What Claude Code Actually Costs in 2026: Per Token, Per Month, and Two June Deadlines

"Claude code cost per month," "claude code cost per token," "claude code cost vs cursor," "claude code cost reddit." The autocomplete tells the story: people are not asking whether to use Claude Code, they are trying to predict the bill, and in June 2026 that prediction got harder. Two pricing changes land nine days apart, the frontier model bills at double the previous flagship, and the meter that matters, cost per task, does not appear on any plan page. Here is the complete cost picture, sourced, plus the part the pricing tables leave out: how to make a $200 ceiling behave like one.

The two ways to pay, in numbers

Claude Code is billed either as a flat subscription or as metered API usage. Most teams run both: subscriptions for interactive coding, API keys for automation.

Subscription plans (flat monthly):

• Pro, $20/month. Interactive Claude Code, modest daily limits, fine for individual side-project and light professional use.
• Max 5x, $100/month. Roughly five times Pro's usage allowance, the common pick for full-time engineers living in the agent.
• Max 20x, $200/month. Twenty times Pro, for heavy multi-session and long-context work.
• Team Premium, $100/seat/month (5-seat minimum, so $500/month floor). Adds team management and pooled administration.

API pay-per-token (per million tokens, input / output):

• Haiku 4.5: $1 / $5. The workhorse for high-volume, well-scoped tasks.
• Sonnet 4.6: $3 / $15. The default balance of capability and cost.
• Opus 4.8: $5 / $25. The prior flagship, still the value play for hard reasoning.
• Fable 5: $10 / $50. The current frontier, priced at exactly double Opus. The 1M-token context bills at standard rates, with no discount tier.

The subscription-versus-API decision is a volume calculation. Below roughly a few hours of daily agent use, the flat $20 to $200 plans are dramatically cheaper than the equivalent token spend. Above it, or for any unattended automation, the API meter is the only model that scales, and the only one where a runaway loop can produce a four-figure surprise. We walk the crossover point in the cost-per-engineer FinOps playbook.

To make the abstraction concrete, price a single nontrivial coding task. A realistic agentic edit, reading a few files, reasoning, writing a diff, and revising, might consume 200K input tokens (much of it cached repository context) and 30K output tokens across its turns. On Sonnet 4.6 that is roughly $0.60 input plus $0.45 output, about $1.05 a task. On Opus 4.8 the same shape runs near $1.75. On Fable 5 it is closer to $3.50. Run forty such tasks a day and you are at roughly $42, $70, or $140 daily by model, which is exactly why the flat Max plans look like a bargain right up until an unattended job loops and the API meter, with no soft limit, keeps going long after a human would have stopped.

Deadline one: June 15 unbundles the Agent SDK

Starting June 15, 2026, Claude Agent SDK usage and headless claude -p invocations no longer count toward your Claude plan's usage limits. Instead they bill against a separate, API-rate credit pool: $20/month of credits on Pro, $100 on Max 5x, $200 on Max 20x. Your interactive subscription limits stay the same and stay reserved for hands-on use of Claude Code, Cowork, and claude.ai.

Read past the mechanics and the intent is clear, and it matches the pattern across the whole industry right now. High-cost, high-volume, unattended workloads, exactly what the Agent SDK enables, are being moved out of all-you-can-eat subscriptions and onto meters where they earn margin instead of burning it. This is the same template behind the June 15 credit split and Claude 4 retirement, and it reads differently once you know Anthropic filed its S-1 the week before. The practical effect: if you run CI pipelines, scheduled agents, or batch jobs through the SDK, those costs are now visible as a separate line, and they can exceed the bundled credit pool quietly.

Deadline two: June 22 moves Fable 5 to usage credits

Fable 5 has been included on Pro, Max, Team, and seat-based Enterprise plans during its rollout window, but that window closes on June 22, 2026. After June 23, Fable 5 access is expected to require usage credits rather than ride the flat plan, unless capacity allows an extension.

The detail that catches teams off guard: Fable 5 draws down plan usage faster than Opus, so even inside the included window the headroom is smaller than the plan tier suggests. A Max 20x user who switched their default model to Fable 5 is consuming the allowance at roughly twice the rate, which is why some heavy users hit limits weeks earlier than they expected. The full burn-rate mechanics are in how Fable 5 usage limits actually work. The move that protects you is mundane: pin your default to the cheapest model that passes your evals, and reserve Fable 5 for the tasks that genuinely need it.

Claude Code cost vs Cursor, Copilot, and Gemini

The comparison searches are constant, so here is the honest framing. The headline subscription numbers are close enough across the field ($20 entry tiers, $100 to $200 power tiers) that the sticker price is not where the difference lives. The difference lives in how each tool meters overage and how much work it does per dollar on your tasks:

• vs Cursor. Cursor's usage-based overage is where its bills get unpredictable; one team reported $1,400 in a single hour tagging 87 tasks. The mechanics of its overage model are in Cursor usage-based pricing explained.
• vs GitHub Copilot. Copilot moved to usage-based billing with per-model credit multipliers; the premium-request math is what determines whether its $10 to $39 tiers hold, covered in Copilot AI credits cost per model.
• vs Gemini. Gemini's API tiers and the retired free tier change the calculus for high-volume automation; see Gemini API spend caps and tiers and the Pro vs Flash-Lite cost-quality split.

The only comparison that survives contact with your actual workload is cost per completed task, measured on your repos, not a vendor's benchmark slide. A tool that is cheaper per token can be more expensive per task if it needs more turns to finish, which is the entire argument for benchmarking workhorse models by cost per task.

How to make a $200 ceiling actually behave like one

The reason "what does Claude Code cost" has no clean answer is that the flat plans have soft limits and the API has none. The same setup produces wildly different bills depending on four controls, and all four are things you implement, not things you buy:

1. Measure cost per task, per model. The bill is price times volume, and the only number that catches sideways cost creep, longer contexts, more turns, model-mix drift, is cost per completed task at constant work. Teams without it absorb every increase into "we are just using it more" and never see the leak. This is the measurement discipline that makes everything else possible.
2. Route by task, not by habit. Most coding tasks do not need the frontier. A router that sends routine work to Haiku or Sonnet and escalates only hard problems to Opus or Fable 5 is the single biggest lever on the monthly number, and after June 22 it is also your Fable-5-credit defense.
3. Cap the unattended paths. Interactive use self-limits because a human is watching. The Agent SDK and claude -p jobs do not, which is exactly why they got their own meter on June 15. Per-job and per-pipeline hard caps and kill switches are the difference between a bad night and a bad quarter.
4. Watch the meter weekly, not at invoice time. A cost-per-engineer dashboard turns a quarterly surprise into a Tuesday conversation. The full operational setup is the cost-per-engineer playbook.

The honest take

Claude Code in 2026 is still one of the best dollar-for-output deals in coding, and the flat plans remain genuinely cheap for interactive work. But the question people are actually asking, what will it cost me, no longer has a price-page answer, and June made that official: the Agent SDK is on its own meter, Fable 5 is moving to credits, and the frontier costs double the last flagship. The teams that stay in budget are not the ones on the cheapest plan. They are the ones who measure cost per task, route the easy work down, cap the unattended jobs, and look at the meter before the invoice does. Pick the plan that fits your interactive use, then put a meter on everything else, and the $200 ceiling will mean what it says.