Claude Code Pricing for Teams: Plans, Tiers & What to Pick

Claude Code has two access models, and the choice shapes everything downstream:

• Subscription (Claude Pro or Max). Each engineer signs in with a Claude account. Billing is a flat recurring fee per person with generous usage allowances that reset on a rolling basis. Simple, predictable, zero infrastructure.
• API (pay-as-you-go). Usage is metered by tokens through the Anthropic Console. You pay for exactly what you consume, with organisation-level keys, spend limits, and usage analytics.

The trade-off is predictability versus control. Subscriptions cap the downside per seat but bundle usage into allowances; the API exposes true consumption but requires someone to own keys and budgets. Many organisations run both: subscriptions for day-to-day engineering, API keys for CI pipelines, automation, and agents that run unattended. Neither choice is permanent — teams commonly start on subscriptions and add API access as automation grows.

Within subscriptions, the meaningful difference between Pro and Max is usage headroom and model access, not features.

• Pro suits engineers who use Claude Code regularly but not constantly — a few substantial sessions a day, mixed with normal editor work. Occasional power users will sometimes hit usage limits during heavy stretches.
• Max exists for engineers whose workflow is built around the agent: long refactoring sessions, multiple parallel sessions, frequent use of the flagship model. Max tiers offer substantially higher allowances and the most generous access to top-end models like Claude Fable 5.

A practical heuristic: if an engineer complains about hitting limits more than rarely, the productivity lost waiting almost certainly exceeds the tier difference. Start most of the team on Pro, upgrade your heaviest users to Max based on observed usage, and revisit quarterly. Current allowances and terms are on Anthropic’s pricing page.

On the API, you pay per token — text in and text out — with rates that vary significantly by model. That makes model selection your biggest cost lever:

• Claude Fable 5, the flagship, commands premium rates and earns them on complex, long-horizon work: large refactors, unfamiliar codebases, architectural changes.
• Claude Opus 4.8 and Claude Sonnet 5 handle the broad middle of engineering work at lower cost.
• Claude Haiku 4.5 is the economical choice for routine, well-specified tasks and high-volume automation.

The other major driver is context: sessions that repeatedly load large files burn tokens fast, while tight CLAUDE.md files and well-scoped tasks keep sessions lean. Prompt caching also reduces the cost of repeated context substantially. Our Claude model guide covers how to match models to task types in detail.

Skip the theoretical modelling and measure instead. A two-week instrumented pilot answers the cost question better than any spreadsheet:

• Pick 5–10 engineers across seniority levels and give them real backlog tickets to work with Claude Code.
• Track usage honestly — via Console analytics on API keys, or by noting how often subscription users approach limits.
• Segment your users. Every team splits into heavy adopters, steady users, and occasional users. Your cost model is that distribution multiplied by the appropriate tier, not headcount times one price.
• Count the offset. The relevant comparison is not tool cost versus zero — it is tool cost versus the loaded hourly cost of the engineering time it saves. For most teams, a single meaningful task delegated per week changes the arithmetic decisively.

Pilot data also settles the subscription-versus-API question empirically rather than by committee.

Uncontrolled AI spend is a leadership problem before it is a finance problem, and the controls differ by access model:

• On the API, the Anthropic Console provides workspace-level spend limits, per-key budgets, and usage dashboards. Set limits before the rollout, not after the first surprising invoice.
• On subscriptions, cost is capped per seat by design — governance is mostly about seat management: reclaim unused seats, right-size tiers against actual usage.
• In both cases, assign an owner. The teams that manage this well treat Claude Code spend like cloud spend: one accountable person, monthly review, model-choice defaults documented in team configuration.

One cultural warning: do not make engineers afraid to use the tool. Penny-pinching model choice on complex tasks produces worse code that costs more in review time than the tokens saved. Optimise the defaults, not each individual session.