There has never been a worse time to pay full price for a coding model.

Every lab with a checkbook is running the same play: torch money now, buy your habit, invoice you later. Chinese labs are the most aggressive about it, but everyone is in on it. The result is a menu where $6 sometimes buys more real work than $30, and the sticker price tells you almost nothing.

Last time I mapped free coding quotas into real work. This time I climb the whole ladder, from $0 to $100, and ask the same question at every rung.

Not “how much is it?”

The question is: how much actual coding does this dollar survive before it hits the wall?

Because that is the only number that matters. A plan is not “unlimited” if the quota dies mid-agent-session. A cheap plan is not cheap if the model can’t finish a bug hunt. Sticker price is marketing. Work-per-dollar is the truth.

One warning before the ladder. I fact-checked all of this in July 2026, and half of it will be stale by autumn. These labs change prices weekly, quietly kill their best tiers, and quote “credits” that are not dollars and “tokens” that are not tokens. I have corrected the numbers I could verify against official pages and flagged the ones that only exist in marketing copy. Verify before you pay. Treat this as a map, not a receipt.

The Ladder

Price The Pick What You’re Really Buying
Free free-coding-models + OpenRouter / Cerebras / Gemini fuel, not a home
$1 Command Code Go $10 of credits for $1 — if you trust their math
$5 Xiaomi MiMo Lite / Doubao Coding Plan the Chinese price war
$7 StepFun Step Plan decent models, fiddly setup
$10 Copilot Pro / OpenCode Go / MiniMax Starter the mainstream, plus real coding plans
$15 GLM Coding Plan Lite (Command Code Pro is vendor-only)
$18 GLM Coding Plan / MiniMax Plus the blood bath
$30 SuperGrok compliance, not quality
$40 airouter.ch unlimited Qwen and DeepSeek
$50 Alibaba Qwen Pro / Cerebras Code the firehose
$80 MiniMax Max-Highspeed (there is no MiMo plan here — that was wrong)
$100 Claude Max 5x just buy Claude

Free

You already know my stance here, so I will keep it short.

The free tier is real now, but it is fuel, not a home. OpenRouter free models, cto.new, KiloCode’s free “Auto” model, OpenCode’s rotating free Zen models, and ModelScope all give you something to burn.

Two honest updates since last time. cto.new launched as “completely free” and is now ad-supported with rolling daily and weekly caps. And every one of these free lanes logs or may train on your prompts — KiloCode’s free router, OpenCode’s free Zen models, and ModelScope all carve out exceptions during free use. Public code only. Do not be a hero with client code on a free tier.

If you want raw model fuel to route through a tool, the genuinely useful free API tiers right now are:

  • Cerebras — 1,000,000 tokens/day free, absurdly fast, but an 8K context cap on free.
  • Groq — fast free inference, tight tokens-per-minute ceilings that bite before the daily cap.
  • Google AI Studio — free Gemini Flash models (Pro got moved off free around April 2026).
  • Mistral — free Codestral, still one of the best free code-specialist models, throttled to ~2 requests/min.
  • Z.ai — GLM-4.7-Flash and 4.5-Flash are genuinely $0 on the API, with a huge context window.
  • GitHub Models — 45+ models free with a GitHub account, per-model daily caps.

None of them survive a messy agent session alone. You will hit a 429, a dead model, or a daily cap right when the work gets interesting. The trick is not picking one — it is routing across them so a dead provider does not kill your session. The full setup is in the free coding post. Everything below is for when free stops being enough.

$1 — The Loss Leader

Command Code sells a “Go” plan for $1/month that hands you $10 of credits every month. Ten dollars for one. The homepage literally says “Everything unlocked. For a dollar.”

Here is the catch, and it is a real one: I could not independently verify Command Code at all. Every number, every model name, every screenshot traces back to their own site or to affiliate posts parroting their site. The model roster they advertise at higher tiers reads like a wishlist. So take the dollar if you want — the downside is a dollar — but do not build your week on it, and do not be shocked if the $10 of “credits” behaves nothing like $10 of real usage. This is the one plan on the ladder where I trust nothing but the price.

$5–$6 — The Chinese Price War

This is the tier that broke my brain, and it is where the money-burning is loudest.

The Xiaomi MiMo token plan is the poster child. It is not quite $5 — the cheapest tier, Lite, is ¥39 ≈ $6/month (a first-purchase discount drops month one to about $5.28) and gives you 4.1 billion credits for MiMo’s own models. Still the best value-per-dollar on the low end, and Xiaomi is very obviously eating the cost to get you hooked. If you buy one thing off this whole list to feel out cheap coding plans, start here.

But MiMo is not alone anymore. $5–$6 is a knife fight between Chinese labs, all selling near-identical coding plans:

  • Doubao / Volcengine Ark Coding Plan (ByteDance) — ¥40 ≈ $5.24/month, with a ¥9.9 (~$1.30) first month. Roughly 1,200 requests per 5-hour window, and the flagship Doubao-Seed-Code reportedly benchmarks near Claude Sonnet class. This is the one I would actually test against MiMo.
  • iFlytek / Astron Coding Plan — ¥39 ≈ $5.40/month for the pro tier, same rough quota, plus a ¥3.9/month “worry-free” tier with unlimited requests on light models only.

One correction to a claim you will see repeated everywhere: MiniMax does not sell a $5 plan where one package is 1,000 credits. MiniMax’s ratio is 1,000 credits = $1, so $5 buys 5,000 credits — and those are prepaid pay-as-you-go credits, not a coding subscription. MiniMax’s actual coding plan starts at $10. It is further down.

The catch for this whole tier: most of these want a Chinese account and Alipay/WeChat Pay, and some open signups in daily flash-sale batches that sell out. The price war is real. Getting a seat is the hard part.

$7 — StepFun

StepFun’s Step Plan starts at $6.99/month for the Flash Mini tier (four tiers, $6.99 up to $99). The models — Step 3.5 and 3.7 Flash — are genuinely fine.

I originally filed this under “good models, bad reliability.” After actually checking, that is not quite fair: there is no documented uptime problem. The pain is dumber than that. Step Plan keys need a non-standard base URL (/step_plan/v1, not /v1), and if your tool points at the wrong one you get a 401 that looks like a dead key. Pair that with a tight ~100-prompts-per-5-hours entry limit and it feels flaky even though the servers are up. Fixable friction, not an outage. But friction is friction, and for $7 you can do better one tier up.

$10 — The Mainstream Shows Up

$10 is where the names you recognize finally walk in, and where the real coding subscriptions start.

  • GitHub Copilot Pro — $10/month, the safe, boring, IDE-native option your company already trusts. (Note: Copilot moved to usage-based “AI Credits” in June 2026, so heavy agent use now meters.)
  • OpenCode Go — a flat $10/month ($5 first month), not a “$10–12” tier like I said before. The $12 is the usage value you get per rolling 5-hour window (~$30/week, ~$60/month), across ~13 open models from the Chinese labs. Still the most honest paid lane: the math is visible instead of magic.
  • MiniMax Coding Plan — Starter — the real MiniMax plan, $10/month, ~1,500 M2.7 requests per 5-hour window, billed in USD (no China-payment lock). MiniMax markets it as roughly Claude-Code-Max-5x-equivalent, which is a stretch, but the quota is real.
  • Tencent CodeBuddy Pro — $9.95/month, 1,000 credits, a Cursor-style IDE on Tencent’s Hunyuan model plus third-party models.
  • ModelArk Lite (BytePlus, the international arm of Volcengine) also lands around $10 if you want the Doubao stack without the Chinese-payment dance.

This is the tier for people who want a real coding tool without thinking about which lab is burning money this week.

$15 — The Awkward Tier

Nothing clean lives at exactly $15.

Command Code Pro claims this slot at $15/month, the step up from the $1 Go plan. Same warning as before, louder: it is only corroborated by Command Code’s own site, and the model list it advertises (Opus 4.8, GPT-5.5, and friends for fifteen dollars) reads like fan fiction. I am not telling you it is a scam. I am telling you I found zero independent proof, and that is enough to keep my card in my pocket.

The honest $15-ish pick is actually the GLM Coding Plan Lite from Z.ai/Zhipu. Sticker is $18, but a standing 30% promo lands it near $12.60, and unlike Command Code it is a real, widely-used plan with real models (GLM-5.2 and friends) that drop straight into Claude Code. More on GLM in the next tier — it is the sleeper of the sub-$20 bracket.

$18–$20 — The Blood Bath

This is where every lab on earth parks its flagship consumer plan, so this is where the war is loudest — and where the internet’s favorite “insane deal” rumors go to die.

Let me kill two of them first, because you have definitely seen both:

  • The MiniMax “$400 of credits for $20” rumor is false. The $20 ZenMux plan bills in “Flows,” not dollars. At their published rate the tier maxes out around $30 of equivalent API value per month. Good, not $400. The one worth buying is the first-party MiniMax Token Plan — Plus at $20: 4,500 M2.7 requests per 5-hour window, one API key, drops into Claude Code. Heavy agent use can still drain it in an afternoon.
  • Zo Computer is not “unlimited LLM.” It is a real and genuinely cool deal — $18/month for an always-on cloud box with 4 cores and 32GB RAM — but the AI is metered: $10 of credits bundled, then pay-as-you-go or bring your own key. A whole computer for eighteen bucks, yes. Unlimited models, no.

Now the picks I would actually make at this tier:

  • GLM Coding Plan ($18 Lite / ~$12.60 promo) — the best quality-per-dollar in the bracket. GLM-5.2 is a legitimately strong coding model, and the Anthropic-compatible endpoint means Claude Code just works. Watch the peak-hours multiplier (quota burns 2–3× midday China time).
  • MiniMax Token Plan Plus ($20) — the highest raw request quota here if you live in an agent.

Then the crowd you already know, all clustered at the same $19–$20, with corrected prices:

  • Claude Pro ($20) — Claude Code with Sonnet. Still the most pleasant of the Western picks.
  • ChatGPT Plus ($20) — Codex access.
  • Cursor Pro ($20) and Devin Core / “Desktop” ($20).
  • Windsurf Pro — now $20 after its March 2026 overhaul (was $15; old subs grandfathered).
  • Kilo Pass ($19, not $20) and Kimi Code Moderato ($19) — both round-down deals.
  • Ollama Cloud Pro ($20).
  • Featherless — has no $20 tier. It is $10 Basic or $25 Premium, unlimited tokens but capped on concurrent connections.

And the newcomers worth knowing at ~$20:

If you are not sure where to start paying, start here. This tier is the best value on the ladder if you only ever spend once — just buy GLM or Claude and stop reading rumor threads.

$30 — SuperGrok

SuperGrok is $30/month, confirmed, sitting between SuperGrok Lite ($10) and SuperGrok Heavy ($300). One buying tip: the standalone $30 SuperGrok gives you more Grok than the $40 X Premium+ bundle, which dilutes it with social features.

The review is otherwise unchanged. If you want compliance, an uncensored lane, and “it’ll actually answer” energy, Grok delivers. If you want the model to be good at coding, it still trails the field. For $30, most of the tiers below out-code it. Buy it for what it is, not for what the marketing implies.

$40 — Unlimited Qwen And DeepSeek

airouter.ch (AI Router Switzerland) is the standout. Found them through Reddit, which is where half the good deals live now.

Corrected details: it is CHF 39/month (≈ $43–49 depending on the franc), a single flat plan, fair-use “unlimited” access via an OpenAI-compatible API to Qwen3.6 and DeepSeek-V4-Flash. The fair-use caps are generous — 3 parallel requests, 240 requests/minute, 10M tokens/minute — with no per-token billing. If those two models cover your work, and for a lot of real coding they do, this is a very different economy from paying per token everywhere else.

Also parked at $39–$40: GitHub Copilot Pro+ ($39, ~$70 of credits + Opus access), Kimi Allegretto ($39), and MiniMax’s Plus-Highspeed ($40, same quota as Plus but ~2× throughput). Copilot Pro+ is the pick if you want frontier Western models with a credit buffer.

$50 — The Firehose Tier

$50 is where “unlimited” stops being a marketing word and starts being an actual firehose.

  • Alibaba / Qwen Coding Plan — Pro (~$50/month) is the GOATed one, and there is bad news attached: the cheap ~$10 Lite tier was discontinued for new subscribers in March 2026, so Pro is the entry point now. It is still an enormous amount of Qwen — the “90,000 requests a month” figure lives here. At that volume you stop rationing and just work.
  • Cerebras Code Pro ($50) — the speed play. Up to 24M tokens/day and ~1,000 messages/day on GLM-4.7 at ~1,000–2,000 tokens/sec. It sells out regularly and runs near 100% utilization, so peak-time requests can queue, but nothing else codes this fast.
  • MiniMax Token Plan — Max ($50) — 15,000 M2.7 requests per 5-hour window. The top of MiniMax’s standard ladder before you pay for speed.

The one to not trust at this tier is Firepass by Fireworks. It is real — unlimited Kimi/GLM in fast mode — but it is invite-only, early-access, and Fireworks deliberately does not publish a price. Sources split between $7/week (~$30/month) and ~$49/month, which tells you how settled it is. If you have a Firepass invite going spare, my inbox is open. Otherwise, treat it as a rumor with a login page.

$80 — About That “82 Billion Tokens”

I had this tier wrong, so let me correct it in public.

Xiaomi MiMo Max is not an $80 plan, and it does not give you 82 billion tokens. The real MiMo tiers are Lite $6, Standard $16, Pro $50, and Max $100 (¥659). There is no $80 rung. And the famous “82 billion” figure is 82 billion credits, not tokens — and MiMo charges up to 100–200 credits per token on cache misses. Xiaomi’s own cache-friendly estimate is that Max buys you roughly 4–10 billion actual tokens a month, depending on the model. Still a lot. Just not the number on the box.

So what actually lives at ~$80? Mostly MiniMax’s Max-Highspeed at $80 — the $50 Max quota (15,000 requests per 5-hour window) at ~2–3× the throughput. If you are a heavy user who wants volume and speed and you are fine on MiniMax’s models, that is the honest $80 buy. MiMo Max itself is a $100 plan, and it shows up in the next section where it belongs.

$100 — Just Buy Claude

At $100, the whole burn-the-VC-money calculus stops mattering, because you are finally paying for models that finish the hard task without you babysitting them.

The picks:

One correction: there is no ~$100 Cursor or Devin plan. Cursor goes Pro $20 → Pro+ $60 → Ultra $200; Devin goes Core $20 → Team $500. Do not go looking for a hundred-dollar tier that does not exist.

If you do want to spend $100 on something other than Claude, the real field is:

But honestly, at this number, stop optimizing. Claude Max or Codex, buy it, ship.

The Stuff That Bites

The same traps showed up on almost every pricing page. Learn them once and you will read all of these faster.

  • Credits are not dollars, and neither are tokens. MiMo’s “82 billion” is credits, and it charges up to 200 credits per token. GitHub’s “AI Credits” are worth a cent each. MiniMax credits are 1,000-to-the-dollar. Every lab minted its own funny money. Find the conversion before you get excited.
  • The window is five hours now, not a month. Most of these meter you in rolling 5-hour buckets, often with a weekly cap on top. “15,000 requests” reads huge until you learn it refills every five hours — and you can still wall yourself in one afternoon.
  • “Unlimited” means fair-use. airouter, Featherless, Ollama Cloud, Chutes, NanoGPT — every “unlimited” plan is really capped on concurrency, tokens, or speed. Chutes killed its true-unlimited tier in February. NanoGPT killed unlimited input tokens. The word is decoration.
  • Peak hours cost double. The Chinese plans, GLM especially, burn quota 2–3× during China working hours. Code at night and the same subscription lasts longer. Literally.
  • Free tiers read your code. Free OpenRouter, KiloCode’s free router, OpenCode’s free Zen models, cto.new, ModelScope — all reserve the right to log or train on your prompts. Public code only.
  • The best deals are geo-locked. The $5 price war is real, but most of it wants a Chinese account, Alipay/WeChat, and sometimes a flash-sale signup slot that sells out at 10:30am Beijing time.
  • Tiers die quietly. Alibaba killed its $10 Lite plan. Windsurf got swallowed into Devin. Gemini Pro left the free tier. Whatever you buy on my word, check the page first.

The Full List — Everything We Checked

The ladder up top is the opinionated version. This is the whole ledger: everything that survived fact-checking in July 2026, grouped by what it actually is, with a direct link and the one thing you need to know before clicking. Monthly USD unless noted. Bookmark it, argue with it.

Free Fuel

Route these through a tool; none survive a real session alone.

Plan Price The catch
OpenRouter free models $0 20/min, 50/day → 1,000/day after a one-time $10 top-up
Cerebras $0 1M tokens/day, absurdly fast, hard 8K context cap
Groq $0 fast, but tokens-per-minute ceilings bite before the daily cap
Google AI Studio $0 free Gemini Flash only — Pro left the free tier in April
Mistral Codestral $0 best free code model, throttled to ~2 req/min
Z.ai GLM-4.7-Flash $0 genuinely $0 on the API, ~200K context
GitHub Models $0 45+ models, tight per-model daily caps
ModelScope $0 2,000 calls/day, but needs Alibaba real-name verification
cto.new $0 now ad-supported with rolling caps; reads your code
Free tool tiers $0 Copilot Free, Cursor Hobby, Warp Free, Zed Personal, JetBrains AI Free, Cline, Amp, Aider — thin but real

The Chinese Coding Plans (The Price War)

The most work-per-dollar on the board, if you can pay in yuan.

Plan Price The catch
Xiaomi MiMo Token Plan $6 / $16 / $50 / $100 credits, not tokens; coding tools only; CN payment
Doubao / Volcengine Ark ~$5 (¥40), ¥9.9 first mo Doubao-Seed-Code benchmarks near Sonnet; CN payment
iFlytek / Astron ~$5 (¥39) also a ¥3.9 unlimited-light tier; CN payment
StepFun Step Plan $6.99 – $99 needs a non-standard base URL or you get a fake 401
Tencent CodeBuddy Pro $9.95 Cursor-style IDE on Hunyuan; USD billing
Tencent Hunyuan Coding Plan ¥40 / ¥200 Pro ~90k req/mo; CNY only
MiniMax Coding Plan $10 / $20 / $50 M2.7; USD billing; +highspeed $40/$80/$150
GLM Coding Plan (Z.ai) $18 / $72 / $160 GLM-5.2; quota burns 2–3× at peak China hours
Alibaba Qwen Coding Plan ~$50 Pro ~90k req/mo; the $10 Lite tier was killed in March
Kimi Code (Moonshot) $19 / $39 / $99 / $199 K2.6/K2.7; Agent Swarm up to ~300 subagents

Western Tools & IDEs

The names you know, with corrected prices and the odd surprise.

Plan Price The catch
Trae (ByteDance) $3 / $10 / $30 / $100 unlimited autocomplete, token-metered agent
ChatGPT Go $8 light Codex, not for all-day dev
GitHub Copilot Free / $10 / $19 / $39 / $100 moved to metered “AI Credits” in June
OpenCode Go $10 flat $10, ~$12 value/5h — the most honest math
Zed Free / $10 / $30 fast native editor, edit-prediction focused
JetBrains AI Free / $10 / $30 credits per 30 days; completions stay free
Command Code $1 / $15 only their own site confirms any of it
Amazon Q Developer $19 Pro ~1,000 agentic req/mo, IP indemnity
Cosine Genie $20 80 async tasks/mo, opens its own PRs
BLACKBOX AI $10 / $20 / $40 300+ models, use-it-or-lose-it credits
Cursor $20 / $60 / $200 no $100 tier exists; Ultra is $200
Claude $20 / $100 / $200 the pleasant one; Max 5x is the $100 pick
ChatGPT / Codex $20 / $100 / $200 the other one that actually competes at $100
Google AI / Antigravity $20 / $100 / $200 Gemini Pro/Ultra plus the new agentic IDE
Devin $20 / $200 / Teams Windsurf got folded in as “Devin Desktop”
Warp Free / $20 / $50 / $200 agentic terminal, credit-metered
Replit $25 / $100 build-in-browser, credit-metered
SuperGrok (xAI) $10 / $30 / $300 compliance yes, coding still weak
Qodo $30 team agentic PR-review focus
Factory Droid $20 / $100 / $200 token-metered coding agent
Augment Code $100 Business flat, up to 50 seats — cheap for a team
Amp (Sourcegraph) free + PAYG passes model cost through at cost, no markup
Cline free + BYOK / $20 team OSS extension, first 10 seats free
Aider $0 + BYOK OSS CLI pair programmer, you pay the API

Aggregators, Routers & “Unlimited”

One key, many models — or one flat fee and a fair-use asterisk.

Plan Price The catch
OpenRouter PAYG no markup, one API for 300+ models, 28+ free
airouter.ch CHF 39 fair-use unlimited Qwen + DeepSeek, Swiss-hosted
Cerebras Code $50 / $200 fastest coding anywhere; sells out; daily token cap
Synthetic.new ~$30/pack flat open-source LLMs, 500 req/5h per pack
Fireworks Fire Pass ~$7/wk or ~$49? invite-only, price unpublished, unlimited Kimi Turbo
NanoGPT $8 / $12 200+ OSS models, fair-use (unlimited tokens killed Feb)
Ollama Cloud Free / $20 / $100 GPU-time metered, not truly unlimited
Featherless $10 / $25 (+$100/$200) unlimited tokens, capped on concurrent connections
Chutes.ai PAYG + subs cheap OSS; unlimited killed Feb (now 5× PAYG cap)
Requesty PAYG +5% 400-model router, 200 free req/day
Together AI PAYG 200+ models, $5 minimum to start
DeepInfra PAYG ~90 open models, cached-input discounts
Novita AI PAYG 200+ models plus rentable GPUs
SiliconFlow PAYG 200+ models, contexts up to 1M tokens
NVIDIA NIM Free dev / PAYG free tier is dev-only; prod needs an enterprise license

How To Actually Pick

Do not read this ladder top to bottom and buy the most expensive thing you can stomach. Read it against your own week.

  • You code a little. Free stack, plus the $1 Command Code trial if you feel like gambling a dollar.
  • You want the best value on the board. The $5–$6 Chinese coding plans — MiMo Lite or Doubao. It is not close.
  • You want one clean subscription and no thinking. Somewhere in the $18–$20 blood bath — GLM Coding Plan, MiniMax Plus, or Claude Pro if you want the comfortable Western option.
  • You live inside an agent all day. $50 Alibaba Qwen Pro or Cerebras Code, where the quota stops being a leash.
  • You just want the best model and don’t care about the game. $100, buy Claude, move on.

The old instinct was that more money buys a better model. Right now, more money mostly buys you out of the money-burning carnival happening under $100. Everything below that line is a lab paying for your attention — and half of them are quoting you credits that aren’t dollars and tokens that aren’t tokens.

So take the subsidy. Just don’t confuse the sticker price with what you’re actually getting — and don’t get loyal to a plan that is only cheap because someone upstream is bleeding to keep it that way.