Pricing and Plan Selection for Design Teams

This is the most perishable page on this site. The dollar amounts below were re-read on 9 October 2026 from the vendors' own pricing and help pages, not from the June snapshot this article used to carry. Regional prices, tax, promotions, and which tiers are open for new sign-ups move without a blog post. Treat the table as a dated reading plus a way of choosing, and open the linked page before anyone is billed.

Two scope notes. This is not procurement, financial, or legal advice. And it covers money and plan fit only. Sandboxing, instruction files, skills, MCP, and churn live in the platform-comparison article.

The re-check method, when this date is old: open the vendor page, compare the thing you actually keep (the repository, the hosted app, and the marketing site are different products), and score the plan on context, computer use, MCP, and seats. Do not carry forward a dollar amount this page has stopped printing.

The useful shape is still the metering model, not a single sticker price. Subscription windows (Claude) give a rolling allowance — the plan page still describes five-hour session windows shared across the apps and Claude Code — and a heavy screenshot day fails by making you wait. Credit metering (Codex on ChatGPT plans) draws usage against the plan's included allowance and then against purchased credits, so the loop does not stall and the meter can. Bring-your-own-key (OpenCode, or a direct API route) prices the model rather than the seat, and fails when nobody owns the bill.

The professional rungs still cluster near $20, $100, and $200 a month on Claude, ChatGPT Pro, and Google AI Ultra, as re-read on 9 October 2026. That is no longer the whole ladder. Google AI Plus is $4.99. ChatGPT Pro also lists a $500 tier. ChatGPT Go exists as a lower-cost individual plan, and its help article tells you to read the pricing page for the amount — that page did not render a number we could quote, so this note does not print one.

Vendors still describe many limits relative to another tier — 5x Pro, 20x Pro — rather than as a fixed message count. A hard number of messages per tier, quoted without a vendor page, is not a fact.

• Subscription windows (Claude): predictable monthly cost, shared rolling allowance; heavy days hit the window and stall the loop.
• Credit metering (ChatGPT plans for Codex): included usage first, then credits where the plan allows them; the loop keeps going and the meter moves.
• BYOK / API (OpenCode, direct API): pay for the model, and assign someone to own keys and cost tracking.
• Relative limits (5x, 20x) are what the vendors publish. This note does not invent absolute token or message counts.

Every amount below is a US list price from the linked page on that date. Tax and local currency differ. Where a page named a tier but not a dollar amount, the tier is named and the amount is omitted.

Reading it as a designer: image input and computer use burn allowance faster than chat. Team and business tiers are mostly seats, billing, and admin. The BYOK route is the one where spend tracks tokens rather than seats.

Plan pages are written for general subscribers, so the design-relevant dimensions are easy to miss. Four of them decide more than the sticker. Context: whether a design-system packet fits in one session, which you check on the current model card, not from memory. Computer use: browser and screenshot loops, which are what burn windows and credits. MCP: whether the plan's agent can reach Figma, a canvas, and a browser server. Seats: central billing, SSO, and the ability to mix light and heavy seats once more than a couple of people are involved.

Habits still stretch every plan: crop screenshots to the component, keep context on the task, use a cheaper model for drafts when the platform allows it, and compact long sessions. Measure a normal week before you climb a rung.

Compare the product you keep — repository agent, hosted app, or marketing site — then score seat, context, computer use, and MCP. Dollar amounts stay in the dated table above. This figure has none.

Match usage intensity to a rung, then confirm the amount on the vendor page the day you decide. The heuristic is meant to survive the next price change. The table above will not.

If the organisation already pays for Claude or ChatGPT, evaluate the agent included in that plan before adding a vendor. Treat the first month as the measurement period.

• Occasional sessions: Claude Pro or ChatGPT Plus. Upgrade when you actually hit the allowance, not when a comparison article says you will.
• Daily design-to-code work, with screenshot loops: a $100-class rung in the ecosystem you already use — Claude Max 5x, ChatGPT Pro 100, or Google AI Ultra 5x — after you have seen a week of your own meter.
• All-day agent work, parallel agents, large design-system context: a $200-class rung, ChatGPT Pro 500 if that tier is what the work actually needs, or API/BYOK when a flat seat fights a spiky month.
• A mixed team: mix standard and premium seats, or entry tiers for occasional people and a higher rung for the people whose meter says so.
• Compliance or a controlled budget: enterprise or the API route, where admin controls and spend limits live. That is a procurement conversation.
• Spend must track usage: OpenCode or another BYOK route, with a named owner for keys and reporting on day one.

Output quality, sandboxes, and what ports between platforms are the platform article's job. This note stays out of negotiated discounts and anything that resembles procurement advice.

Prices omitted on purpose, because they were not on a page we could quote on 9 October 2026: ChatGPT Go's monthly amount, and any claim that Google Stitch replaced Gemini CLI. Gemini CLI's individual cutoff is documented; Stitch is not that announcement. Windsurf's rebrand to Devin Desktop is also outside this price table — no article on this site was still selling Cascade as the current local agent, and this note does not invent Devin prices.

If the date at the top is more than a quarter old, read the vendor pages first. The metering models and the four plan dimensions are the part that should still be useful.