Paid vs. Free
What ChatGPT, Claude and Gemini cost as of October 2026, how usage limits work, when the free tier is enough, and the two billing surprises (the Agent SDK change and the AWS Bedrock trap) that hit agent builders.
On this page
- What the plans cost, October 21, 2026
- ChatGPT (chatgpt.com/pricing)
- Claude (claude.com/pricing)
- Gemini (Google AI plans)
- How usage limits actually work
- When free is enough
- Running an agent on a subscription: the 2026 billing saga
- The AWS Bedrock billing trap
- What burns money
- The low-cost stack
- Related
In September 2025 the group admin wrote: "Every shliach must have at least one paid AI plan. $20 is באין ערוך." A poll the same day came back 24 paid, 13 free only. A year later the question has changed. Free tiers got much better, the top plans now run from $100 to $500 a month, and the real cost question for many members is usage limits, not the sticker price.
Nonprofit and teacher pricing is on its own page: Nonprofit Discounts. A registered Chabad House can get Claude or ChatGPT team seats for $8 a user per month.
What the plans cost, October 21, 2026
US prices, checked on each company's pricing page. Prices outside the US differ.
ChatGPT (chatgpt.com/pricing)
| Plan | Price | What changes |
|---|---|---|
| Free | $0 | Text chat with GPT-6 (OpenAI calls it unlimited, within abuse limits). Limited uploads, image creation, voice, deep research, Codex, and ChatGPT Work on desktop. |
| Go | $8/month | More messages with tools, more uploads and images, longer memory. May show ads. |
| Plus | $20/month | GPT-6 Astra and the Sol models, ChatGPT Work on desktop, web and phone, more Codex, projects and scheduled tasks. |
| Pro | from $100/month | Three usage tiers, Astra's Pro reasoning, longer Work and Codex sessions, Dots (OpenAI's always-on agent), unlimited image creation within abuse limits. |
Claude (claude.com/pricing)
| Plan | Price | What changes |
|---|---|---|
| Free | $0 | Chat, file creation, Claude Design. No Claude Code, no Claude in Chrome, no Opus. |
| Pro | $20/month ($17 billed yearly) | Opus, Claude Code, Claude in Chrome, Projects. Fable only on usage credits. |
| Max 5x | $100/month | 5 times Pro usage, Fable up to half of weekly limits, $100/month in API credits. |
| Max 20x | $200/month | 20 times Pro usage, Fable up to half of weekly limits, $200/month in API credits. |
| Team | $25/seat monthly, $20 yearly (Premium seat $125 or $100) | 2 to 150 seats; nonprofits pay $8 (Standard) or $40 (Premium). |
Gemini (Google AI plans)
| Plan | Price | What changes |
|---|---|---|
| Free | $0 | The Gemini app with a lighter Flash model, image generation, Deep Research with limits. |
| Google AI Plus | lower than Pro; price varies by country | Twice the free usage, a newer Flash model, 400 GB of storage. |
| Google AI Pro | $19.99/month | Gemini 4 Argon and Gemini 3.1 Pro, four times the free usage, Gemini in Gmail and Docs, 5 TB of storage. |
| Google AI Ultra | from $99.99/month (5x Pro usage); $199.99 for 20x | Highest limits, Gemini Spark (Google's always-on agent), 20 TB or more of storage. |
A registered Chabad House gets the Gemini app free through Google Workspace for Nonprofits; see Nonprofit Discounts.
How usage limits actually work
A paid plan is "not a bottomless well," as one member put it in September 2026. You get an allowance. When you use it up, the assistant pauses until the allowance resets, unless you buy extra usage.
- Claude has a rolling five-hour limit and, on paid plans, a weekly limit. Bigger models use the allowance faster. Fable 5.1 uses it fastest; in September 2026 one member "maxed out two accounts" on it.
- Agent modes burn faster. Claude Cowork and ChatGPT Work run many steps, sometimes many helpers at once. In August 2026 the admin traced members' limit problems to Cowork starting lots of agents. Anthropic's own Cowork page says it uses limits faster than chat.
- One account, one allowance. If your personal agent runs on your Claude account, everything it does comes out of the same allowance you chat with. On a Claude Team plan, each seat has its own allowance, as long as the admin has not turned on extra usage for the organization (the admin, August 2026).
- Members' consensus (August to September 2026): when you keep hitting limits, move to a higher plan rather than buying extra credits; use plain chat before agent mode; match the model and effort level to the task.
Limits also move. On May 6, 2026 Anthropic doubled Claude Code's five-hour limits on paid plans after a compute deal with SpaceX, and the Opus 5.5 launch on Sep 22, 2026 raised five-hour limits again.
When free is enough
The counter-position has held since 2025: "Don't feel pressured." One member dropped his paid plan for low use and works fine on free tools. If you are not making flyers, letters and spreadsheets every week, start free and upgrade the month you hit the wall. "Once you understand the limits, you'll upgrade naturally."
What the free tiers handle well in October 2026:
- ChatGPT Free: everyday questions and drafting on GPT-6.
- Claude Free: layout work. A UK member found free Claude did a page rewrite better than his paid ChatGPT (October 2026), and Claude Design is included on the free plan.
- Gemini Free: quick questions and image generation, within daily limits.
Where free runs out: uploads of long PDFs, batches of flyers, Deep Research, and any agent work.
The September 2026 advice still applies to the plan you pay for: "Stick to one and get really good at it. When its deficiencies really bother you, switch."
Running an agent on a subscription: the 2026 billing saga
Members who run a personal agent (see GabAI, Hermes Agent and Agent Hosting) want it to run on a flat monthly subscription instead of paying per token. Paying per token surprised people: one member spent about $40 in under a week in March 2026, and another burned about $50 of credit in 20 minutes on Opus 5 in September 2026.
What happened with Claude, in order:
- April 2026. Members shared an Anthropic email saying that "third-party harnesses" connected to a Claude account would draw from extra usage instead of the subscription. Builders using Anthropic's official Agent SDK understood they were not affected.
- May 2026. Anthropic announced that from June 15, Agent SDK use,
claude -p, and third-party apps built on the SDK would stop counting against subscription limits and run on a separate monthly credit instead. Members read the email as $100 a month for Max 5x and $20 or $100 for Team seats. - June 15, 2026. Anthropic paused the change. Its help page now says: "You can still use the Claude Agent SDK, claude -p, and third-party apps with your subscription limits" (support article).
- October 7, 2026. Max and Team plans began including monthly API credits: $100 on Max 5x, $200 on Max 20x, and up to $500 pooled on Team, according to claude.com/pricing.
The group lead's advice through the summer was to keep a fallback ready in case the subscription route closes. Several builders moved some agents to ChatGPT through a Codex sign-in, which uses a ChatGPT subscription instead of an API key (June to July 2026). Members said the agent felt "a bit dumber" but worked.
The AWS Bedrock billing trap
In July 2026 a shliach's real AWS invoice showed two bills. The server itself was covered by nonprofit AWS credits. A separate AWS Marketplace bill for Claude, called through Amazon Bedrock, ran several hundred dollars a month at full price. His agent had been set up to use Bedrock without his noticing, and AWS support told him the credits did not cover Marketplace charges. AWS's nonprofit credit terms do list AWS Marketplace as excluded.
If you run an agent on AWS:
- Open AWS Cost Explorer, group costs by service, and look for a Claude or Marketplace line.
- Ask whoever set up your agent (or ask Claude Code on the server) whether the agent's settings point at Bedrock.
- If you already pay for a Claude plan, point the agent at that subscription instead, then restart it and confirm it still answers.
- Track when your credits expire; a member warned that the renewal date is easy to miss.
Lesson the group drew: free credits do not follow the model. Check which products a credit actually covers.
What burns money
- The biggest model by default. Fable 5.1 and Opus at high effort. Builders set Claude Code's helper agents to never use Fable.
- Agent modes left running. Cowork, ChatGPT Work, Codex and self-hosted agents.
- API keys instead of a subscription, and cloud model routes like Bedrock that your credits do not cover.
- Deep Research and very long documents.
- Paid app builders. Members criticized Base44 for token cost in August and September 2026. See Vibe Coding.
The low-cost stack
- A nonprofit Claude or ChatGPT seat at $8 a month, or a teacher program if you qualify.
- Google Workspace for Nonprofits for the free Gemini app, NotebookLM and storage.
- Google AI Studio for trying Gemini models in the browser.
- The free tiers of ChatGPT and Claude for second opinions.
- Free dictation on your own computer: see the offline dictation guide.