenAI GPT OSS 120B gpt-oss-120b Production $0.35 $0.75 ~3,000 131K 40K Text
Google Gemma 4 31B gemma-4-31b Preview (NEW) $0.99 $1.49 ~1,850 131K 40K Text + Image
Z.ai GLM 4.7 zai-glm-4.7 Preview $2.25 $2.75 ~1,000 131K 40K Text

Sources: Inference-Docs – Overview , Inference-Docs – Gemma 4 31B , Inference-Docs – Openai Oss , Inference-Docs – Zai Glm 47

Rate Limits by Tier

Model Tier Requests/min Input Tokens/min Daily Tokens
GPT OSS 120B Free 5 30K 1M
GPT OSS 120B Developer 1,000 1M N/A (undisclosed)
Gemma 4 31B Free 5 30K 1M
Gemma 4 31B Developer 300 500K N/A (undisclosed)
GLM 4.7 Free 5 30K 1M
GLM 4.7 Developer 500 500K N/A (undisclosed)

Sources: Inference-Docs – Gemma 4 31B , Inference-Docs – Overview

Dedicated Endpoint Models (Enterprise Only)

The dedicated-endpoint catalog spans 30+ models across 10+ families. Key coding-relevant families:

Model Family Key Models
Alibaba Qwen Qwen3-235B-A22B, Qwen3-Coder-480B-A35B, Qwen3-32B, Qwen3-30B-A3B
OpenAI OSS GPT-OSS-120B, GPT-OSS-20B
Moonshot AI Kimi-K2.6, Kimi-K2.5, Kimi-K2-Instruct, Kimi-K2-Thinking
Z.AI GLM-5.1, GLM-5, GLM-4.7, GLM-4.7-Flash, GLM-4.6
DeepSeek DeepSeek-V3.2, DeepSeek-V3.1, DeepSeek-V3
Meta Llama-4-Maverick (402B), Llama-4-Scout (109B), Llama-3.3-70B
Mistral Mistral-Large-3-675B, Devstral-Small-2-24B, Codestral-22B
MiniMax MiniMax-M2.5, MiniMax-M2.1
ByteDance Seed-OSS-36B
ServiceNow Apriel-1.6-15B-Thinker

Dedicated-endpoint pricing is not publicly listed. Source: Inference-Docs – Overview

Model Performance / Benchmarks

Cerebras publishes speed (verified by Artificial Analysis) but does not publish its own coding-quality benchmarks for served models; quality numbers below come from the model creators.

Model Metric Value Source
Gemma 4 31B Output speed (Artificial Analysis) 1,851 tok/s (35x typical GPU, 18x Haiku 4.5) Cerebras – Gemma 4 On Cerebras The Fastest Inference Is Now Multimodal
Gemma 4 31B AA Intelligence Index (quality) 29 (vs Haiku 4.5 at 30) Cerebras – Gemma 4 On Cerebras The Fastest Inference Is Now Multimodal
Gemma 4 31B Time-to-first-token (incl. reasoning) 1.5 seconds Cerebras – Gemma 4 On Cerebras The Fastest Inference Is Now Multimodal
Kimi K2.6 Output speed on Cerebras 981 tok/s Cerebras – Which Is Faster Gemini 3 5 Flash Or Kimi K2 6 On Cerebras
Kimi K2.6 Time-to-first-token (aiewf-eval voice benchmark) 452ms (first frontier model under 500ms with CoT on) Cerebras – Which Is Faster Gemini 3 5 Flash Or Kimi K2 6 On Cerebras
Kimi K2.6 SWE-Bench Pro (model-level) 58.6% (vs Gemini 3.5 Flash 55.1%) Cerebras – Which Is Faster Gemini 3 5 Flash Or Kimi K2 6 On Cerebras
Gemini 3.5 Flash Output speed (Artificial Analysis, reference) 181 tok/s Cerebras – Which Is Faster Gemini 3 5 Flash Or Kimi K2 6 On Cerebras
GPT OSS 120B Output speed (Cerebras measurement) ~3,000 tok/s Inference-Docs – Openai Oss
GLM 4.7 Output speed (Cerebras measurement) ~1,000 tok/s Inference-Docs – Zai Glm 47

Note: SWE-Bench Pro and AA Intelligence Index scores are model-level benchmarks published by model creators or Artificial Analysis. Cerebras contributes inference speed, not model capability. The blog comparison "Qwen-3.6-27B and Gemma-4-31B with reasoning, both beating last year's SOTA Sonnet-4" is a Cerebras-authored claim without a numeric benchmark citation. Source: Cerebras – The Economics Of Ai Reasoning

Latest News

Gemma 4 31B Launch, First Multimodal Model (June 29, 2026)

Gemma 4 31B launched in public preview as the first Google DeepMind model and the first multimodal (image + text) model on Cerebras. Artificial Analysis measured 1,851 tok/s, positioning it as 18x faster than Claude Haiku 4.5 at comparable quality (AA Intelligence Index 29 vs 30). Pricing is $0.99/$1.49 per MTok. Cerebras recommends it as the "reference medium-size model" and an open-weight Haiku/GPT-OSS/Llama alternative. Image inputs are base64 only (no URLs), max 5 images / 10MB per request, Chat Completions endpoint only. Source: Cerebras – Gemma 4 On Cerebras The Fastest Inference Is Now Multimodal

First Quarterly Earnings as a Public Company (June 23, 2026)

Q1 2026 revenue was $193.4 million, up 92% year-over-year from $99.5 million. Net loss narrowed to $14 million from $23.9 million a year earlier (loss per share 22 cents vs 46 cents). Core gross margin was 46.5% in Q1 but guided to 36-38% for Q2 and 38-41% for the full year, with core revenue guided to roughly $914 million (88% YoY growth) for Q2 and $855.5M-$865M (69% growth at midpoint) for the full year. The stock fell ~10% in extended trading and closed the next session at $226.72, down 28% from its $350 post-IPO opening print. Source: Cnbc – Cerebras Cbrs Q1 Earnings Report 2026

CEO Defends Margin Guidance (June 24, 2026)

CEO Andrew Feldman told CNBC the margin forecast was "misunderstood," stating Cerebras is beating the plan it laid out pre-IPO, and attributed part of the margin compression to a plan to rent back equipment from one of its largest clients. He also noted the staggered lock-up (about 28 million Class A shares becoming tradeable after earnings) and that Cerebras does not face the HBM or TSMC advanced-node supply pressure confronting Nvidia. Mizuho and Wedbush raised estimates following the call. Source: Cnbc – Cerebras Cbrs Stock Earnings

India/UAE Sovereign AI Supercomputer (reported June 1, 2026)

A May 15 agreement will deploy a 64-system Cerebras AI supercomputer in India, installed and operated by G42's Core42 unit with Cerebras providing technical support; all data remains under Indian governance. India is the first country on G42's "Intelligence Grid." Rest of World reported G42 and MBZUAI together accounted for 86% of Cerebras' 2025 revenue, and that G42 became Cerebras's largest customer in 2021. Source: Restofworld – India Uae G42 Cerebras Ai Sovereignty

Kimi K2.6 vs Gemini 3.5 Flash Comparison (June 5, 2026)

Ahead of the Gemini 3.5 Flash launch context, Cerebras published a head-to-head. On the Artificial Analysis 10K-input/500-output test, Kimi K2.6 on Cerebras completed in 5.6 seconds at 981 tok/s vs Gemini 3.5 Flash at 17.5 seconds and 181 tok/s. On the aiewf-eval voice benchmark Kimi K2.6 hit 452ms time-to-first-token vs Flash's 960ms. On SWE-Bench Pro, Kimi K2.6 scored 58.6% vs Gemini 3.5 Flash 55.1%. Source: Cerebras – Which Is Faster Gemini 3 5 Flash Or Kimi K2 6 On Cerebras

Economics of AI Reasoning (June 17, 2026)

A Cerebras-authored analysis argued reasoning produces ~6x more tokens and 7-11x longer completion times for a ~10-20% quality gain on coding/agentic benchmarks, and that 87.5%+ of tokens for a model like Qwen-3.6-27B are reasoning tokens. The recommendation is to treat reasoning as a cost/speed toggle and disable it for the ~50% of simple agentic steps. Source: Cerebras – The Economics Of Ai Reasoning

Never Loop Without Verifiers (June 24, 2026)

A practical post on building verified agent loops, demonstrating Gemma 4 producing a new STEP (CAD) file roughly every 1.2 seconds at ~1,500 tok/s, with the loop rewriting its own prompt across five iterations to clone a physical object from a photo. Source: Cerebras – Never Loop Without Verifiers

Cybersecurity Inference Edge (June 15, 2026)

A post positioning fast inference for security workflows, naming Armis (AI for Security: unified application-security context and remediation) and Operant AI (Security for AI: inline runtime defense for prompts, tool calls, and data flows) as Cerebras partners. Source: Cerebras – Ai Inference Cybersecurity

Community Signals

Gemma 4 launch thread: fast, but unclear who pays for a fast medium model

The HackerNews thread for the Gemma 4 launch reached 23 points and 8 comments. Commenters were impressed by the speed but questioned the value proposition of a fast, smaller, cloud-hosted model. Source: News – Item

  • tmanchester: "Okay this is actually pretty cool. Gemma 4 is a nice little model and I've really enjoyed playing around with it. At 1800 tok/s turns are essentially instant, it's a bit of a trip." News – Item
  • simianwords: "I just tried it on their website and it is extremely fast. I wonder what is the value prop of this? Where would I want 1. a smaller model 2. also non local, hosted on cloud. I can't think of any case." News – Item
  • anthonypasq: "speed is always better. if you have ever used a coding agent with 1000 tps going back to 50 seems like walking through sludge. for simple question i hate waiting 2 minutes for opus to loop 50 times just to read some files and answer a question." News – Item
  • simianwords (reply): "Of course I think speed is preferable but I don't see myself paying for a fast Gemma." News – Item
  • johntash: "OCR is a decent use-case for smaller models. I've had good experience using gemma for OCR'ing handwritten stuff that tesseract doesn't do so well on." News – Item

Earnings reaction: minimal developer-community engagement

The CNBC story on Cerebras's post-earnings stock drop ("Cerebras CEO says margin forecast misunderstood as stock plummets after earnings") drew only 8 points and zero comments on HackerNews, consistent with the May IPO submission (3 points, 0 comments). The developer community remains focused on product availability, not Cerebras's financials. Source: News – Item

Sovereign AI deal drew more attention than the IPO or earnings

The India/UAE sovereignty story ("India, UAE partner on AI sovereignty to bypass Google, Microsoft") reached 36 points and 12 comments, the highest-engagement Cerebras-adjacent HN submission of June. The framing around bypassing US cloud dominance and the 86% customer-concentration figure drove the discussion. Source: News – Item

Cerebras cited as the speed benchmark in a local-inference "Ask HN"

In an Ask HN about fast coding models on Apple Silicon (M4 Max, 128GB), the author used Cerebras as the reference point: "Cerebras offers gpt-oss-120b at over 1000t/s, but it's so expensive and also isn't able to properly call tools most of the time." This echoes the recurring May theme that Cerebras's speed advantage is real but model quality/tool-calling reliability on open-weight models trails Claude. Source: News – Item

Sold-out Cerebras Code subscriptions remain an unaddressed frustration

No June Cerebras blog post, press release, or pricing-page update announced a reopening of the Pro ($50/mo) or Max ($200/mo) Cerebras Code tiers. The earlier August 2025 launch thread (449 points, 172 comments) remains the locus of ongoing frustration, and the topic carries forward unresolved into June. Source: News – Item

Enterprise Readiness

Feature Available? Details
SSO (SAML/OIDC) Undisclosed Not mentioned in pricing, docs, or enterprise tier description. Contact sales.
SCIM Undisclosed Not mentioned in public documentation.
Audit logs Undisclosed Not mentioned. Dedicated-endpoint metrics are available in Prometheus format. Source: Inference-Docs – Metrics
IP indemnity No Cerebras is an inference provider, not a model creator; indemnity would depend on the underlying model. Not mentioned in Cerebras's own terms.
Data residency Partial Sovereign AI initiative supports on-premises deployment in specific countries (US, UAE, India); the June-reported India deal keeps data under Indian governance. Cloud inference data residency is otherwise undisclosed. Source: Restofworld – India Uae G42 Cerebras Ai Sovereignty
HIPAA Undisclosed Not mentioned in public documentation.
Air-gapped/on-prem Yes Cerebras CS-3 systems are sold as hardware for customer datacenters; the India/G42 deployment is a 64-system on-sovereign-soil example. Source: Cerebras – Ai Supercomputer
SLA Undisclosed Enterprise tier mentions "response time guarantees" for support, but inference SLA (uptime, latency) is not publicly documented.
Admin controls (RBAC) Partial Cloud Console supports Projects for organizing workloads and team access. Full RBAC details not publicly documented. Source: Inference-Docs – Projects

Transparency Gaps

Gap Details Severity
Cerebras Code plans still sold out (2 months) Both Pro ($50/mo) and Max ($200/mo) remain sold out as of June 30 with no reopening timeline. This blocks the primary flat-rate access point for individual developers for the second consecutive month. High
Customer concentration G42 and MBZUAI accounted for 86% of 2025 revenue (per SEC filing). The margin-guidance drop is partly tied to renting equipment back from a large customer, but which customer and the contract terms are not disclosed. High
Prompt caching pricing Prompt caching is listed as a supported capability (including on Gemma 4 31B), but no price discount for cached input tokens is published. Competitors (Anthropic, Google) make this explicit. High
Developer tier daily token cap Rate-limit tables list "N/A" for daily tokens on Developer tier for all three public models. Either there is no cap or it is not disclosed. Medium
Enterprise / dedicated-endpoint pricing No pricing, rate limits, or SLA terms are public for the Enterprise tier or the 30+ dedicated-endpoint models. Buyers must engage sales. High
Gross-margin mix detail The 46.5% to 36-38% Q2 margin drop is attributed partly to an equipment rent-back from a large customer and "data center pass-through revenues," but the breakdown between inference-API margin and hardware/systems margin is not separated in public disclosures. Medium
SSO/SCIM/audit logs/HIPAA None of these enterprise-compliance controls are mentioned in pricing, docs, or the enterprise tier description. High
Lock-up mechanics ~28M Class A shares became tradeable under a staggered schedule after June 23 earnings, but the full staged-release calendar is not summarized on the investor page. Low
Model quality benchmarks Cerebras publishes speed benchmarks but no independent coding-quality benchmarks (SWE-Bench, LiveCodeBench) for the models it serves. Quality claims rely on model creators or Artificial Analysis. Medium
Gemma 4 "preview" status Gemma 4 31B is labeled preview ("may be discontinued on short notice") with no production commitment or GA date. GLM 4.7 carries the same preview caveat. Medium

---

*Sources: All pricing and plan data from Cerebras – Pricing and Inference-Docs – Overview (accessed 2026-06-30). Financial data from CNBC Q1 earnings coverage (June 23-24, 2026). Sovereign-AI data from Rest of World (June 1, 2026). Blog data from Cerebras – Blog. Community signals from Hacker News.*