Tony Nguyen · research note · 2026-08-10
Same model, ten times the price.
What does it cost to buy a model’s intelligence via API, and what does it cost via a subscription? This note compares both across providers on a single metric: cost per standardized task (AA Intelligence Index) against the intelligence index. The short answer: GLM-5.2 through the GLM Coding Plan off-peak costs $0.035 per task — the cheapest path to GLM-5.2 intelligence, cheaper than the Z.AI API ($0.30). And DeepSeek V4 Flash delivers nearly the same intelligence (II 52 vs 53) at 6–20x less.
Decision tool
Model parity across providers — cost per task × intelligence
Top-left = the most intelligence for the least money. GLM-5.2 via the GLM plan off-peak ($0.035/task) is the cheapest path to GLM-5.2 intelligence — cheaper than the Z.AI API ($0.30).
The table is wide — scroll it sideways; the first column stays put.
| Provider / plan | Recurring price / mo | Quota unit / baseline | Reset / window | API status | Availability | Confidence / note | Off-peak discount |
|---|---|---|---|---|---|---|---|
| Mainland China and global token plans | |||||||
| Baidu QianfanToken Plan Personal | ≈ $1.46–88.63 (¥9.9–600; promo ¥4.9–299.9 ≈ $0.72–44.30) | tokens / month (shared across models) | monthly | subscription · OpenAI + Anthropic-compatible endpoint · dedicated key · interactive tools only · Hermes Agent (official integration) | mainland China — account with a mainland number and identity verification | 5-off promo is checkout-only — likely still running (3P, no official end date) | 5-off promo (50%) — checkout-only, no official end date |
| VolcanoArk Coding Plan | Lite ¥9.9 / Pro ¥49.9 promo (2.5x off, first 2 months) · regular ¥40 / ¥200 · campaign until 2026-08-27 | requests — Lite ≈ 1,200 / 5h, 9,000 / week, 18,000 / month (Pro 5x) | rolling 5h window from first request · weekly reset Monday 00:00 · monthly reset on day 1 of the subscription month | subscription · /api/coding (Anthropic) and /api/coding/v3 (OpenAI) endpoints · Codex, Claude Code, OpenCode, OpenClaw, Hermes Agent, TRAE, OpenViking | mainland China | GLM-5.2 4x benefit is contradictory (docs say it ended 08-08, the campaign still advertises it) — verify in console | — (2.5x-off campaign for the first 2 months — not off-peak) |
| ZhipuGLM Coding Plan | CN ¥118 / 538 / 1,078 · international z.ai $18 / 80 / 168 (yearly −30%, quarterly −20%) | credits — 2,000 / 12,000 / 28,000 per 5h; weekly 10,000 / 60,000 / 140,000 · GLM-5.2 coef. 6.9/1.7/24 | 5h refresh 5 hours after consumption (dynamic) + 7-day cycle from order date | subscription · Anthropic and OpenAI endpoints · tool whitelist (Claude Code, OpenCode, Cline, Cursor; Hermes = general agent, secondary) | mainland China + international z.ai (region-specific keys) | peak Mon–Fri 14:00–18:00 UTC+8 = 1x, off-peak = 0.5x credits · GLM-5.3 unconfirmed (leak only) | off-peak 0.5x credits — outside Mon–Fri 14:00–18:00 UTC+8 |
| XiaomiMiMo Token Plan | $6 / 16 / 50 / 100 / mo (Lite–Max; ¥39 / 99 / 329 / 659) · yearly −12% · first-purchase −12% | credits (4.1–82B) · official per-1M coefficients — not raw tokens | monthly · 0.8x coefficient 00:00–08:00 Beijing (= UTC 16:00–24:00) | subscription · OpenAI/Anthropic-compatible (token-plan-cn/-sgp/-ams) · Claude Code, OpenClaw, OpenCode, Kilo Code, Cline, Hermes Agent, CodeBuddy | mainland + Singapore and Europe endpoints | foreign payment via Waffo/Stripe · official per-1M coefficients exist — credits are still not raw tokens | 0.8x coefficient 00:00–08:00 Beijing (= UTC 16:00–24:00) |
| TencentToken Plan Personal Edition | $7–103 / mo (1,000–15,900 credits) | token-weighted credits · official per-1M coefficients — not raw tokens | subscription month · no rollover | subscription · sk-tp-… key · /plan/v3 (OpenAI) and /plan/anthropic endpoints · Hermes, OpenCode, Claude Code, Cursor, Cline, Codex CLI | Singapore region (international, English docs) | interactive tools only — automation banned · MiniMax-M3 dual coefficient (≤512k 47/188/10, >512k 94/375/19) · DeepSeek V4 = Vendor Direct, no SLA | — |
| KimiKimi Code | international $19–199 / mo (Moderato $19 · Allegretto $39 · Allegro $99 · Vivace $199; yearly $15–159) · CN ¥49–699 (Andante–Allegro) | 1x–30x multiplier · rolling 5-hour window ≈ 300–1,200 requests, 30 concurrent | quota refreshes every 7 days | subscription · api.kimi.com/coding/v1 (OpenAI) and api.kimi.com/coding/ (Anthropic) endpoints · official third-party: OpenCode “Kimi For Coding”, Claude Code, Codex · max 5 keys · Extra Usage with a spending cap | international + CN — region-dependent pricing | region-dependent — verify at checkout | — |
| MiniMaxToken Plan | CN ¥49 / 119 / 469 · international $20 / 50 / 120 (platform.minimax.io) | 5-hour + weekly windows · token-based metering since M3 | 5h + weekly | subscription · sk-cp-… key · OpenAI and Anthropic endpoints (api.minimaxi.com / api.minimax.io) · Claude Code, Codex, Cursor, TRAE, Hermes, OpenClaw, OpenCode, Pi (Cline undocumented) | CN + international platform — region-specific keys, not interchangeable | only official number: Ultra ≈ 7.1B tokens/month · RPM/TPM 200/10M (M3) · individual interactive use | — |
| OpenCodeGo | $10 / mo ($5 first month) | dollar caps: $12 / 5h · $30 / week · $60 / month · per-model allocation $60/$15 | 5h / weekly / monthly caps | subscription · /zen/go/v1/chat/completions (OpenAI), /responses (Luna), /messages (MiniMax/Qwen, Anthropic-style) endpoints · official OpenCode provider | global — US hosting per docs · test mainland latency and payment | $5 first-month promo, then $10 · only 1 subscriber per workspace · Luna 2x usage is a marketing banner (not in the table) | $5 first month (promo, then $10) |
| Qwen CloudQwen Token Plan (Personal / Team) | Personal Lite $6 · Standard $18 · Pro $68 · Team $20–200 / seat (5h limits temporarily lifted) | credits — Personal 7-day 2,500 / 10,000 / 40,000 (5h temporarily unlimited) · Team 25k / 100k / 250k / seat | Personal: 5h (temporarily lifted) + fixed 7-day window · Team: monthly billing cycle, no windows | subscription · compatible-mode/v1 (OpenAI) and /apps/anthropic endpoints · tools: Hermes (official guide), OpenCode, Cline, Claude Code, Codex, Cursor, Qwen Code, OpenClaw · interactive-only, explicit automation ban | Singapore only (ap-southeast-1) | exact-string model allowlist (qwen3-coder is NOT included) · qwen3.8-max-preview = Team 10x promo · Personal concurrency 1–2 / 3–4 / 6–8 agents | nightly 50% credits 22:00–08:00 UTC+8 (qwen3.8-max) |
| Western and global subscriptions | |||||||
| OpenAIChatGPT + Codex | $20–200 / mo (Plus $20 · Pro 5x $100 · Pro 20x $200; new Go $8) | dynamic limits / credits — no fixed token quota | dynamic · shared across ChatGPT and Codex | subscription (Codex web / CLI / IDE / iOS) · pay-as-you-go API is separate · third-party OAuth via ChatGPT = ToS violation, 3P suspensions | supported countries — CZ yes · mainland China is not listed | API key ≠ subscription credit · mainland officially unsupported | — |
| AnthropicClaude Pro / Max + Claude Code | $20–200 / mo (Pro $20, $17 annual effective · Max 5x $100 · Max 20x $200) | rolling 5-hour window + weekly limits · no message count | rolling 5h + weekly | subscription · Claude Code in terminal and IDEs · ANTHROPIC_API_KEY switches to API billing · third-party OAuth (OpenCode, Cline, RooCode, OpenClaw, OMP) = ban — official Claude Code only | supported locations — CZ yes · mainland China is not listed | not a reusable API key · usage credits after quota = API pricing (opt-in, daily cap $2,000) | — |
| GitHubCopilot Pro / Pro+ / Max | $10–100 / mo (Pro $10 · Pro+ $39 · Max $100) | AI credits 1,500 / 7,000 / 20,000 / mo (1 = $0.01) · completions unlimited | monthly at 00:00 UTC · no rollover | subscription · VS Code, JetBrains, Neovim, Zed, CLI · delegates to Claude and Codex · no public endpoint — automation via Copilot CLI -p only (PAT with “Copilot Requests”) | global (GitHub regions) | chat/agents/CLI consume credits — completions do not | — |
| CursorPro / Pro Plus / Ultra | $20–200 / mo (Pro $20 · Plus $60 · Ultra $200; included usage $20 / $70 / $400) | included dollar usage · cloud agents at API pricing | monthly | editor subscription · integrates inside Cursor · not an external provider | global | credits do not work outside Cursor (Cline/OpenCode/aider) | — |
| ReplitCore / Pro | $25–100 / mo ($20 / $95 annual) · $25 / $100 app credits | app / platform credits · Pro: up to 10 parallel agents | monthly | hosted environment · not a portable API key | global (hosted) | usage depends on Agent mode and effort | — |
| xAISuperGrok / X Premium+ + Grok Build | X Premium+ $40 / mo ($395 / year) · SuperGrok: Lite $10 · $30 ($300/year) · Heavy $300 (3P prices; Heavy existence official) | one shared weekly usage pool across products — no public numbers | weekly pool (reset per account UI) | X / SuperGrok subscription · OAuth into Hermes, OpenCode, Kilo, OpenClaw, Warp · Grok Build CLI (headless -p too) · automation OK · xAI API separate (PAYG) | global (X regions) · EU API Console since 2026-07-17 | the pool is not unlimited — validate usage in your own account · SuperGrok prices are 3P — verify checkout · OAuth 403 gating on some tiers | — |
| ZedPersonal / Pro (editor) | no public monthly seat price · Pro includes $5 of hosted tokens (+ API list price +10%) | $5 hosted tokens · free 2,000 edit predictions | — | editor · own API keys or external agents (Claude Agent, Codex CLI) | global | pricing exposes no stable monthly seat price — I am not inventing one | — |
| GoogleGemini CLI → Antigravity CLI | Google AI Plus $4.99 · Pro $19.99 · Ultra $99.99 (5x) / $199.99 (20x) — none provides an API endpoint | quotas only in AI Studio and Antigravity — no public numbers · AI credits (overage) purchasable | Pro/Ultra: 5h refresh until weekly cap · Free/Plus: weekly | consumer Gemini CLI ended 2026-06-18 · Antigravity CLI · Antigravity OAuth in third-party clients = ToS violation (FAQ) · legal path: GEMINI_API_KEY + OpenAI-compat base_url generativelanguage.googleapis.com/v1beta/openai/ (pay-per-request) | global (Google AI plan regions) — CZ supported | I do not present the old Gemini CLI subscription path as current — the subscription only changes AI Studio/Antigravity quotas | — |
| ClineClinePass | $4.99 first month → $9.99 / month · Cline usage-billing credits separately (100+ models, free models) | 2–5x usage vs standard API · quota in 5h / weekly / monthly windows — no published numbers | 5h / week / month — numbers unpublished | subscription · official API api.cline.bot/api/v1/chat/completions (OpenAI format) · key from app.cline.bot · automation allowed · outside Cline: custom OpenAI-compatible route only | global — no official country list; US export controls | ToS: personal / internal business use · Stripe + fee (~$0.61 3P) · no published quota numbers | $4.99 first month (promo, then $9.99) |
| GitLabGitLab Duo (Agent Platform) | composite: platform tier (Premium $29) + Duo credits $1/credit on-demand (promo 12 / 24 included) · legacy add-on Duo Pro $19 / Enterprise $39 | credits at $1/credit — every request has a credit multiplier · 160 calls / 8h / user rate limit (aiAction) | monthly — included credits do not roll over | REST POST /api/v4/chat/completions + Duo CLI (GA, headless too) · NOT an OpenAI/Anthropic-compatible endpoint · automation OK · OpenCode “GitLab Duo” provider (experimental) | global — no official country table | composite pricing and a strict rate limit — expensive for solo devs · models: Claude Sonnet 4.6 default, Gemini Flash, GPT-5.x, Codestral | — |
Practical picks
AI subscription stacks by budget
Free
- Nous Portal — $0
They always have free models available inside Hermes Agent. Just download Hermes, sign in with Nous Portal, and start using them for free.
$5
- DeepSeek Flash — $5
$10
- Command Code GOAT — $10
$20
- ChatGPT Plus — $20
Probably the best value at this price right now. ChatGPT is effectively unlimited for normal chat, while Codex + Luna can handle around 90% of tasks.
$50
- ChatGPT Plus — $20
- Factory Droid Pro — $20
- DeepSeek Flash — $10
Connect DeepSeek Flash to Droid and you have a very strong daily stack with plenty of usage.
$100
- Cursor Pro Plus — $60
- ChatGPT Plus — $20
- Factory Droid Pro — $20
Cursor gives you heavy Grok usage, plus $60 of API credit for stronger models when you need planning, reviews, or difficult tasks. Grok 4.6 and Composer 3 should make this stack even stronger.
$150
- Cursor Pro Plus — $60
- ChatGPT Plus — $20
- Factory Droid Pro — $20
- DeepSeek Flash — $20
- Command Code GOAT — $10
- Extra API budget — $20
At this point, DeepSeek Flash becomes your daily worker for automation and long-running Hermes tasks, while the extra API budget covers Cursor or Factory when needed.
$200
- ChatGPT Pro — $100
- Cursor Pro Plus — $60
- Factory Droid Pro — $20
- DeepSeek Flash — $20
Basically the $150 stack, but upgrade ChatGPT Plus to Pro.
Best picks
OrchestratorKimi K3 inside Cursor or Droid.
WorkersDeepSeek Flash + Luna. Those two can handle roughly 95% of day-to-day work with great speed, capability, and cost.
Practical recommendations, not quota measurements: DeepSeek Flash, Luna and Kimi K3 are in the graph above; ChatGPT Plus, Cursor and Grok in the subscription table. Products without a table row (Nous Portal, Command Code GOAT, Factory Droid Pro, Composer 3) are not quota-measured in this note — verify prices and limits at checkout.
Method & sources
Official pages, docs and the AA Data API.
Snapshot as of 2026-08-10. Quality: AA Intelligence Index v4.1.1 (max effort) from the Artificial Analysis Data API (free tier, 595 models). Price: AA cost per Intelligence Index task = API list-price cost of one standardized task — not a subscription price. Effective $/1M of subscriptions: (plan price × credits per 1M of mixed usage) ÷ monthly credits; baseline mix 90.9% cache-hit input, 9.1% cache-miss input, 20% output (GLM-5.2, DS V4 Flash, MiMo, MiniMax-M3, Kimi K3, Grok 4.5 from value-data; Baidu plain tokens; OpenCode Go effective per-model allocation). Subscription dots in the graph are derived estimates — same task, same tokens, different unit price; they are not measurements on a subscription quota. Unknown conversions are marked "unverified" and get no dot. II ≠ success on a specific task. USD is indicative (1 USD = 6.77 CNY). Every plan measures consumption differently — none is "unlimited".
The intelligence index and cost per task come from Artificial Analysis — I have not benchmarked any model myself. For independent model-quality and price benchmarking: Artificial Analysis
Each dot's sources are in the selected-dot detail.
Historical indicative rate 1 USD = 6.77 CNY, last updated 2026-08-10: exchangerates.org.uk