What one month of AI costs, side by side.

Every product here sells the same underlying thing — tokens in, tokens out — and every one of them charges for it differently. So a price column alone is misleading: $14.99 and $19.99 can buy you the same compute and still be wildly different deals, depending on what happens when the number runs out.

Prices re-checked at each vendor's own pricing page on 9 October 2026 (USD, US market, excluding tax). Where a vendor does not publish a figure — or publishes it behind a bot wall — the cell says so rather than guessing. This market reprices monthly; treat it as a snapshot, not a quote.

Rows are plans, not vendors. Each vendor's cheapest published paid plan has its own row; a second row where a vendor's step-up tier is the one people are actually quoted, marked step-up. Sorted by monthly price, cheapest first, with the three pay-per-token platforms that sell no subscription pinned last. Column 2 is the price of the plan named in that row. An em dash means no annual option is published; where a figure is shown it is the vendor's own published form (monthly equivalent “/mo”, or the annual total “/yr”). List prices, excluding tax and any carrier or platform pass-through. Full notes for every row are in the fine print below. flat quota credit pot hybrid pay per token
Product	Tier price /mo	Billed annually	Overage model	Free tier
Google AI Plus	$4.99/mo	—	flat quota	yes
ChatGPT Go	$8/mo	—	flat quota	yes
Mistral Pro	$14.99/mo	—	hybridflat + $25.50 credits	yes
Perplexity Pro	$17/moannual billing only	$17/mo	hybridcredits + rate-limited search	yes
Z.ai GLM Coding Plan	$18/mo· $12.60/mo annual	$12.60/mo Lite tier	credit potweekly, 10,000 credits	no
Kimi (Moonshot)	$19/moModerato· $15/mo annual	$15/mo	credit potone shared pool	limited
Google AI Prostep-up	$19.99/mo· $199.99/yr	$199.99/yr	flat quota	yes
Microsoft 365 Premium	$19.99/mo· $199.99/yr	$199.99/yr	flat quota	yes
Claude Pro	$20/mo· $17/mo annual	$17/mo	flat quota	yes
ChatGPT Plusstep-up	$20/mo	—	flat quota	yes
Ollama Pro	$20/mo· $200/yr	$200/yr	credit pot$60/mo of credits	yes
xAI Grok	$30/moSuperGrok	—	flat quota	yes
Claude Maxstep-up	from $100/mo	—	flat quota	n/a
OpenRouter	no subscription	—	pay per token5.5% platform fee	yes
Together AI	no subscription	—	pay per token	no
DeepInfra	no subscription	—	pay per token	no

The details

Background and per-product notes, in fine print, in table order. The table above is the price; everything worth reading around it is here.

Three billing models hide inside these price tags

Two plans at the same price are not comparable until you know which of the three you are buying.

Model 1 — Flat-rate quota: One price, a usage ceiling, then degrade or stop — the model gets slower, the context window shrinks, or the assistant refuses until the period resets. You are not billed more when you hit the wall. Predictable, and the safest thing to buy on a fixed budget; the cost is that the ceiling is deliberately vague (“more usage”, “expanded limits”, “2×”) and you find the real limit by hitting it. Nor can two flat tiers be compared to each other: every one describes its ceiling as a multiple of its own free tier, and every free tier differs, so 2× of a tiny allowance can be smaller than 1× of a large one. The only real numbers on a flat tier carry units — terabytes, projects, concurrent tasks, hours per week. Treat the rest as a direction of travel. Here: Claude Pro and Max, ChatGPT Go and Plus, Google AI Plus and Pro, Microsoft 365 Premium, xAI SuperGrok.
Model 2 — Credit-based: A monthly stipend of pay-as-you-go credits, metered by tokens: it looks like a subscription and behaves like a prepaid balance. Burn it in a week and you are done for the month, or you top up and keep going. This is the model most often described wrongly — a “plan with $60 of credits” is not unlimited access with a soft cap, it is $60 of token spend that happens to renew. There is no ceiling protecting you and no unlimited access hiding behind the wording; when the pot is empty, the meter stops. Read the Ollama and Mistral rows side by side: one is a subscription with credits attached, the other is credits with a subscription attached. Here: Ollama Pro ($60/mo of credits), Mistral Pro ($25.50/mo of API credits), Perplexity Pro (Computer credits, alongside rate-limited search), Kimi Moderato and up (one shared credit pool), Z.ai GLM (10,000 credits a week).
Model 3 — Pure pay-as-you-go: No subscription at all. You pay per token and nothing else — no monthly floor, no seat, no ceiling. The only recurring charge is a platform fee on top of the token price. This is the honest floor for every comparison on this page. Before you renew anything, price your actual month of work here; that is the number the subscription has to beat. Here: OpenRouter (5.5% above token price), Together AI, DeepInfra. Also how Ollama’s Free tier and every vendor’s API console work.

Per-product notes

↑ tableGoogle AI Plus the floor of the paid market

What you get: 400 GB of storage across Gmail, Drive and Photos; 2× access to Gemini compared with no plan; Gemini 3.1 Pro and Deep Research; Gemini in Gmail and the Google apps; image, music and video generation in Gemini and Search; Gemini Notebook; family sharing with up to 5 others.
Overage: Flat-rate quota — a fixed multiple of the free Gemini allowance. Hit it and you fall back to the free tier until it resets; no metered bill.
Free tier: Yes — Gemini free tier plus the 15 GB that comes with a Google Account.
Best for: The least you can spend to be inside a major assistant’s paid tier at all — the cheapest entry plan on this page.

↑ tableChatGPT Go the ad-supported entry

What you get: Everything in Free, plus more messages with tools, more uploads, more image creation, more voice chats, longer memory. OpenAI states the plan may include ads.
Overage: Flat-rate quota — exhaust the allowance and you drop back to the Free limits. No overage purchase on this tier.
Free tier: Yes — ChatGPT Free, $0: unlimited text chats with GPT-6, subject to abuse guardrails.
Best for: Cheapest way into the OpenAI line. No advanced reasoning models and no Codex headroom, so it is a chat plan, not a work plan.

↑ tableMistral Pro a subscription that hands back credits

What you get: Full access to Vibe: all-day coding in the CLI, IDE or web, long-running tasks, faster responses, spreadsheet and slide generation. Plus $25.50/mo in API credits to spend on anything in the Mistral API.
Overage: Hybrid — flat price plus a credit stipend. The $25.50 is a genuine credit pot and can be spent anywhere. The Vibe side is described as “more access and usage”, i.e. a ceiling you hit and wait out. The two halves run out independently.
Free tier: Yes — limited messages and web searches, limited coding sessions, basic image generation, 100+ connectors.
Best for: Coding all day in one tool while also getting a monthly pot of raw API credits. The most credit per dollar of the subscription tiers here.

↑ tablePerplexity Pro research with an answer engine attached

What you get: Expanded Perplexity Computer access; 4,000 bonus credits (limited-time offer); deep research; access to the top AI models and the ability to pick between them; polished documents and apps; higher usage limits and memory; 50 file uploads per project.
Overage: Credit-based plus quota — Computer runs on credits: 4,000 bonus on Pro against the 10,000/month stated on Max, so the paid tier is a pot, not a ceiling. Ordinary searches are limited by rate rather than metered.
Free tier: Yes — Free, $0/month: search with citations, basic models, limited daily usage.
Best for: Cited research across several frontier models, running on a monthly pot of Computer credits. The step-up is Max at $167/mo annual.

Not verified: the pricing page quotes annual billing only; the monthly-billed rate is not shown there (US Apple IAP has been observed at $20).

↑ tableZ.ai GLM Coding Plan credits by the week

What you get: 10,000 credits per week; rolling access to the latest flagship GLM models; works with 20+ agent tools including ZCode, Claude Code, Codex, Cursor and OpenClaw; default data privacy.
Overage: Credit-based, weekly — a credit pot that refills each week rather than each month. Spend the week’s credits and the plan waits for the reset; there is no overage purchase on the individual tiers.
Free tier: No free coding tier — the subscribe page prices three paid tiers only (Lite, Pro, Max).
Best for: Cheapest way to run a coding agent inside somebody else’s CLI, if you work in bursts that fit inside a week.

↑ tableKimi (Moonshot) one pool for chat, agent and code

What you get: About 60 agent uses a month from one shared credit pool; 2 parallel agent tasks; Agent Swarm; Dream Memory and self-evolving skills; Kimi Code, with its own 5 h/week cap; 20 projects and 20 GB project storage; 2,000 professional-database calls; file processing, deep research and website deployment with a database.
Overage: Credit-based — every feature (agent, deep research, docs, Kimi Code, K3) draws on one token-metered credit pool. Empty it and features stop until renewal; there is no overage charge. Kimi Code additionally has a 5 h/week ceiling that buying credits does not lift.
Free tier: Chat is usable without a plan; no priced free tier is listed.
Best for: Replacing four subscriptions with one credit pool covering chat, agent tasks, documents and code.

Not verified: the international page starts at Moderato ($19/mo). The China-locale page lists a cheaper tier below this one (¥49/mo) with no USD price.

↑ tableGoogle AI Pro step-up: the bundle play

What you get: 5 TB of storage; 4× Gemini access; Gemini 3.1 Pro, Deep Research and Deep Search; Gemini in Gmail, Docs and Sheets; YouTube Premium Lite; $10/mo in Google Cloud credits; Google Health Premium; Chrome auto browse; family sharing with up to 5; higher limits in AI Studio, Antigravity and Jules.
Overage: Flat-rate quota — fall back to the free allowance at the ceiling. The only hard number is the storage; everything AI-side is “expanded” or “4×”, which is a multiple, not a quantity.
Free tier: Yes — 15 GB of account storage and the Gemini free tier.
Best for: People already paying Google for storage. The subscription is storage-first; the AI comes bundled on top of it.

↑ tableMicrosoft 365 Premium the old Copilot Pro, folded in

What you get: Word, Excel, PowerPoint, Outlook and OneNote with Copilot for the subscription owner; Copilot Chat with the latest models, image and voice generation; exclusive features such as Researcher and Analyst; up to 6 TB of OneDrive storage (1 TB per person, up to 6 people); Teams, Defender, Designer, Clipchamp.
Overage: Flat-rate quota — Copilot usage is limited by credits at the tier; Microsoft calls Premium’s the highest consumer limits. Only the account owner gets the AI; the other five members do not.
Free tier: Yes — free Copilot web chat with a Microsoft account: web chat only, no AI inside the Office apps.
Best for: Anyone who wants Copilot in the desktop Office apps. Standalone Copilot Pro is no longer sold to new customers; this is its successor, and buying it costs less than the old add-on plus a base plan did.

↑ tableClaude Pro the cheapest tier with an agent in it

What you get: Everything in Free, plus Claude Code included; more usage; Projects; more Claude models; Claude in Chrome and Microsoft 365; hand-off and scheduled tasks; Claude Dashboards (beta); Claude Science.
Overage: Flat-rate quota — a usage ceiling you hit and wait out; no overage invoice. The published escape hatch is a bigger tier, not a bigger bill: Claude Max starts at $100/mo for 5× or 20× Pro usage.
Free tier: Yes — Claude Free, $0: chat on web, desktop and mobile, search, files, code, memory, artifacts, up to 5 projects.
Best for: Someone whose first $20 should include an actual coding agent. Claude Code in the price is the differentiator against every other chat tier here.

↑ tableChatGPT Plus step-up: the default, priced like it

What you get: Everything in Go, plus advanced reasoning models with GPT-6; expanded messages and uploads; more complex and accurate image creation; expanded deep research; expanded memory and context; Projects, scheduled tasks and custom GPTs; expanded Codex usage; early access to new features.
Overage: Flat-rate quota — “Limits apply.” No overage purchases on the consumer tiers; the step-up is Pro, from $100/mo, not an extra charge on Plus.
Free tier: Yes — ChatGPT Free, $0: unlimited text chats with GPT-6 under abuse guardrails, limited tools, uploads, image and voice.
Best for: The general-purpose paid assistant: a mix of chat, images, research and Codex in one seat, at the price most people are already anchored to.

Still $20, not $30 — the OpenAI page shows $20.

↑ tableOllama Pro credits, not unlimited access

What you get: $60 of usage credits per month; access to the larger pro models; run multiple models concurrently; no service fees; the local runner stays free.
Overage: Credit-based — say it plainly. The $60 is a monthly stipend of pay-as-you-go credits, not a usage ceiling and not unlimited access. Burn it in a week and you top up or stop. Rates are per million tokens: deepseek-v4.1-flash $0.30 in / $1.20 out, dropping to $0.15 / $0.60 off-peak (off-peak applies outside 12:00–18:00 UTC on weekdays and all weekend). The Team tier is the same shape at $500/mo with $1,000 of credits.
Free tier: Yes — Free, $0: run models locally, starter usage credits, access to starter models, no service fees.
Best for: Trying many open models through a single meter. The right choice if you would rather pay per token than rent a seat — just do not read “$60 of credits” as a cap you cannot cross.

↑ tablexAI Grok the generous free tier

What you get: Higher rate limits and access to frontier models; Grok 4-class reasoning; live web and X search; image and video generation; Plus adds significantly higher usage, 1080p video and priority access.
Overage: Flat-rate quota — the paid tier buys headroom (higher limits), not a different product. No metered overage is quoted.
Free tier: Yes — Free, $0/month: “generous limits” on Grok itself, with real-time web and X search.
Best for: Live social and web grounding. The free tier covers a lot of that, so $30 buys capacity rather than capability.

Not verified: the pricing page’s plan matrix also lists a SuperGrok Lite tier with no published price (checked 9 October 2026); the $30 SuperGrok is the cheapest tier that page prices.

↑ tableClaude Max step-up reference, not a floor

What you get: Everything in Pro, plus 5× or 20× Pro usage; higher output limits for all tasks; early access to advanced features; priority access at high-traffic times.
Overage: Flat-rate quota — same shape as Pro at five or twenty times the size. Still a ceiling, still no metered bill.
Free tier: n/a — an upward step from a paid tier.
Best for: A reference point for scale: five times the Pro ceiling costs roughly five times a $20 seat. Useful for judging whether the first $20 was really the limit you hit.

↑ tableOpenRouter no subscription — the honest floor

What you get: 500+ models across 80+ providers; one API key and one bill; auto-routing and preferred-vendor selection; budgets and spend controls; prompt caching; data-policy-based routing; activity logs and export.
Overage: Pure pay-as-you-go — no floor and no ceiling. The only recurring number is the platform fee: 5.5% on Standard, 8% on Business, discounted on Enterprise.
Free tier: Yes — 25+ free models from 4 free providers, with chat and API access, but no routing, budgets or caching.
Best for: The baseline every subscription above should be measured against. Model shop across vendors without signing anything.

↑ tableTogether AI per-token, open weights

What you get: Open models behind one API; DeepSeek V4.1 Flash $0.30 in / $1.20 out; MiniMax M3 $0.30/$1.20; GLM-5.3-Flash $0.15/$0.50; gpt-oss-120B $0.15/$0.60; cached input is cheaper again on most rows.
Overage: Pure pay-as-you-go — nothing recurring except what you spend. Provisioned throughput (reserved capacity in PTUs) is a separate enterprise contract: an alternative to metering, not a subscription tier.
Free tier: No — no free tier is published; the cheapest models are simply very cheap.
Best for: Running an open model at volume when per-token beats per-seat. Compare a $20 seat against roughly 50 million DeepSeek V4.1 Flash output tokens here.

↑ tableDeepInfra the cheapest tokens on this page

What you get: No contracts and no upfront cost; DeepSeek-V4-Flash-0731 $0.06 in / $0.18 out; DeepSeek-V3.2 $0.26/$0.38; Kimi-K3 $2.85/$14.25; Qwen3.5-397B $0.45/$3.00; cached input priced separately on most rows.
Overage: Pure pay-as-you-go — pay for what you use, with no subscription and no minimum. A 1M-token context is available on the Flash models.
Free tier: No — no free tier is published.
Best for: The lowest per-token floor here for a general-purpose model. If your workload is ordinary chat and summarising, this is the number your subscription has to beat.

Bottom line

The cheapest entry plan in this market is $4.99 (Google AI Plus), then $8 (ChatGPT Go). The $20 you probably have in mind buys a different product depending on where you put it:

Claude Pro ($17 annual, $20 monthly) if the first $20 should include a coding agent — Claude Code is in the price.
ChatGPT Plus ($20) for the broadest single consumer product: chat, images, research and Codex.
Mistral Pro ($14.99) if you want a subscription that also hands back $25.50 of raw API credits every month.
Ollama Pro ($20) if you would rather hold $60 of tokens than rent a seat — knowing it is a stipend, not a ceiling.
Google AI Plus ($4.99) or AI Pro ($19.99) if what you actually want is storage with an assistant attached.
Price the same month of work on OpenRouter or DeepInfra before renewing anything. That is the honest floor, and if a subscription cannot beat it, the subscription is the wrong buy.

Whichever you pick: verify the price at the vendor's own page before you pay. Every figure here was correct on 9 October 2026 and this market reprices monthly.