inference on tomrochette.com

Cerebras Code

tom@tomrochette.com (Tom Rochette) — Sat, 26 Sep 2026 00:00:00 +0000

Cerebras Code is a subscription from chipmaker Cerebras that sells fast inference on one open coding model at $50/month (Pro) and $200/month (Max). The pitch is speed, and the record shows the speed claim and the quota fine print are the two things to verify before paying.

What it is
#

A hosted coding-inference plan on Cerebras wafer-scale hardware, consumed by pointing any OpenAI-compatible editor or agent (Cline, OpenCode, Crush, Cursor) at a Cerebras API key. It launched August 1, 2025 with Qwen3-Coder-480B advertised at up to 2,000 tokens per second and a 131k context window. As of 2026-09-26 the product page promotes GLM 4.7 at “1,000 tokens+ per second”, so the headline model has already been swapped once.

Status
#

Launched August 1, 2025 and drew 449 points and 172 comments on Hacker News the same day. Launch windows sold out repeatedly, and as of 2026-09-26 both Pro and Max are marked “sold out” on cerebras.ai/code, with a limited free trial still open. The model changed from Qwen3-Coder to GLM 4.7 between launch and now, which shows the plan follows whichever open model is fastest rather than committing to one family. I could not verify funding or subscriber counts.

Strengths
#

Speed is genuinely differentiated: even its harshest reviewer calls Cerebras the fastest provider of its model, bar none.
Flat monthly pricing with published daily allowances (24M tokens Pro, 120M Max) instead of per-token anxiety.
Bring-your-own-editor stance with OpenAI-compatible endpoints, no proprietary IDE lock-in.
By October 2025 the original tokens-per-minute caps had been raised in response to criticism.

Cautions
#

The marketing number did not survive contact: InfoWorld measured well under 500 tokens/second and often under 100, against the “up to 2,000” claim.
Undocumented throttles drove the experience: 300k TPM on Pro and 400k on Max produced 429 errors mid-session, and an early buyer reported a 7.5M-token daily cap hidden behind an advertised 1,000-request limit.
131k context is about half the model’s native window and demands careful context management.
At launch there was no prompt caching, which made agent loops expensive at the $2/1M API rate, and the usage console had defects (a Max purchase provisioning as Pro).
Cerebras declined InfoWorld’s request for comment on these issues.

Pricing
#

Pro costs $50/month with up to 24M tokens/day, and Max costs $200/month with up to 120M tokens/day. Both plans were marked sold out as of 2026-09-26; a free tier with limited tokens remains for connection testing. The underlying API price at launch was $2 per 1M input and $2 per 1M output on Qwen3-Coder. The current per-token table on cerebras.ai/pricing renders client-side and I could not extract it, so treat current API rates as unverified.

Price history
#

Date	Plan	Change	Source
2025-08-01	Pro / Max	Launched at $50/month (24M tokens/day) and $200/month (120M tokens/day)	Cerebras blog and HN thread
2025-09-15	Pro / Max	TPM caps (300k/400k) documented as the binding limit; prompt caching promised	InfoWorld review
2025-10-28	Pro / Max	Caps reported improved; Qwen3 deprecated in favor of GLM-4.6 from November	InfoWorld follow-up
2026-09-26	Pro / Max	Prices unchanged at $50/$200 but both marked sold out; model now GLM 4.7	cerebras.ai/code

Compared to
#

- Chutes is the cheap multi-model pay-as-you-go option; choose Chutes for price and variety, Cerebras when seconds per response matter.
- GLM Coding Plan undercuts Cerebras on quota per dollar and now serves the same GLM family; choose it for volume, Cerebras for raw tokens per second.

Bottom line
#

Recommended for developers whose bottleneck is iteration latency and who code in bursts that fit the daily token allowance. Not for heavy all-day agent runs, where the TPM throttles and 131k context cut the advertised advantage. Claim to disagree with: at the documented throttles, Max at $200 was worse value than four Pro accounts, because 4x300k TPM beats 400k TPM.

Changes
#

2026-09-26 - Created.

References
#

https://www.cerebras.ai/blog/introducing-cerebras-code - launch post: $50/$200 tiers, 24M/120M tokens/day, 2,000 tok/s claim, 131k context (fetched, HTTP 200).
https://www.cerebras.ai/blog/qwen3-coder-480b-is-live-on-cerebras - Qwen3-Coder launch, $2/1M API rate, “20x higher coding speed” claim (fetched, HTTP 200).
https://www.cerebras.ai/code - current product page: GLM 4.7 at 1,000+ tok/s, Pro and Max both marked “sold out” (fetched, HTTP 200, as of 2026-09-26).
https://news.ycombinator.com/item?id=44762959 - launch-day reception: 449 points, 172 comments, top comment flags missing caching (fetched, HTTP 200).
https://www.infoworld.com/article/4055909/ - critical review, Sep 15, 2025: disputed tok/s claims, TPM caps, 131k context, billing mixups, no vendor comment (fetched, HTTP 200).
https://www.infoworld.com/article/4075825/how-to-vibe-code-for-free-or-almost-free.html - follow-up: caps improved, Qwen3 deprecated for GLM-4.6, “fastest bar none” (fetched, HTTP 200).
https://www.reddit.com/r/LocalLLaMA/comments/1mfeazc/ - critical post, Aug 2, 2025: advertised 1,000 requests/day actually a 7.5M-token cap, “request” defined at ~8k tokens (retrieved, HTTP 200, via the Arctic Shift archive API).

Chutes

tom@tomrochette.com (Tom Rochette) — Sat, 26 Sep 2026 00:00:00 +0000

Chutes is a model-access platform built on Bittensor (subnet 64) that sells per-token inference on open-weight models, optional monthly plans, and private GPU deployments. Its per-token prices were the lowest I verified in this category, and price is the main reason to pick it.

What it is
#

Chutes serves open-weight models (Qwen, DeepSeek, Kimi, GLM, Mistral, Gemma) from hosted instances and charges by the million tokens, with no subscription required. Featured models run on confidential TEE compute, and billing accepts USD or TAO from a Bittensor wallet. A CLI can also deploy a private chute on self-serve GPU capacity billed per second, with a one-time deployment fee scaled to the GPU choice.

Status
#

Active and shipping changes as of September 2026, with a live pricing page, status page, and regular community announcements. 2026 was a repair year for the business model: on February 27 the team retired the free 200-requests-per-day Early Access perk, capped subscriptions at 5x pay-as-you-go value, and cut frontier models from the $3 Base tier after its tables showed top users extracting up to 324x their subscription price. A March 20 update reports tokens served down 45% over six weeks while revenue per million tokens rose 37.7% since February 1, after ending roughly 10B tokens/day of free-quota serving to about 11,000 users. I could not verify funding or traffic figures.

Strengths
#

Lowest per-token list prices I found this run, for example DeepSeek-V3.2 at $1.00/$1.00 per 1M in/out and Mistral-Nemo at $0.0245/$0.0978.
Many model families behind one OpenAI-compatible key, so switching models costs nothing.
Pay-as-you-go works with no plan attached, keeping commitment near zero.
TEE enclaves on featured models, a privacy posture most cheap providers lack.

Cautions
#

Terms changed three times in early 2026 (free tier removed, 5x cap, Base tier model cuts), so treat any quota as provisional.
InfoWorld’s reviewer found performance underwhelming, called the privacy policy ambiguous, and needed a VPN workaround to sign up.
The March 2026 update cites a global GPU shortage limiting inventory, so capacity depends on miner participation.
The 6%/10% plan discounts apply only beyond the bundled daily quota, so heavy users are pushed back to full PAYG rates.

Pricing
#

PAYG is the default as of 2026-09-26: GLM-5.2 $1.25 in / $3.95 out per 1M, Kimi-K3 $3.00/$15.00, Qwen3-235B-A22B-Thinking $0.2989/$1.1957, DeepSeek-V3.2 $1.00/$1.00. Plus costs $10/month for a bundled daily quota plus 6% off PAYG beyond it, and Pro costs $20/month for a larger quota plus 10% off. Since February 27, 2026, every subscription is capped at 5x the equivalent PAYG value, with overflow billing at standard PAYG. Private chutes run on an RTX Pro 6000 at $1.80/hour plus a $5.40 one-time deployment fee (3x the hourly rate).

Price history
#

Date	Plan	Change	Source
2025-07-31	Subscriptions	Pricing tiers announced, “coming next Monday” (live early August 2025)	Chutes news post, July 31, 2025
2026-02	Base / Plus / Pro	Baseline: $3 / $10 / $20 per month tiers existed	Feb 27, 2026 community announcement usage tables
2026-02-27	Early Access	Free 200-requests/day perk cut off from TEE models, fully retired March 15, 2026	same announcement
2026-02-27	All subscriptions	Value capped at 5x PAYG equivalent; GLM-5, Kimi K2.5, Qwen 3.5, MiniMax M2.5 removed from Base	same announcement
2026-09-26	Base	No longer listed on the pricing page, which shows only Plus $10, Pro $20, and Enterprise	chutes.ai/pricing

Compared to
#

- Cerebras Code sells raw speed on one coding model at $50/$200; choose it when latency dominates, Chutes when price dominates.
- GLM Coding Plan sells a large fixed quota on one model family; choose it if you will burn the quota, Chutes for variety without commitment.

Bottom line
#

Recommended for hobbyists and cost-driven developers who want many open models cheaply and can tolerate churn in terms and capacity. Not for teams needing SLAs, contractual stability, or closed frontier models. Claim to disagree with: after the 5x cap, Plus and Pro make little sense for coders, because PAYG with a small prepaid balance delivers the same upside without a recurring bill.

Changes
#

2026-09-26 - Created.

References
#

https://chutes.ai/pricing - current PAYG model rates, Plus $10 / Pro $20 with 6%/10% PAYG discounts, private GPU pricing (fetched, HTTP 200, as of 2026-09-26).
https://chutes.ai/news/community-announcement-february - February 27, 2026 changes: Early Access retirement, 5x subscription cap, Base tier model removals, abuse tables (fetched, HTTP 200).
https://chutes.ai/news/from-volume-to-value-building-a-sustainable-ai-inference-platform-2 - March 20, 2026 economics: tokens down 45%, revenue per token up 37.7%, free-tier costs (fetched, HTTP 200).
https://chutes.ai/news/coming-soon - July 31, 2025 post announcing the first pricing tiers (fetched, HTTP 200).
https://chutes.ai/docs/cli/deploy - private chute deployment mechanics and deployment fee structure (fetched, HTTP 200).
https://www.infoworld.com/article/4075825/how-to-vibe-code-for-free-or-almost-free.html - critical third-party review: underwhelming performance, ambiguous privacy policy, signup bugs; confirms the $3 entry tier in October 2025 (fetched, HTTP 200).

inference on tomrochette.com

Cerebras Code

What it is
#

Status
#

Strengths
#

Cautions
#

Pricing
#

Price history
#

Compared to
#

Bottom line
#

Changes
#

See also
#

References
#

Chutes

What it is
#

Status
#

Strengths
#

Cautions
#

Pricing
#

Price history
#

Compared to
#

Bottom line
#

Changes
#

See also
#

References
#

inference on tomrochette.com

Cerebras Code

What it is #

Status #

Strengths #

Cautions #

Pricing #

Price history #

Compared to #

Bottom line #

Changes #

See also #

References #

Chutes

What it is #

Status #

Strengths #

Cautions #

Pricing #

Price history #

Compared to #

Bottom line #

Changes #

See also #

References #

What it is
#

Status
#

Strengths
#

Cautions
#

Pricing
#

Price history
#

Compared to
#

Bottom line
#

Changes
#

See also
#

References
#

What it is
#

Status
#

Strengths
#

Cautions
#

Pricing
#

Price history
#

Compared to
#

Bottom line
#

Changes
#

See also
#

References
#