Cerebras

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 (third consecutive month). Source: Cerebras – Pricing
• Prompt Caching -- Stores and reuses previously processed prompt tokens to cut latency on repeated queries. Cerebras caching is automatic (128-token block matching, 5-minute guaranteed TTL). Critically, cached tokens are billed at the same rate as fresh tokens (no discount). Source: Inference-Docs – Prompt Caching
• Disaggregated Inference -- Splitting the two phases of LLM inference across different hardware: prefill (prompt processing, compute-heavy) on one engine and decode (token generation, memory-bandwidth-heavy) on another. The AMD+Cerebras partnership uses AMD Helios GPUs for prefill and Cerebras WSE for decode. Source: Cerebras – Amd And Cerebras Announce Industry Leading Ultra Low Latency And High Throughput Ai Inference
• Uncached vs. Total Tokens (dual-bucket rate limiting) -- New rate-limit model where each organization has an uncached TPM limit (primary constraint) and a total TPM limit set at 3x the uncached limit. Cached tokens count toward total but not uncached, so higher cache hit rates effectively increase throughput capacity. Rolling out gradually; visible to all by August 17, 2026. Source: Inference-Docs – Rate Limits
• 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: Artificialanalysis

Latest Changes

Changes since the 2026-06 report.

June Watch Items: Verification

• Cerebras Code subscriptions reopening: FAILED. Both Pro ($50/mo) and Max ($200/mo) remain listed as "sold out" as of July 31, 2026, with no reopening date announced. This is the third consecutive month of sold-out status (May, June, July). Source: Cerebras – Pricing
• Q2 earnings results: STILL PENDING. Cerebras reported Q1 2026 earnings on June 23, 2026. Q2 2026 earnings have not been released as of July 31, 2026. Based on the Q1 reporting cadence (roughly 6 weeks post-quarter-end), Q2 results are expected in mid-to-late August 2026. Source: Investors
• New model additions: PARTIAL. No new models were added to the shared public API in July (still 3: GPT OSS 120B, Gemma 4 31B, GLM 4.7). However, GLM 4.7 deprecation was announced for August 17, 2026, which will reduce the shared catalog to 2 models. The GPT-5.6 blog confirmed Cerebras serves GPT-5.6 Sol at 750 tok/s for OpenAI Codex subscribers, and the Upstage partnership added Solar 31B at ~2,000 tok/s for enterprise customers. Neither is on the shared public API. Sources: Inference-Docs – Overview , Cerebras – Getting The Most Out Of Gpt 5 6 Sol Terra And Luna , Cerebras – Cerebras And Upstage Bring Ultra Fast Ai To Korea
• Prompt caching pricing disclosure: CONFIRMED (unfavorable outcome). The prompt caching docs now explicitly state: "There is no additional fee for using prompt caching. Input tokens, whether served from the cache or processed fresh, are billed at the standard input token rate for the respective model." This means cached tokens receive zero price discount, unlike Anthropic (90% cheaper cached input) and OpenAI (90% cheaper cached input per their docs). A new dual-bucket rate limit system partially compensates by exempting cached tokens from the uncached TPM limit (total TPM is 3x uncached). Sources: Inference-Docs – Prompt Caching , Openai – Prompt Caching

New July 2026 Items

• AMD partnership (major infrastructure): On July 23, 2026, AMD and Cerebras announced a disaggregated inference solution combining AMD Helios GPU rackscale systems (for high-throughput prefill) with the Cerebras WSE (for ultra-low-latency decode). Targeted at up to 5x higher tokens/sec/watt compared to Cerebras WSE-only. Cerebras plans to deploy AMD Helios in its data centers, with availability through Cerebras Cloud in H2 2026. Announced at AMD's Advancing AI 2026 event. Dr. Lisa Su (AMD CEO) and Andrew Feldman (Cerebras CEO) both quoted. Source: Cerebras – Amd And Cerebras Announce Industry Leading Ultra Low Latency And High Throughput Ai Inference
• GLM 4.7 deprecation announced: The model catalog page now displays a deprecation notice: "Z.ai GLM 4.7 is scheduled for deprecation on August 17, 2026." This is one of only three shared-API models. No replacement model has been announced. After deprecation, the shared API will serve only GPT OSS 120B (production) and Gemma 4 31B (preview) unless a new model is added. Source: Inference-Docs – Overview
• GPT-5.6 Sol on Cerebras (via Codex): A July 27 blog post, "Getting the most out of GPT-5.6: Sol, Terra, and Luna," confirmed that GPT-5.6-Sol runs on Cerebras at 750 tokens/sec (up to 10x faster than regular mode per Artificial Analysis). This is available through OpenAI's Codex subscription, not the Cerebras shared API. The blog also referenced Kimi K2.7 Code and GLM-5.2 as external models usable in Codex via custom providers. Source: Cerebras – Getting The Most Out Of Gpt 5 6 Sol Terra And Luna
• Upstage/South Korea partnership: On July 10, Cerebras announced a collaboration with Upstage (Korean AI company). Upstage's Solar 31B model runs on Cerebras at up to 2,000 tokens/sec. Targeted at Korean enterprise markets (finance, insurance, healthcare, manufacturing). A demo showed Solar 31B processing 246 sources in a deep-research query, described as completing in seconds vs. 8 minutes with Sonnet 4.6. Source: Cerebras – Cerebras And Upstage Bring Ultra Fast Ai To Korea
• Gemma 4 app performance update: A July 8 "First Look" blog demonstrated Gemma 4 31B at ~2,300 tok/s in multimodal application demos (higher than the 1,850 tok/s official benchmark). Demonstrations included: a 60-page DeepSeek-V4 technical report analyzed in 1.79 seconds on Cerebras vs. ~25 seconds on GPU providers (17x speedup); a rental car damage inspector processing 17 video frames with structured JSON bounding boxes in 5 seconds total (~300ms per 3-image batch). Source: Cerebras – First Look Gemma 4 On Cerebras 3 Fast Multimodal Apps We Built
• Free trial structure changed: The Free Trial tier now provides "$5 in free credits after making an account" rather than the previous token-based allowance. Per the rate-limits FAQ: "New accounts receive $5 in free credits after adding a verified payment method. These credits expire 30 days after they're granted." If no payment method is added, Playground and API access remain inactive. Source: Cerebras – Pricing , Inference-Docs – Rate Limits
• Dual-bucket rate limiting (new system): Rate limits now use two independent buckets per organization: uncached TPM (primary constraint) and total TPM (set at 3x uncached). Cached tokens count toward total but not uncached. Rolling out gradually; visible to all organizations by August 17, 2026. Example: with 1M uncached TPM limit and 70% cache hit rate, effective total throughput is up to 3M TPM. Source: Inference-Docs – Rate Limits
• prompt_cache_key parameter: New optional routing hint that tells the system which requests share a common prompt prefix, improving cache hit rates for multi-turn conversations and agentic workflows. Requires account enablement (contact sales). Does not affect billing. Source: Inference-Docs – Prompt Caching
• Prompt caching supported models clarification: The prompt caching FAQ lists zai-glm-4.7 and gpt-oss-120b as supported. Gemma 4 31B lists "Prompt Caching" as a capability on its model page, but it is absent from the FAQ's supported-models list (discrepancy). Source: Inference-Docs – Prompt Caching , Inference-Docs – Gemma 4 31B
• Developer-guidance blogs: "How We Built Our Knowledge Base" (July 15) and "Hiring Engineers for an AI-Native World" (July 9) were internal/culture posts with no product or pricing impact. Source: Cerebras

Plans

Inference API Access

Terms explained:

• RPM -- Requests per minute, the standard API rate-limit unit.
• $5 in free credits -- New accounts must add a verified payment method to receive the $5 credit. Credits expire 30 days after being granted. If no payment method is added at sign-up, Playground and API access remain inactive. Source: Inference-Docs – Rate Limits

Cerebras Code (Subscription)

Both plans remain unavailable for purchase (third consecutive month: May, June, July). Source: Cerebras – Pricing

API Pricing

Per-token pricing on the shared public endpoint. All prices are per million tokens. No price changes from June 2026.

Sources: Inference-Docs – Openai Oss , Inference-Docs – Gemma 4 31B , Inference-Docs – Overview , Inference-Docs – Prompt Caching

Note on prompt caching pricing: Unlike Anthropic (cached input at 10% of standard rate, i.e., 90% cheaper) and OpenAI (cached input at 10% of standard rate), Cerebras charges the full input rate for cached tokens. The benefit is rate-limit headroom (cached tokens don't count toward uncached TPM) and lower latency, but not lower cost. Sources: Inference-Docs – Prompt Caching , Openai – Prompt Caching

Rate Limits by Tier

Sources: Inference-Docs – Rate Limits , Inference-Docs – Openai Oss , Inference-Docs – Gemma 4 31B

Gemma 4 31B image limits: Free Trial allows 2 images/request (4 MB payload). Developer allows 10 images/request (10 MB payload). Source: Inference-Docs – Gemma 4 31B

Dedicated Endpoint Models (Enterprise Only)

The dedicated-endpoint catalog spans 30+ models across 10+ families. Key coding-relevant families (unchanged from June):

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.

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 AMD 5x T/s/W figure is based on AMD Performance Labs and Cerebras modeling in July 2026 using Kimi K2.6 1T, not independently verified. Source: Cerebras – Amd And Cerebras Announce Industry Leading Ultra Low Latency And High Throughput Ai Inference

Latest News

AMD and Cerebras Announce Disaggregated Inference Partnership (July 23, 2026)

AMD and Cerebras announced a technical partnership to combine AMD Helios GPU rackscale solutions with the Cerebras Wafer-Scale Engine in a single disaggregated inference workflow. AMD Helios handles high-throughput prefill (prompt processing); the Cerebras WSE handles ultra-low-latency decode (token generation). The joint solution is targeted at up to 5x higher tokens/sec/watt compared to a Cerebras WSE-only configuration (based on July 2026 modeling with Kimi K2.6 1T). Cerebras plans to deploy AMD Helios in its data centers, with availability through Cerebras Cloud in H2 2026. Announced at AMD's Advancing AI 2026 event, with quotes from Dr. Lisa Su (AMD CEO) and Andrew Feldman (Cerebras CEO). Source: Cerebras – Amd And Cerebras Announce Industry Leading Ultra Low Latency And High Throughput Ai Inference

GLM 4.7 Deprecation Announced (July 2026)

The Cerebras model catalog page now displays a deprecation notice for Z.ai GLM 4.7, scheduled for August 17, 2026. GLM 4.7 is one of only three models on the shared public API (alongside GPT OSS 120B and Gemma 4 31B). No replacement model has been announced. After deprecation, the shared API will serve two models unless a new one is added before or by that date. This follows a pattern noted in a January 2026 "Tell HN" post where an enterprise customer reported Cerebras frequently terminates and replaces models, leaving subscribers without alternatives. Sources: Inference-Docs – Overview , News – Item

GPT-5.6 Sol on Cerebras via OpenAI Codex (July 27, 2026)

A Cerebras blog post, "Getting the most out of GPT-5.6: Sol, Terra, and Luna," provided guidance on using the three GPT-5.6 variants within OpenAI's Codex subscription. The post confirmed GPT-5.6-Sol runs on Cerebras at 750 tokens/sec, described as "up to 10x faster speed advantage than regular mode of Sol" per Artificial Analysis. This access is through the Codex subscription product, not the Cerebras shared API. The blog also discussed reasoning levels (Light through Ultra, with each step up in Sol increasing average cost per task by ~50% per Artificial Analysis July 17 testing), prompt cache TTL (~30 minutes for GPT-5.6 models, with cached input 90% cheaper than fresh on OpenAI's pricing), and multi-agent workflows using external models like Kimi K2.7 Code and GLM-5.2. Source: Cerebras – Getting The Most Out Of Gpt 5 6 Sol Terra And Luna

Cerebras and Upstage Partnership for South Korea (July 10, 2026)

Cerebras announced a collaboration with Upstage, a Korean AI company. Upstage's Solar 31B model runs on Cerebras Inference Cloud at up to 2,000 tokens/sec. Targeted at Korean enterprise markets across finance, insurance, healthcare, and manufacturing. A demo showed Solar 31B processing 246 sources in a deep-research query in seconds, compared to 8 minutes with Sonnet 4.6 on a comparable setup. Source: Cerebras – Cerebras And Upstage Bring Ultra Fast Ai To Korea

Gemma 4 Multimodal App Demos (July 8, 2026)

A "First Look" blog demonstrated three multimodal applications built with Gemma 4 31B on Cerebras. Key results: (1) Document analysis of the 60-page DeepSeek-V4 technical report completed in 1.79 seconds on Cerebras vs. ~25 seconds on GPU providers (17x speedup); (2) An image search agent scanning 80 images with plain-language queries; (3) A rental car damage inspector processing 17 video frames into structured JSON with bounding boxes in 5 seconds total (300ms per 3-image batch). The post also shared 7 practical tips for fast multimodal apps (use official chat templates, enable thinking mode intentionally, place images before text, etc.). Source: Cerebras – First Look Gemma 4 On Cerebras 3 Fast Multimodal Apps We Built

Prompt Caching Pricing Disclosed (July 2026, docs update)

The prompt caching documentation was updated to explicitly state the pricing model: "There is no additional fee for using prompt caching. Input tokens, whether served from the cache or processed fresh, are billed at the standard input token rate for the respective model." This confirms zero price discount for cached tokens. The rate-limits documentation was simultaneously updated with a new dual-bucket system: uncached TPM (primary constraint) and total TPM (3x uncached). Cached tokens count toward total but not uncached. The prompt_cache_key parameter was also documented as a new optional routing hint. Prompt caching is confirmed supported on gpt-oss-120b and zai-glm-4.7. Sources: Inference-Docs – Prompt Caching , Inference-Docs – Rate Limits

Community Signals

AMD partnership thread: cautious interest, architecture questions

The AMD and Cerebras partnership announcement reached 27 points and 9 comments on HackerNews. The discussion focused on understanding the disaggregated architecture and skepticism about timing.

• ckrapu: "They're a bit late to the party." News – Item
• techgnosis: "I'm not sure I understand. It's Helios but with WSE attached? And it uses one or the other depending on some criteria?" News – Item
• wtallis (reply explaining disaggregation): "LLM inference is a two-phase process. The first phase is prompt processing aka prefill. It's compute-heavy but requires relatively low memory bandwidth. The second phase is token generation aka decode, which doesn't require much in the way of FLOPs but wants as much memory bandwidth as possible. This announcement is for a system to do the first phase on Helios and the second phase on WSE." News – Item
• sccvcxv: "lol when is the music gonna stop?" News – Item
• JSR_FDED: "It's disaggregated so you know it's good" (sarcasm). News – Item
• rbanffy (reply to JSR_FDED): "The two major sides of the system excel at different things that happen to be complementary in AI inference." News – Item

Source: News – Item

Sold-out Cerebras Code subscriptions remain unaddressed

No July 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 original August 2025 launch thread (449 points, 172 comments) was last updated July 2, 2026. The topic carries forward unresolved into a third month. Source: News – Item

Model deprecation pattern flagged by enterprise customer (context from January, still relevant)

A January 2026 "Tell HN: Avoid Cerebras if you are a founder" post (35 points, 14 comments) described an enterprise customer being forced off the platform when their subscribed model (Llama 3.3 70B) was terminated with no alternative plan offered. The poster stated Cerebras support staff told them to migrate to Groq. This pattern is directly relevant to the GLM 4.7 deprecation announced this month. Source: News – Item

Cerebras cited as the speed benchmark in local-inference "Ask HN" (updated July 29)

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." The thread was updated as recently as July 29, 2026, confirming the recurring community view that Cerebras speed is real but tool-calling reliability on open-weight models trails Claude. Source: News – Item

Enterprise Readiness

Transparency Gaps

---

*Sources: All pricing and plan data from Cerebras – Pricing and Inference-Docs – Overview (accessed 2026-07-31). Rate limit and prompt caching data from Inference-Docs – Rate Limits and Inference-Docs – Prompt Caching. AMD partnership from Cerebras – Amd And Cerebras Announce Industry Leading Ultra Low Latency And High Throughput Ai Inference (July 23, 2026). Blog data from Cerebras – Blog. Community signals from Hacker News.*