Charm Hyper coding plans

Best for

5 events in the last 90 days — 1 worse, 1 better.

• Ceiling20 Oct 2026

  Charm Hyper switched the $20 plan’s credit cadence from daily to weekly. The docs’ "Hypercredits Available and Refresh" section now reads, verbatim: "Subscription Hypercredits refresh every week on Sunday at 00:00 UTC, and the five-hour rolling window means part of your weekly allocation becomes available again over time. One Hypercredit equals five cents of inference." The credit grant itself is unchanged (250 Hypercredits on both reads), so the same grant now covers a week instead of a day — a sevenfold tightening of the period the subscriber gets it in — and the plan gains its first published sub-window limits (150 Hypercredits/5h and 3 concurrent requests), which are shared credit ceilings on the plan, not per-model caps. The calculator’s monthly-equivalent budget for every model on the plan moves $375 (30 × $12.50/day) → $54.17 (52/12 × $12.50/week) with it.

  Verified 2026-10-20 · source

• Launch20 Oct 2026

  Charm Hyper restructured its personal line from one plan into five Shred tiers, so the catalogue tracks four new records (hyper-shred-diet, hyper-shred-energy, hyper-shred-nitro, hyper-shred-animal) in ascending price order around the retained hyper-dev id. Every tier is billed monthly but credited weekly at the published 1 Hypercredit = 5¢ — Diet ≈$9.00/week through a 100-Hypercredit/5h window with 2 concurrent requests, Classic ≈$12.50/week through 150/5h with 3, Energy ≈$20.00/week through 250/5h with 4, Nitro ≈$34.50/week through 400/5h with 8, Animal ≈$60.00/week through 700/5h with 12 — and each card lists "All available models" plus "Prism routing model". Out of scope, unchanged: the Team tiers (Business Casual, Business Class, Business Platinum, Diamond, Champagne, $30–$220/mo/user — Team is not tracked) and pay-as-you-go (no subscription), as are the free tier and prepaid bundles.

  Verified 2026-10-20 · source

• Price20 Oct 2026

  One merged event for three per-1M rate moves read off the provider’s own /v1/models list (mixed directions, so neutral). DeepSeek V4.1 Flash’s row rose $0.30/$1.20 → $0.33/$1.31 input/output (cached read $0.03 unchanged). The plain deepseek-v4-pro and deepseek-v4-flash rows vanished, leaving only their dated checkpoints, so DeepSeek V4 Pro re-points at deepseek-v4-pro-0813 ($2.40/$4.80/$0.20 → $1.437216/$4.311648/$0.047907, roughly 40% cheaper input) and DeepSeek V4 Flash at deepseek-v4-flash-0731 ($0.20/$0.40/$0.04 → $0.44/$1.32/$0.044, more than double) — each dated row the version label of the same canonical model, not a separate entry, and the old plain-row rates no longer exist on any row. The dated deepseek-v4.1-flash-1017 checkpoint duplicates the plain id’s rates and stays excluded as a label.

  Verified 2026-10-20 · source

• Models20 Oct 2026

  Inkling joins every Charm Hyper Shred roster. The model registry already carries Inkling (several other tracked plans list it), so the earlier "identity unconfirmed" exclusion no longer applies and the stored rate is the row’s own ($1.0888 input / $4.40964 output / $0.185096 cached read per 1M). The "+5 more" set becomes "+4 more": deepseek-v4.1-flash-1017 is a dated label of the plain V4.1 Flash row (identical rates) and stays excluded; prism stays out because it is a router, not a model — the landing page reads "Prism is a decision model, with a full 8B parameter backbone, which reads the prompt like an LLM … routes them to the most appropriate Hyper model for the job" and both plan surfaces list it as a plan feature ("Prism routing model") rather than a selectable model; kimi-k2-thinking and gemma-4-26b-a4b-it stay out because the model registry has no entry for either.

  Verified 2026-10-20 · source

• Launch28 Sept 2026

  Charm Hyper joins the catalogue as a coding-focused inference provider from Charmbracelet: the $20/month Dev plan carries 250 Hypercredits that refresh every 24 hours, so a subscriber gets about $12.50 of inference a day — 1 Hypercredit is published as 5¢ on both the FAQ and the models docs, the landing pricing card reads “250 Hypercredits, refreshing daily”, and the docs’ Plans & Billing page prints “250 Hypercredits / 24 hours”. The docs prose still describes plan credits as a monthly allocation, so the two concrete product surfaces decide the cadence and the daily figure is the one stored. The free tier (100 Hypercredits a month) and the prepaid credit bundles are separate purchases and are not tracked. Access is an OpenAI-compatible (chat completions and responses) and Anthropic Messages API with plain API-key auth and setup pages for Crush, Claude Code, Codex, OpenCode, Pi and Zed, so the plan is catalogued as usable from any tool with no client lock. The 18 stored models come from the provider’s own public model list, each at its own published per-1M rate, and the calculator budgets every one of them at $375/month — the monthly equivalent of spending the full daily allowance for 30 days, named as an estimate because the credits refresh on their cadence rather than banking. Five catalogue entries are not stored: the dated DeepSeek checkpoint labels (0731 and 0813) are version labels of the plain V4 Flash and V4 Pro models rather than separate canonical models, the site’s model registry has no entry for Kimi K2 Thinking or Gemma 4 26B A4B, and Inkling is left out until its identity can be confirmed from a first-party source.

  Verified 2026-09-28 · source

Head-to-head

Charm Hyper compared side by side

Two plans side by side — shared models, BYOK reachability and value per $1, generated from the same verified data as the calculator.

• CommandCodevsCharm Hyper

Flagship rank · Kimi K3: #25 (Max config) on the Agent Arena agent leaderboard — 0.57% net improvement (board fetched 2026-11-09).

Common questions