What Claude Actually Costs a Team in 2026, and How Memory Cuts the Bill

TL;DR

• As of October 2026, Claude Team costs $25 per Standard seat or $125 per Premium seat on monthly billing, and $20 or $100 billed annually, for 2 to 150 seats; Enterprise is $20 per seat per month billed annually, with a 20 seat minimum, plus usage at API rates (claude.com/pricing).
• API tokens are billed per million: Opus 5.5 at $4 input and $20 output, Sonnet 5.5 at $2 and $10, Haiku 5.5 at $0.10 and $0.50 for prompts up to 100,000 tokens ($0.50 and $2.50 above that), Haiku 4.5 at $1 and $5, and Fable 5.1 at $10 and $50 (Anthropic pricing docs).
• Output tokens cost 5x input on every current model, and Claude 4.7 and later models count approximately 30% more tokens for the same text, so a lower headline rate can still produce a higher bill.
• Prompt caching (cache reads at 0.1x input, 0.05x on Opus 5.5 and Sonnet 5.5), the 50% Batch API discount and routing work to Haiku 5.5 each lower the price of tokens, but none of them lowers how many tokens each call carries.
• Shared organization memory removes repeated context at the source: Sentra measured 41.2% fewer tokens and 72.6% lower model cost on Terminal-Bench 2.1, a Sentra internal evaluation pending official verification, run with Codex CLI and GPT-5.5 xhigh rather than Claude.

What Claude actually costs, plan by plan

The individual tiers set the entry price. Free costs nothing and covers everyday questions under limits that reset every five hours. Pro runs $20 a month, or $17 a month billed annually ($200 up front), and adds Claude Code, Research and more Claude models. Max sits above Pro at $100 a month for 5x Pro's usage per session and $200 for 20x, billed monthly only, aimed at people who hit Pro's ceiling daily. Sentra, the managed organization memory for teams and agents, publishes this guide because its measured effect on token bills, covered below, only makes sense against the current rate card. As of October 2026, source: claude.com/pricing. Seat prices for every Claude plan.

Team pricing is published, so there is no need to reconcile third-party snapshots. Standard seats cost $20 per seat per month billed annually or $25 monthly, and Premium seats, with five times a Standard seat's usage, cost $100 or $125 (claude.com/pricing). Team needs at least 2 members and tops out at 150, and both seat types include Claude Code and Opus. Enterprise adds role-based access, SCIM, audit logs, the Compliance API, custom data retention, IP allowlisting and a HIPAA-ready option, and it bills every token at API rates on top of the seat.

The context window is a cost driver, not a comfort feature. Every Claude plan now lists a context window of up to 1M tokens, varying by model, and on the API Claude 4.6 and later models other than Haiku 5.5 bill a 900,000 token request at the same per-token rate as a 9,000 token one, while Haiku 5.5 charges five times its base rate once a prompt passes 100,000 tokens (Anthropic pricing docs). A larger window lets a session hold more history, and that history gets re-read on every call, so the seat price you budget against is only the floor. The real bill depends on how many tokens each conversation carries, which the plan tier alone never tells you.

How API token pricing actually works

The API charges you per million tokens, and on every current model output costs five times more than input. As of October 2026, source: claude.com/pricing and Anthropic's pricing docs, the rates are these.

Context length multiplies the input side because every token you send in the prompt gets billed on every call. When a conversation history grows or you paste in long files, that input count climbs with each turn, and the meter runs on the full window each time.

Three workloads show the range at current rates. A startup chatbot on Sonnet 5.5 at 5M input and 2M output tokens a month costs (5 × $2) + (2 × $10) = $30, or about $22.40 if 4M of those input tokens are cache reads at $0.10, before cache writes. An enterprise knowledge assistant moving from Opus 5 to Opus 5.5 at 10M input and 4M output drops from (10 × $5) + (4 × $25) = $150 to (10 × $4) + (4 × $20) = $120 a month, a 20% cut, because both models use the same newer tokenizer. High-volume content generation at 20M input and 10M output costs (20 × $1) + (10 × $5) = $70 on Haiku 4.5, or $35 with the Batch API, and on Haiku 5.5, with every prompt under 100,000 tokens, it costs (20 × $0.10) + (10 × $0.50) = $7, or $3.50 with the Batch API.

Four modifiers raise costs without warning. Fast mode on Opus 5.5 charges 2x standard pricing ($8 input and $40 output per million), fast mode on Opus 5 and Opus 4.8 charges $10 and $50, and fast mode is not available with the Batch API. A one-hour cache write costs 2x the base input rate against 1.25x for five minutes. US-only inference on Claude 4.6 and later models adds a 1.1x multiplier to every token category. And the newer tokenizer on Claude 4.7 and later counts more tokens for the same text, covered in the next section.

Why did our Claude API bill go up after moving to a newer model?

A Claude bill can rise after an upgrade to a model with a lower headline rate because Claude 4.7 and later models use a newer tokenizer that produces approximately 30% more tokens for the same text, and Anthropic says the exact increase depends on the content and workload shape (pricing docs). The rate per token falls, but the number of tokens rises, and the bill is the product of the two.

• A workload of 10M input and 4M output tokens on Opus 4.6 costs (10 × $5) + (4 × $25) = $150, and the same text on Opus 5.5 at about 13M input and 5.2M output tokens costs (13 × $4) + (5.2 × $20) = $156, which is 4% more despite a 20% lower rate.
• The same workload on Sonnet 4.6 costs (10 × $3) + (4 × $15) = $90, and on Sonnet 5.5 at 13M and 5.2M tokens it costs (13 × $2) + (5.2 × $10) = $78, which is still 13% less because the rate cut is larger than the token increase.
• A move between two models that already share the newer tokenizer, such as Opus 5 to Opus 5.5, passes the full rate cut through, because the token count does not change.

The decision rule: before comparing models by sticker price, count tokens on a sample of your own prompts with each model, then multiply by its rate.

What does Claude cost a 10-person and a 50-person team per month?

Under the seat mixes and API workloads below, a 10-person team pays about $600 a month and a 50-person team about $4,200 a month, at October 2026 prices with seats on annual billing. Seats cover people using Claude and Claude Code; the API workload stands for automated agents and pipelines billed by token on Sonnet 5.5 at $2 input and $10 output per million.

• A 10-person team with eight Standard and two Premium seats pays (8 × $20) + (2 × $100) = $360 a month on annual billing, or (8 × $25) + (2 × $125) = $450 on monthly billing.
• The same 10-person team running 80M input and 8M output API tokens a month pays (80 × $2) + (8 × $10) = $240, for a combined $360 + $240 = $600 a month.
• A 50-person team with 40 Standard and 10 Premium seats pays (40 × $20) + (10 × $100) = $1,800 a month on annual billing, or (40 × $25) + (10 × $125) = $2,250 on monthly billing.
• The same 50-person team running 800M input and 80M output API tokens a month pays (800 × $2) + (80 × $10) = $2,400, for a combined $1,800 + $2,400 = $4,200 a month.
• If 480M of the 50-person team's 800M input tokens are cache reads at $0.10 per million, its API line falls to (320 × $2) + (480 × $0.10) + (80 × $10) = $1,488, before cache writes.
• On Enterprise instead of Team, the 50 seats cost 50 × $20 = $1,000 a month and all usage, including Claude and Claude Code usage by people, is metered at API rates on top.

At a 10-to-1 input-to-output ratio, Sonnet API spend overtakes seat cost above 120M input tokens a month for the 10-person team and above 600M for the 50-person team. Past that line, token volume per request matters more than the seat mix.

Every real way teams cut Claude token spend

Five techniques cut Claude token spend in practice, and each carries a discount you can model from Anthropic's published rates before you commit to it.

• Model routing cuts the rate itself: Sonnet 5.5 costs half of Opus 5.5 and Haiku 5.5 costs 97.5% less on prompts up to 100,000 tokens, so a 70/20/10 Haiku, Sonnet and Opus 5.5 mix has a blended input rate of (0.7 × $0.10) + (0.2 × $2) + (0.1 × $4) = $0.87 per million, 56.5% below all-Sonnet and 78.25% below all-Opus, provided the cheaper models meet each task's bar. If the Haiku share runs on prompts over 100,000 tokens, its rate is $0.50 and the blend is (0.7 × $0.50) + (0.2 × $2) + (0.1 × $4) = $1.15.
• Prompt caching bills a cache read at 0.1x the base input rate on most models, 0.05x on Opus 5.5 and Sonnet 5.5 and 0.025x on Fable 5.1, against a write at 1.25x for five minutes or 2x for an hour, so caching pays off after one read on the five-minute cache or two reads on the one-hour cache (Anthropic docs).
• The Batch API takes 50% off every input and output token for asynchronous work, with most batches finishing within an hour and all results available within 24 hours (Anthropic docs), and its discount stacks with caching.
• Session hygiene, meaning fresh sessions per task, pasting snippets instead of whole files and asking for diffs instead of rewrites, removes tokens inside one conversation, but it is manual and resets every time a developer forgets it.
• Shared organization memory removes the repeated context that every call would otherwise re-send, and Sentra measured 41.2% fewer tokens and 72.6% lower model cost on Terminal-Bench 2.1, a Sentra internal evaluation pending official verification, run with Codex CLI and GPT-5.5 xhigh rather than Claude (results).

The first three lower the price of the tokens you send. The last two lower how many you send, and only shared memory keeps doing so without daily discipline.

Why re-explaining context is the hidden cost driver

Claude reprocesses the entire conversation history on every new message, not just your latest input. That reprocessing is where the bill compounds. On Sonnet 5.5, a turn that carries 100,000 input tokens costs $0.20 in input alone, twice the $0.10 of a turn at 50,000 tokens, and every later turn carries at least as much, so a session's cost grows faster than its length.

Three habits keep that record bloated. You re-read files Claude already processed, which adds duplicate tokens for content that hasn't changed. You re-paste earlier code or re-explain the project setup at the start of each message, even though the session already holds it. Dead history from failed experiments and abandoned tangents stays in context until you clear it, forcing every new question to be processed against a record you no longer care about.

Session hygiene helps, and clean sessions plus snippet-only pasting are free to adopt. Trimming still fights the same problem every session, because you restart with an empty window and rebuild the same context by hand.

Persistent, write-time memory breaks the compounding instead of managing it turn by turn. Sentra comprehends and stores your org context once, then retrieves only the relevant subset per call rather than replaying full history. The redundant input tokens that drove the doubling never enter the window, so cost grows with new work rather than with accumulated history.

Comparison: memory approaches and their effect on token spend

Session hygiene and vector-search RAG both cut some waste, and both are cheap to start, but neither stops your Claude bill from re-processing stale context. Sentra comes out ahead on token reduction because it comprehends facts at write time and knows when each one stopped being true.

The bi-temporal row explains the token gap. Vector search retrieves passages by similarity, so a deprecated policy from last quarter scores as relevant and gets fed back to the model as current, and the model spends tokens reconciling it. Sentra records when each fact became true and when it was superseded, so it filters out the stale version before the query ever reaches Claude. That is how it removes redundant tokens that trimming a session cannot reach. The Terminal-Bench 2.1 result is Sentra's own internal evaluation, 445 trials at 88.31% mean reward against the published 83.37% baseline, pending official verification, and it was run on Codex CLI with GPT-5.5 xhigh, so it shows the mechanism rather than a Claude-specific saving.

How to choose for your team's Claude bill

Start with the plan that matches your team size, then optimize the API spend underneath it. A five-person team fits Team Standard at $20 per seat billed annually or $25 monthly, with Claude Code and Opus included. Developers who keep hitting Standard limits get Premium seats at $100 or $125, and the rest stay on Standard. Above a few hundred million tokens a month, the API bill dwarfs seat costs, and that is where the real work begins.

Layer the mechanical wins first. Turn on prompt caching for stable context, route asynchronous jobs through the 50 percent Batch API discount, and split traffic across models so Haiku 5.5 handles the bulk; on our arithmetic above, a 70/20/10 Haiku, Sonnet and Opus mix costs 56.5% less than all-Sonnet and 78.25% less than all-Opus on input when the Haiku prompts stay under 100,000 tokens.

Add shared organization memory once repeated context, not new work, dominates your token count. Sentra sits underneath Claude and the agents you already run, feeding each call the relevant facts instead of the full history. It complements your stack rather than replacing Claude, Cursor, or Slack, and it serves the agents already doing the work. Our Claude Code pricing guide covers plan break-evens, and the Claude API pricing guide lists every model's rate.

Disambiguation note

Sentra at sentra.app is the managed organization memory for teams and agents described throughout this article. It is a different company from Sentra at sentra.io, which sells data security posture management. It is unrelated to the Nissan Sentra or any cleaning brand of the same name.

FAQ

Bottom line

Your plan choice sets the floor on your Claude bill. Team Standard at $20 to $25 a seat and Sonnet 5.5 at $2 input and $10 output per million tokens are fixed costs you budget against. Repeated context sets the ceiling. Every call that re-injects the same system prompt, profiles, and history pays full input rates for content that never changed. Caching, batching, and model routing trim that waste session by session. Shared organization memory removes it, and Sentra measured 41.2% fewer tokens and 72.6% lower model cost on Terminal-Bench 2.1, a Sentra internal evaluation pending official verification, run with Codex CLI and GPT-5.5 xhigh rather than Claude.