Same tokens, same model, up to a 40x price gap: that is Claude pricing in August 2026.

Agentic coding is where large language models found product-market fit: agents burn vastly more tokens than chat, and they became daily drivers for some of the best-paid professionals in the world. A year ago $20 a month was plenty; today $100 is the entry bar for serious work. Marty Kausas, CEO of Pylon, admitted: “I accidentally spent $4,000 in 3 days in Claude Code.” Uber rolled out Claude Code in December 2025 and burned its entire 2026 AI coding budget by April. And it is industry-wide: The Information reports multiple enterprises facing bills 2-3x higher.

Seats with usage included vs pay per token

The same work, priced per seat (Max, Team) and per token (API, Enterprise):

Case study: author’s own 24-hour session, replayed at API list prices Cost over timeline ($ / time) $5 $10 $15 18:14 20:00 21:00 22:00 06:14 08:00 09:00 10:00 8.2h overnight 65 min gap 117 min gap biggest call: a skill load accidentally pulled the whole claude-api reference into context, 326k tokens cached at once for $6.65 up to 5 parallel Opus 5 subagents building features in the background Context length on same timeline (k tokens / time) 250k 500k 0 cache expired compact One project in 24h, $180.91, almost the equivalent of a monthly subscription Cache reads: $85.95 Subagents (Opus 5): $49.73 Cache writes (1h TTL): $27.46 Output: $17.77 Typed or dictated input: $0.01 Naive estimate: $2.24 A real Claude Code day on an internal web app, July 2026. Fable 5 main loop, Opus 5 subagents. Each square is $1 of tokens; every API call re-reads the whole conversation, so columns grow with context. Cache reads · $0.44 · 18:14–18:29Each call re-read the ~73k-token conversation at $1/MSetting up tooling, reading the code Cache writes (1h TTL) · $1.31 · 18:14–18:29New tokens written to the 1h cache at $20/MSetting up tooling, reading the code Cache reads · $7.10 · 18:29–18:44Each call re-read the ~392k-token conversation at $1/MDrafting plans and docs Cache writes (1h TTL) · $7.01 · 18:29–18:44New tokens written to the 1h cache at $20/MDrafting plans and docs Output · $1.04 · 18:29–18:44Model output tokens at $50/MDrafting plans and docs Cache reads · $8.56 · 18:44–18:59Each call re-read the ~452k-token conversation at $1/MDrafting plans and docs, reading the code Output · $0.82 · 18:44–18:59Model output tokens at $50/MDrafting plans and docs, reading the code Subagents (Opus 5) · $0.72 · 18:44–18:592 Opus 5 agents working in parallelDrafting plans and docs, researching online Cache reads · $3.28 · 18:59–19:14Each call re-read the ~470k-token conversation at $1/MDrafting plans and docs Output · $0.50 · 18:59–19:14Model output tokens at $50/MDrafting plans and docs Cache reads · $6.25 · 19:14–19:29Each call re-read the ~482k-token conversation at $1/MDrafting plans and docs Output · $0.55 · 19:14–19:29Model output tokens at $50/MDrafting plans and docs Cache reads · $7.25 · 19:29–19:44Each call re-read the ~320k-token conversation at $1/MDrafting plans and docs, implementing a feature Cache writes (1h TTL) · $2.03 · 19:29–19:44New tokens written to the 1h cache at $20/MDrafting plans and docs, implementing a feature Output · $2.62 · 19:29–19:44Model output tokens at $50/MDrafting plans and docs, implementing a feature Subagents (Opus 5) · $1.71 · 19:29–19:444 Opus 5 agents working in parallelReading the code Subagents (Opus 5) · $0.55 · 19:44–19:595 Opus 5 agents working in parallelReading the code Subagents (Opus 5) · $5.72 · 20:14–20:294 Opus 5 agents working in parallelImplementing a feature Subagents (Opus 5) · $8.45 · 20:29–20:444 Opus 5 agents working in parallelReading the code, implementing a feature Cache reads · $1.64 · 20:44–20:59Each call re-read the ~126k-token conversation at $1/MDrafting plans and docs, testing the app Cache writes (1h TTL) · $2.37 · 20:44–20:59New tokens written to the 1h cache at $20/MDrafting plans and docs, testing the app Output · $0.82 · 20:44–20:59Model output tokens at $50/MDrafting plans and docs, testing the app Subagents (Opus 5) · $1.70 · 20:44–20:593 Opus 5 agents working in parallelReading the code Cache reads · $0.91 · 20:59–21:14Each call re-read the ~155k-token conversation at $1/MImplementing a feature Cache writes (1h TTL) · $0.41 · 20:59–21:14New tokens written to the 1h cache at $20/MImplementing a feature Subagents (Opus 5) · $6.17 · 20:59–21:141 Opus 5 agent working in the backgroundImplementing a feature, testing the app Cache reads · $3.61 · 21:14–21:29Each call re-read the ~182k-token conversation at $1/MImplementing a feature, reading the code Cache writes (1h TTL) · $0.69 · 21:14–21:29New tokens written to the 1h cache at $20/MImplementing a feature, reading the code Output · $0.87 · 21:14–21:29Model output tokens at $50/MImplementing a feature, reading the code Subagents (Opus 5) · $1.24 · 21:14–21:291 Opus 5 agent working in the backgroundReading the code Cache reads · $9.10 · 21:29–21:44Each call re-read the ~235k-token conversation at $1/MImplementing a feature Cache writes (1h TTL) · $1.52 · 21:29–21:44New tokens written to the 1h cache at $20/MImplementing a feature Output · $2.62 · 21:29–21:44Model output tokens at $50/MImplementing a feature Cache reads · $6.98 · 21:44–21:59Each call re-read the ~321k-token conversation at $1/MImplementing a feature Cache writes (1h TTL) · $1.52 · 21:44–21:59New tokens written to the 1h cache at $20/MImplementing a feature Output · $1.69 · 21:44–21:59Model output tokens at $50/MImplementing a feature Subagents (Opus 5) · $4.17 · 21:44–21:592 Opus 5 agents working in parallelImplementing a feature, reading the code Cache reads · $12.75 · 21:59–22:14Each call re-read the ~389k-token conversation at $1/MImplementing a feature Cache writes (1h TTL) · $1.53 · 21:59–22:14New tokens written to the 1h cache at $20/MImplementing a feature Output · $2.40 · 21:59–22:14Model output tokens at $50/MImplementing a feature Cache reads · $0.85 · 22:14–22:29Each call re-read the ~424k-token conversation at $1/MShipping commits Cache reads · $0.67 · 06:14–06:29Each call re-read the ~72k-token conversation at $1/MTesting the app Cache writes (1h TTL) · $1.07 · 06:14–06:29New tokens written to the 1h cache at $20/MTesting the app Cache reads · $0.74 · 06:29–06:44Each call re-read the ~97k-token conversation at $1/MPlanning with subagents, reading the code Output · $0.73 · 06:29–06:44Model output tokens at $50/MPlanning with subagents, reading the code Subagents (Opus 5) · $2.68 · 06:29–06:443 Opus 5 agents working in parallelReading the code Cache reads · $7.71 · 06:44–06:59Each call re-read the ~155k-token conversation at $1/MImplementing a feature, reading the code Cache writes (1h TTL) · $1.21 · 06:44–06:59New tokens written to the 1h cache at $20/MImplementing a feature, reading the code Output · $1.00 · 06:44–06:59Model output tokens at $50/MImplementing a feature, reading the code Subagents (Opus 5) · $1.25 · 06:44–06:591 Opus 5 agent working in the backgroundTesting the app Cache reads · $2.49 · 06:59–07:14Each call re-read the ~194k-token conversation at $1/MTesting the app Cache writes (1h TTL) · $0.69 · 06:59–07:14New tokens written to the 1h cache at $20/MTesting the app Output · $0.86 · 06:59–07:14Model output tokens at $50/MTesting the app Subagents (Opus 5) · $14.87 · 06:59–07:145 Opus 5 agents working in parallelImplementing a feature, reading the code Cache reads · $4.04 · 07:14–07:29Each call re-read the ~214k-token conversation at $1/MShipping commits, reading the code Output · $0.41 · 07:14–07:29Model output tokens at $50/MShipping commits, reading the code Cache reads · $0.69 · 09:44–09:59Each call re-read the ~223k-token conversation at $1/MTesting the app Cache writes (1h TTL) · $4.05 · 09:44–09:59New tokens written to the 1h cache at $20/MTesting the app 65 min without a call1h cache expired: resume re-wrote 93k tokens for $1.87 117 min without a call1h cache expired: resume re-wrote 201k tokens for $4.02 489 min without a call1h cache expired: resume re-wrote 51k tokens for $1.01Also compacted: 424k context summarized down to 71k CompactContext summarized from 504k down to 56k tokens

These are list-price counterfactuals, not real bills. Self-reported numbers on Hacker News land in the same 12-50x band: from $1,850 a month at half the limits of a $100 Max 5x, up to “$15k in the past 30 days” on roughly $300 of subscriptions.

Organizations where many people use Claude irregularly, or for non-agentic work, see more favorable math. But among the companies I have talked to, everyone who moved from seats to per-token Enterprise saw the bill at least double, and most reported roughly 3x.

The realistic price for agentic coding

Full details on Anthropic’s pricing page. The same models can also be bought per token through the major clouds (AWS Bedrock, Google Vertex AI, Microsoft Foundry). The non-obvious part of per-token billing is the cache: agents re-read their whole context on every step, so long sessions are mostly prompt-cache traffic. Across all my sessions, 82% of the API-equivalent cost was cache related.

How companies cap AI spend

A per-engineer cap with an override path is the current norm among heavy adopters:

These are the heavy adopters, not the average: in the Pragmatic Engineer survey, the typical company-funded plan is $100-200 per engineer per month, and Gartner found nearly a quarter of tech leaders spending $200-500 per developer per month on AI coding tokens, with only about 6% above $2,000.

Why Anthropic leaves money on the table

The best explanation: the seat plans are a subsidy and market segmentation. Cheap seats let people learn how great agentic coding is, get good at it, and shape both the product and their own preferences, so that adoption later happens at much bigger scale. Depending on what you count, subscriptions bring in only 5-15% of Anthropic’s revenue: SemiAnalysis estimates consumer subscriptions alone at ~5%, while Sacra puts all subscription plans combined at 10-15%. The metered side, dominated by enterprises, is where the money is made.

So enjoy the buffet while it is open. Those multi-billion-dollar data centers full of chips are not cheap: data center capex surged 57% in 2025 and is forecast to top $1 trillion in 2026. Anthropic already tried to meter Fable 5 for subscribers before settling on including it in Max plans at up to half the weekly limits, and it will likely try again when competition allows.

What Claude Enterprise actually is

Claude Enterprise today is a roughly $20 per seat license that includes no usage at all. Every token bills at standard API rates on top: self-serve customers prepay into a shared credit pool, sales-assisted customers get monthly invoices in arrears. There is no published volume discount, though there are rumors of discounts to smooth the price-hike transition.

When Claude Enterprise is a must

Seat pricing stops working in three places:

• The 150-seat limit of Claude Team. The one hard block, and a deliberate cliff. Past it you must move to Enterprise, where Anthropic earns the most. Big companies tend to be more productive with the tool and can afford higher prices.
• Enterprise-only features. SSO and SCIM provisioning, audit logs, per-seat spend limits, and compliance and analytics APIs for exporting per-user usage live only on the Enterprise tier. Anthropic also has a habit of parking capabilities there: a longer context window was an Enterprise exclusive when the tier launched, and today Mythos, the ungated sibling of Fable for approved use cases such as cybersecurity, goes only to approved enterprise customers.
• Gotchas of consumer plans at work. Individual Max accounts mean individual invoices, which procurement hates. Consumer plans lack commercial terms and a data processing agreement, the model can be trained on your data if you click the consent prompt the wrong way, and in the EEA and Switzerland the consumer terms even carry a non-commercial-use clause, mostly a liability limitation (discussion).

How to manage Claude Code costs

Start with Claude Max. Give people the tool and think of it as training, like a conference ticket. Many early agentic projects underdeliver against the executive scope. That is fine: using the tool is the only way to learn it. Fund it as a perk or reimburse the subscription, side projects included: you are learning on the subsidized tier instead of at API prices. Cancel the seats nobody uses.

Then Team, then Enterprise. Grow organically from Max to Team, and move to Enterprise only once you genuinely outgrow 150 seats. If your company has multiple divisions, buying multiple Team workspaces is also a great path; centralization is an anti-pattern.

Prefer more accounts over API overages. When someone hits the limits and cannot work, the first instinct should be another subscription, not API tokens for the overflow: credits bill per token at roughly the prices above, and the up-to-30% prepaid-bundle discount does not close a 15x gap. Plenty of people run more than one: a Max for experimentation and a Team seat for commercial work.

Use a second vendor to stay under 150. Some companies split deliberately: core engineers get Claude Code, everyone else uses OpenAI Codex or another provider. Both groups keep seat pricing, and the multi-vendor setup is negotiation leverage besides.

Get good at cost monitoring before Enterprise. Once you pay per token, per-seat budgeting is crude. Agentic spend varies wildly, and your biggest spenders are often the people leveraging the tool the most, which is exactly the behavior you want. At the same time, we are all still bad at measuring the actual impact. A weekly budget per person, actually measured, plus a lightweight process to raise limits for the people who deserve it, beats any flat cap.

Claude Code budgeting antipatterns

The powerful tool with a tiny budget. If you hand people the latest Fable model with ultracode multi-agent tooling and a small allowance, they will burn the weekly budget in a few hours. That is a terrible first experience. Better to default to a slightly weaker model, or a lower reasoning effort, that people can use all week than something that regularly cuts them off mid-task. Claude Pro and small Team plans are the official-packaging version of the same mistake: at those limits agentic coding is barely usable, and Claude Code likely stays on Pro mostly for PR reasons (Anthropic tried removing it in April 2026 and reversed within a day after backlash).

If money is short, restricting Fable and defaulting to Opus helps, but there is a limit to how much you can save inside Anthropic’s price list: Anthropic in 2026 is Apple, not cheap, but many like it the most. There is a whole competitive field beyond it, including cheap Chinese models (DeepSeek, Kimi, Qwen, GLM) hosted by Western companies and alternative harnesses (Pi, OpenCode), but that deserves a blog post of its own.

Restricting everyone to Haiku. The cheapest model has real uses: permission checks and simple mechanical tasks, where it is unbeatable per dollar. For long-horizon agentic coding it is not great, and people restricted to it get a poor experience of the whole technology.

The free-for-all. The opposite failure mode is letting everybody burn the whole budget in a short time with no visibility, and then discovering the money is gone without knowing what it went to. That is how Uber blew its 2026 AI coding budget in four months, with its CTO telling The Information: “I’m back to the drawing board, because the budget I thought I would need is blown away already.” The typical answer is a per-engineer weekly/monthly cap.

Heavy, centralized procurement. The AI world moves fast and repricing already cascades through it: plans change, promos lapse, models leapfrog each other within months. You cannot watch and wait from the sidelines. Patterns that work elsewhere, like centralizing procurement to save money, can backfire badly here.

Token economics matters

Agentic coding gives companies a huge transformation potential, and we are still early in discovering what it can do. Think of how the iPhone enabled businesses like Uber and DoorDash years after launch; with AI we will keep making discoveries like that for years.

But the economics matter now. These days you can spend more on tokens than on the engineers driving them, and spend can scale to almost any number if nobody is watching. The tooling for visibility still lags what enterprises need.

If you are wrestling with an AI bill, I would love to hear your story: contact me through e-mail or a call.