What a Heavy AI Habit Actually Costs

A Claude Max subscription is the cheapest way to rent a frontier model, one of the top-tier AI models like Claude Opus, right now. Three months of my own Claude Code usage prove it: $19,765.21 of tokens at Anthropic’s published rates from July to September 2026, on three A$160 bills. Seems a great deal, but the hidden cost is the data you hand over, and where it then sits. Everything you type into a closed model lives on servers you don’t control, under data terms the company writes and can rewrite, with no way to verify what’s kept or check what’s done with it.

The compute is subsidised and temporary. Handing your thinking to someone else’s infrastructure is the part that lasts.

I’ll put numbers on both halves. The bill that hits my account is in Australian dollars, so that’s where the cash side starts, with US figures alongside.

The number, and how I got it

The figure comes from ccusage, a tool that reads Claude Code’s local logs and values every tokenTokenA digital unit of value or access rights tracked on a blockchain. Tokens can represent ownership in a project, a right to use a service, a share of future revenue, or simply a tradable asset with no underlying claim.Like a physical poker chip a casino issues. The chip itself has no value. What makes it worth something is what it lets you do at the casino, what the casino has promised, and how much other people will pay you for it.Token in the glossary → at Anthropic’s own APIAPIApplication Programming Interface. A structured way for one piece of software to talk to another. In DeAI, APIs let applications request inference from a model without running the model themselves.Like a waiter in a restaurant. You don't walk into the kitchen and cook your own meal. You tell the waiter what you want, they tell the kitchen, the kitchen cooks it, and the waiter brings it back. The API is the waiter.API in the glossary → rates. Tokens are the small chunks of text a model reads and writes, and they’re the unit AI companies bill by. API rates are the per-token prices Anthropic charges developers who pay as they go. I pay A$160 a month for Claude Max 5x, the GST-inclusive Australian price of the US$100 tier, and I stayed on that tier for all three months.

Against the A$160 I actually pay, July alone came back about 44 times over. August more than doubled July’s tokens, and September ran to the 30th.

These figures are a floor: they count Claude Code only, and the web and desktop apps never expose token counts, so my actual usage is higher. The second caveat has nothing to do with the money. Claude Code was quietly deleting my own transcripts after 30 days until I changed a setting to keep them.

Hold that one. It comes back.

Claude Code pricing, plan by plan

Claude Code has no price of its own. You get it with a paid Claude plan or run it on an API key, and the plan sets how much of it you get. Anthropic’s pricing page and help centre, read 30 September 2026:

Local prices differ. In Australia the Max 5x tier bills A$160 including GST.

Claude Pro vs Max

Pro and Max run the same Claude Code. What changes is how much of it you get before the allowance runs out, plus two extras listed only for Max. From Anthropic’s pricing page and help centre, read 30 September 2026:

How I’d choose:

• Pro if you hit the limit now and then. Usage credits cover the odd heavy week at API rates.
• Max 5x if you hit Pro’s limit most days. That’s Anthropic’s own advice: move to Max 5x if you consistently hit limits.
• Max 20x if you still run out on Max 5x.

Claude Max 5x vs 20x

Max 20x costs twice as much as Max 5x for four times the session allowance. The help centre states the 5x and 20x multiples for the five-hour session only; none of the pages I read gives the weekly cap as a number. Moving up is charged pro rata for the rest of your billing cycle, so trying 20x costs part of a month. I stayed on Max 5x for all three months at the volumes above.

Claude Code usage limits

Anthropic sets out the limits in its help centre rather than on the pricing page:

• Five-hour sessions. Your session allowance resets every five hours.
• A weekly cap. Pro and Max both carry a weekly limit across all models, resetting at a fixed time assigned to your account.
• One shared pool. Usage in claude.ai, Claude Desktop and Claude Code all counts against the same limit.
• Usage credits. Pro, Max 5x and Max 20x subscribers can switch on credits to keep working past the limit, billed at standard API rates.

Anthropic’s own advice for stretching the allowance is to type /clear between tasks, which starts a fresh conversation, and /compact in long sessions, which shrinks the history to a summary. My logs show why. Every time Claude Code replies, it re-reads the whole conversation so far, and those re-reads (cache reads) made up 98% of my Claude Code tokens in both August and September. The longer the session, the more each reply costs.

The cheapest way to run Claude Code

It depends on how hard you run it:

• Occasional use: Pro, with usage credits switched on for the odd heavy week.
• Daily, heavy use: Max. At my August volume the API would have charged $8,618.85 for what the plan covered at US$100.
• Automated or bursty work: an API key, where you pay for exactly the tokens you use and nothing waits on a reset.

Change the model and the per-token price falls further. Open-weight coding models, the ones anyone can download and run, such as Kimi, GLM and DeepSeek, list on Venice at a fraction of Opus’s rate, and Venice runs them on its Private tier. Our guide to running Claude Code and opencode on open-weight models covers the set-up, tested on the Morpheus network.

Why it’s this cheap

Anthropic is losing money on me, and it isn’t close. For a heavy user the flat subscription breaks, and the pattern is well documented across coding tools: the top tenth of users burns most of the compute, and a plan that prints money on a casual user runs deep in the red on a power user. TechCrunch flagged the thin margins across AI coding startups in August 2025, and Replit’s chief executive conceded flat-rate pricing was unsustainable before moving to usage-based billing.

The same holds at the company level. Anthropic told investors it expects to burn about $3 billion in 2026, down from roughly $5.6 billion in 2025, and to stop burning cash in 2027 (reported by The Information, carried via Reuters). Revenue is climbing fast. Today’s prices are still funded by investors, not by profit on the inference.

Fact: My Claude Code usage from July to September was worth $19,765.21 at Anthropic’s rates against three A$160 bills, and the company is burning billions a year while it grows. Take: That gap is the product. Frontier tokens are this cheap because someone else is paying the difference to win the market, which is a fine deal to take and a poor thing to depend on. A subsidised price is a promotion, not a floor.

What the same habit costs on the alternative

Move the habit to Venice and the economics flip in a way most people get wrong. Venice runs two separate surfaces. On its own hosted models, the open-weight ones like Llama, Qwen and DeepSeek, paid tiers give unlimited text in the web app with no per-token meter. On frontier closed models, premium media, and all API use, you pay per token in credits, even inside the chat app.

So Claude through Venice sits outside the unlimited bucket: it’s metered per token, at roughly 20% over Anthropic’s direct rate. On 30 September 2026 Venice listed Claude Opus 5.5 at 1.2 times Anthropic’s own API price, on input and output alike. For frontier coding, Venice is the more expensive way to buy the same tokens, and I’m not going to dodge that.

Venice draws the free-versus-paid line precisely:

• Web app plus specific free models: no credit drawn. This is the only unlimited usage.
• Web app plus pay-per-use models (the frontier ones): credits drawn per token.
• The API, any model, open or closed: always metered. There is no free programmatic tier.

Generous as a chat app, metered the moment you build or code on it. That’s the shape of the thing.

Where the subscription can’t follow

On a closed model, your prompts sit under a policy. Anthropic’s is decent today: as of March 2026 it doesn’t train on consumer chats unless you opt in. But a policy is a promise the company can revise, safety-flagged conversations get reviewed, and you can’t independently check what’s retained or walk out with it. You’re trusting the terms, not verifying them.

Venice sells the opposite: privacy you can check, in the modes that carry attestation. It runs four modes, and only one of them is the weak one you’d use for frontier tokens.

Venice says Private keeps inferenceInferenceRunning a trained AI model to produce an answer. Inference is what happens when you type a prompt into ChatGPT and get a response. The model takes your input, computes a best guess, and returns it.Like asking an expert for their opinion. The training was the decades they spent becoming an expert. The inference is the 30 seconds it takes them to answer your specific question.Inference in the glossary → on zero-retention infrastructure, which you have to take on trust. TEETEETrusted Execution Environment. A hardware-secured region of a CPU or GPU where code runs in isolation, so even the machine's operator can't read what's happening inside. TEEs give decentralised AI inference privacy guarantees.Like a bank vault inside a bank. The bank owns the building, staffs the lobby, and runs the security cameras. But what's inside the vault is invisible to everyone, including the bank staff, unless the customer opens it.TEE in the glossary → runs it inside hardware enclaves with remote attestationAttestationA cryptographic proof that a piece of code is running on a specific hardware enclave in an unmodified state. Attestation lets remote users verify that a service is genuinely running what it claims to be running.Like a tamper-evident seal on a medicine bottle. The seal itself doesn't make the medicine safe, but it gives you a way to verify that nobody opened the bottle and swapped the contents before you bought it.Attestation in the glossary → you can check yourself. E2EE encrypts your prompt on your device, so neither Venice nor the GPU provider ever sees it. Anonymous, the mode that reaches Claude and GPT, only proxies you and hides your identity, which is why frontier privacy on Venice is thinner than the marketing implies.

Now remember Claude Code deleting my transcripts. That’s the small version of the same problem: on the closed stack, the terms of what happens to your thinking aren’t yours to set. On the TEE and E2EE modes above, they’re enforced by architecture you can inspect rather than a policy page you have to trust.

Owning the access instead of renting it

The second thing you can’t rent is the access itself. With Claude Max you pay monthly for a seat, at a price Anthropic sets and can change.

With Venice you can hold the access as an asset. StakeStakingLocking up a cryptocurrency to help secure a blockchain network, usually in exchange for rewards. The locked tokens act as a security deposit that can be taken away if the staker misbehaves.Like putting down a large rental deposit for an apartment. You get the money back if you behave, you earn interest while it's locked, and the landlord takes it if you trash the place.Staking in the glossary → VVV and you get complimentary Pro for open-model chat, plus a share of DIEM, the platform’s credit. Each DIEM you stake is worth $1 a day of perpetual, renewing inference against any model, and you reclaim your stake when you exit.

That’s a real, if partial, form of owning your inference rather than renting it. You hold the key, it’s portable, and the per-unit terms have held since DIEM launched. Be precise about that last one though: $1 a day is Venice policy resting on the Venice balance sheet, not a contract you could enforce. Erik Voorhees calls each DIEM a liability of the company, which is the right way to read it and the reason the supply is capped by design.

It isn’t trust-free, and I won’t pretend otherwise. VVV and DIEM carry token-price and emissions risk, the staking yield comes from new supply, and you’re still relying on Venice to run the proxy honestly, since it controls the pipe and can block. So the claim is narrow: you own the access and can walk with it, not that you’ve escaped trusting anyone. On the closed subscription there’s nothing to own in the first place.

Both subsidies, and what’s left when they stop

I went further into this in Staking for Inference: Morpheus vs NEAR vs Venice, and what came out of it sharpens the choice above rather than settling it.

The staked-access model runs on somebody else’s money too, though not uniformly. Morpheus pays its compute providers from a token emission pillar rather than from user fees, and NEAR routes your staking rewards straight to NEAR AI as payment for the service, saying so in its own terms. Both of those are printing.

Venice is the exception, and it’s the one I actually use. It serves DIEM credit off its own capacity and wears the cost on its P&L, funded by subscriptions, which is closer to an airline handing out seats than to a protocol minting tokens. What’s emission-funded there is the staking yield, not the compute.

How much the revenue actually covers

Revenue helps on the other two as well, just indirectly and not by much. Buybacks funded by customer money offset some of the dilution the emissions cause, and if a network ever burned as many tokens as it issued, the customer’s cash would have cancelled the tokens the providers were paid in and the revenue would genuinely have funded the compute. Coverage is nowhere near that: roughly 15% at Venice, about 4% at Morpheus, and NEAR’s is held rather than burned.

So the comparison is two subsidies with different funders and different degrees of self-funding, a distinction that carries more weight than the sustainability label ever did.

Anthropic’s is funded by investors and lasts until they stop writing cheques. The DeAI version is funded by issuance, which means token holders cover it through dilution. Hold the token and you’re partly subsidising your own inference, which is a stranger arrangement than it sounds and a more honest one: the people drawing the cheap compute are the people paying for it.

What separates them is what’s in your hand when the subsidy tapers. A promotion ending leaves you with a lapsed seat and a chat history the vendor still keeps. An emission taper leaves you holding the token, and you can still unstake it.

Compute gets dearer either way. Only one of them leaves you with an asset.

That piece also puts a floor under the Venice case specifically. The obligation is capped by design, and buybacks funded by subscription revenue already offset a growing share of it, which is the mechanism that would eventually turn the subsidy into a business. On that measure Venice is further along than the other two.

So, is Claude Max good value?

On cash, obviously, and I’m keeping mine. For raw value it’s the best deal in AI right now, and it isn’t close. I’d tell any heavy user to take it while it lasts.

Just be clear about what “while it lasts” means. The price is a promotion funded by investor money, the privacy is a policy you’re trusting rather than checking, and the access is a monthly rental you can’t hold.

The durable bet is the other stack: open models for the unlimited work, private and attested modes when the content matters, and staked access you own and can carry. It costs close to market rate instead of a subsidised discount, and in exchange it’s private, portable, and yours. The cheapest tokens today and the ones you’ll still own tomorrow come from different places, and I know which one I’m building on.