We'll estimate what an agent session costs, explain why prompt caching matters so much, and choose between a subscription and the API for a given use. Every price here comes from the vendors' own pages, checked on 8 November 2026. Prices change often, sometimes monthly. The arithmetic still works when the prices change. The reading is module 12. Students, read Foundations, the price snapshot, and the worked example. Engineers, all of In practice. Leads, the worked example, cost per task, team budgeting, and the counterpoints to my take.

Those two numbers come from the worked example later: one hour, 60 model calls, Claude Opus 5.5 list prices as of November 2026. The only difference between them is prompt caching. Compare cost per finished task and choose a billing model that fits your use. Subscriptions suit individual interactive work. The API suits anything that runs in production, in CI, in batch, or for other users.

Chat and pro subscriptions run roughly $20 to $200 a month for individuals, more for some tiers. They include the lab's coding agent: Claude Code with a Claude plan, Codex with a ChatGPT plan. The cap is stated as a multiple ("5x Pro usage"), not as a number of tokens. Tool vendors buy model access from the labs and resell it inside their product. Some include their own or partner models in the flat fee and charge frontier lab models at API list price. With the API, you get a key and pay for exactly what you use, with no cap beyond rate limits. You can use any harness, build your own tools, run in CI, and ship features to customers.

A token is roughly three-quarters of an English word on average. Prices are quoted per million tokens, written "MTok". First, tokenizers differ. The same text becomes a different number of tokens on different models. Anthropic notes that its Claude 4.7-and-later tokenizer produces about 30% more tokens for the same text than earlier Claude models. A lower per-token price on a wordier tokenizer can cost the same. Second, thinking is output. A reasoning model that thinks for 5,000 tokens before writing a 300-token answer bills you for 5,300 output tokens. Effort or thinking-budget settings change cost directly. Some bills also add surcharges for long prompts above a threshold on some models, and for regional data residency (Anthropic's US-only inference is 1.1x).

This explains almost every surprising AI bill.

A model API is stateless: it remembers nothing between calls. A coding agent works in a loop. The model asks to run a tool, the harness runs it, and the result goes back to the model along with everything that came before. Every turn re-sends the full context. The numbers in the block are the assumptions from the worked example: a 20,000-token start (system prompt, tool definitions, project instructions) and 2,500 tokens added per call. A full test log, a large file read, or a web page can add 10,000 to 100,000 tokens in one step. Retries and dead ends re-send everything again. Subagents each start their own context, often re-reading the same files. Parallel agents multiply all of it. Thinking adds output tokens on every turn. Keep this in mind for the discussion: how would you design a harness to keep cost closer to linear?

Caches match from the start of the prompt. Anything that changes early, such as a timestamp in the system prompt or a reordered tool list, invalidates everything after it. Put stable content first: tools, system instructions, project docs, then the conversation. On Anthropic, mark cache breakpoints or turn on automatic caching. The cache economics, as of November 2026. Anthropic charges 1.25x input to write a 5-minute cache entry and 2x for 1 hour. Reads cost 0.1x base on most models and less on the newest: 0.05x on Opus 5.5 and Sonnet 5.5, 0.025x on Fable 5.1. A 5-minute write pays for itself after one read. OpenAI caches automatically with no write premium. Google charges for cache storage per hour on models that offer explicit caching. For agent workloads, most input tokens are cache reads, so compare the cache-read price first, not the headline input price.

Go through the arithmetic slowly. It's the core of the session. Context on call i is 20,000 + 2,500 × (i − 1), so call 60 sends 167,500 tokens. Total input = 60 × 20,000 + 2,500 × (0 + 1 + … + 59) = 1,200,000 + 2,500 × 1,770 = 5,625,000 tokens. Total output = 60 × 1,500 = 90,000 tokens. Swap in your own numbers from your logs. The module's Try it shows how to get them.

With caching, each call reads the previous prefix from the cache and writes only the new part. Cache writes: 20,000 tokens on the first call plus 2,500 on each of the 59 later calls = 167,500 tokens. Cache reads: everything else, 5,625,000 − 167,500 = 5,457,500 tokens. Look at what moved. Input fell from $22.50 to about $1.93, while output stayed at $1.80. With caching on, output is now the largest single line.

Same token counts as the previous slide. Cache-read prices used: top tier $0.25, frontier $0.20, mid $0.10 per million tokens. These match Anthropic's Fable 5.1, Opus 5.5, and Sonnet 5.5 as of November 2026. Other vendors price cache reads and writes differently, so recompute with the price table in the module before comparing vendors. The monthly column assumes one agent. Parallel agents multiply it: three agents at once is roughly three times the bill.

This is why a $100–$200 subscription feels like a bargain to heavy users. For light users, Anthropic's own docs use an example one-hour session of 50,000 input and 15,000 output tokens, which costs well under a dollar on Opus 5.5. If you use an agent for a few short tasks a week, pay-as-you-go may beat a $20 plan. On cache misses, Anthropic's default cache lifetime is five minutes. Step away longer and the next call re-writes the whole prefix: 150,000 tokens at $5 per million is $0.75. Harnesses can use the 1-hour cache at twice the write price. Measure instead of guessing. The ccusage tool, in the Try it slide, reads the usage data agent CLIs store locally.

Now let's work out how to pay for it.

Anthropic renamed Max 20x to Ultra on 30 October 2026 at the same $200 price; the old Max 5x is now just Max. That was reported by multiple outlets and is consistent with the pricing page. Ultra also includes $200 a month in API credits. Pro is $17 a month billed annually. OpenAI added the $100 Pro tier on 9 April 2026 and the $500 tier at the end of September 2026, when it also lowered the allowance for new $200 subscribers (reported). Free and Go ($8) tiers give limited Codex access on a small model. Team seats: Anthropic Team is $25 a month standard ($20 annual) and $125 premium ($100 annual). OpenAI Business is $25 ($20 annual). On the tools, Copilot allowances are in "GitHub AI Credits" (1 credit = $0.01). Cursor rebuilt its lineup around a shared subscription with xAI in late October 2026 (reported); its Plus plan is limited to Grok models, and Claude, GPT, and Gemini are billed at API list price. Tool pricing moves fast.

Whether vendors make or lose money on heavy users isn't public. On weekly limits, Anthropic announced on 28 July 2025 that weekly limits would start 28 August 2025 for Pro and Max. It cited users running Claude Code "continuously in the background, 24/7" and account sharing, and estimated fewer than 5% of subscribers would be affected. Its estimates then: Pro about 40–80 hours of Sonnet 4 a week; the $200 Max tier about 240–480 hours of Sonnet and 24–40 hours of Opus. Reporters calculated the "20x" plan gave about 6x Pro's Claude Code hours. On third-party tools, as of November 2026, Anthropic says plan sign-in is for Claude Code and its own apps; products, including ones built with the Agent SDK, must use API keys. It began blocking tools that spoofed the Claude Code harness in January 2026. Running the unmodified Claude Code binary with your own subscription is allowed. OpenAI lets you sign into Codex with a plan or an API key, and in September 2026 opened "Sign in with ChatGPT" so subscribers can spend their allowance in partner tools. These policies changed several times in a year. If your workflow depends on a consumer plan inside a third-party tool, have a fallback.

Z.ai's GLM Coding Plan starts at $18 a month for the Lite tier, with 5-hour and weekly credit limits and half-price off-peak usage. Moonshot's Kimi Code Plan runs its Kimi K3-series models. MiniMax sells an M Plan for its M3-series models. They plug into popular harnesses through API-compatible endpoints. Read the data terms. The "open" part only helps you if you might self-host later (module 04). With routing, send renaming, boilerplate, log summaries, and issue classification to a small, cheap model; keep planning, hard debugging, and review on the frontier model. Many harnesses let you set a different model for subagents. Local models have zero marginal token cost but real hardware and electricity cost, and today's local models are well behind the frontier for agentic coding (module 04).

The number that matters is what it costs to get a correct, merged change. Same growth assumptions as the hourly example: 20,000-token start, 2,500 tokens added per call, 1,500 output tokens per call. Full prices: frontier $4 input, $5 cache write, $0.20 cache read, $20 output; budget $1, $1.25, $0.10, $5. Without caching: $7.65 vs $12.49. Final context: 92,500 vs 242,500 tokens. Three times the turns means about seven times the input tokens, because context grows every turn. The budget model also used three times as much of your attention. That costs more than anything else in the table. The reverse also happens. For well-specified, repetitive work, such as applying the same refactor to 200 files, a small model finishes in the same number of turns, and the 20–40x price gap wins. Thinking budgets follow the same logic: more reasoning per turn often means fewer turns. Try a task at two effort levels and compare total cost.

For production, you need an API key by contract, and you want per-request metering anyway. CI and automation run unattended and shared, which consumer terms don't allow. On data terms, commercial API terms default to no training on your data. Some organizations require zero data retention (ZDR), which is available on commercial terms and not on consumer plans.

All major consoles have spend limits and alerts. Anthropic Enterprise and Copilot let admins set per-user and organization limits. Separate keys let you see who spent what. Trimming tool outputs means truncating logs and summarizing large files before they enter the context. The weekly cost review matters most in the first month of any new agent workflow.

We'll work out what a lead can tell leadership about costs. I'll also give my opinion on subscriptions.

Show your assumptions; leadership will change them, and that's the point. Loaded cost covers benefits, taxes, equipment, and overhead. The hourly figure works out to about $105; round to $100. A $20 seat costs about 12 minutes of an engineer's time a month. The 10-person team's $24,000 a year is about 13% of one loaded engineer. The weak link is self-reporting. In METR's 2025 randomized study, experienced open-source developers using early-2025 AI tools took 19% longer on their tasks while believing they were 20% faster. Tools have changed a lot since, but the lesson about self-reports holds. Measure things you can count before and after: cycle time, review turnaround, escaped defects, rework. On seats vs pooled, seats are predictable but you pay for light users, and heavy users still hit caps. Pooled usage tracks real use but needs budgets and alerts. Many teams mix: a standard seat for everyone, premium seats or pooled API credits for the heaviest users and for automation. Before a rollout, get in writing: training on your content, retention and ZDR, data residency and its price, SSO, SCIM, audit logs, admin spend limits, per-user cost attribution, regulated-data coverage, and renewal pricing and notice for limit changes.

This is my opinion, from daily use. The module's arithmetic supports it for full-time use of frontier models. Even with caching, a single heavy user can spend several hundred dollars a month at API prices, and parallel agents multiply that. There are counterpoints worth considering too.

On caps, heavy parallel use may need the top tier, extra credits, or a second vendor. The terms keep moving. Weekly limits (2025), third-party harness restrictions (2026), and reduced allowances at existing price points (2026) all happened within about a year. Consumer subscriptions are for an individual using the vendor's tools. Production features, CI bots, and shared automation need the API, by contract. The API lets you switch models and harnesses freely and route cheap work to cheap models. For teams, pooled usage plus good caching and routing can beat buying top-tier seats for everyone, especially when usage is uneven. Which of these applies to you?

ccusage reads the usage data that Claude Code, Codex, and many other agent CLIs store locally, and prints token counts with API-equivalent costs. Record total input, cache-read, cache-write, and output tokens, and the API-equivalent cost. Compare that with your subscription price, then recompute it for a model one tier cheaper using the price table. For your most expensive session, decide whether length, big tool outputs, retries, or parallel agents drove it, and write down one habit change that would have halved it. For the team exercise, assume 4 heavy users, 6 moderate, 2 light, plus a PR-review bot on every pull request (about 40 a day). Price three options: all top-tier seats; standard seats plus pooled API credits for the heavy users and the bot; API only, with caching and routing. State every assumption (tokens per session, sessions per day, cache hit rate, PR size). Compute the break-even productivity gain against loaded engineer cost, and list the data-terms questions you'd send to procurement.

Fuller versions from the module: 1. Why does doubling the length of an agent session more than double its cost? How would you design a harness to keep costs closer to linear? 2. A model costs 75% less per token but takes three times as many turns. When is it still the right choice? 3. Should a university course provide students with API keys, subscriptions, or neither? What about the students who can't afford their own? 4. Self-reported productivity gains have been shown to be unreliable. What would you measure to decide whether a $200/month seat is worth it? Two more if time allows: As a team lead, how much should you rely on subscription pricing in a two-year plan? Is it fair for a vendor to restrict a paid subscription to its own tools?

The module links every vendor pricing page used here. Prices were checked on 8 November 2026; re-check them before quoting any number from this deck later.