Manus vs ChatGPT Work vs Claude 2026: Real Cost of an AI Agent
Short Answer
These three "hand off a task" products meter cost in three genuinely incompatible ways, and none of them lets you predict a task's price in advance. Manus is the most transparent about the unit itself — 1 credit = $0.005 flat, confirmed by dividing every Pro tier's price by its credit allocation ($20÷4,000 = $40÷8,000 = $0.005) — but Manus's own documentation explicitly states it will not tell you how many credits a task will cost until the task finishes. ChatGPT Work and Claude's unified agentic mode are cheaper to access on paper ($20/month Plus or Pro) but have no published per-task unit at all: both draw from a variable "agentic usage allowance" sized only in relative terms (multiples of a baseline), not dollars or credits.
Rough cost per task, by type (MODELING derived from Manus's own confirmed credit price and third-party-reported consumption ranges — no equivalent breakdown exists for ChatGPT Work or Claude, which is itself a finding):
| Task type | Manus credits (independent reporting range) | Manus $ equivalent (CALCULATION at $0.005/credit) |
|---|---|---|
| Simple chat/query | 10–50 | $0.05–$0.25 |
| Simple web research (1–3 sites) | 100–300 | $0.50–$1.50 |
| Deep research / multi-step report | 500–900 | $2.50–$4.50 |
| Complex multi-step automation | 1,000–1,500+ | $5.00–$7.50+ |
At 50 tasks/month averaging "deep research" complexity, Manus would need roughly 35,000 credits (CALCULATION) — inside the $200/month Pro tier's 40,000-credit allocation, but with little headroom. ChatGPT Work and Claude have no equivalent number to compare against; this article treats that gap as a real finding, not a research failure to be papered over with a guess.
1. Current Plans and What Consumes Credits/Allowance
Manus — (help.manus.im, manus.im/docs/introduction/plans, cross-checked against numerous convergent third-party trackers dated through September 2026)
| Plan | Price | Credits/month | Notes |
|---|---|---|---|
| Free | $0 | 300/day refresh (does not accumulate) + 1,000 one-time starter | Manus 1.6 Lite only, 1 concurrent task, 2 scheduled tasks |
| Pro (entry) | $20/mo | 4,000/mo + 300/day refresh | Full model access (1.6, 1.6 Max), up to 20 concurrent/20 scheduled tasks |
| Pro (mid) | $40/mo | 8,000/mo + 300/day refresh | Same features, larger pool; 7-day full-access trial offered here |
| Pro (top) | $200/mo | 40,000/mo + 300/day refresh | Same features, largest self-serve pool |
| Team | $20/seat/mo (2-seat minimum) | 4,000/seat/mo, pooled | SSO (free at 30+ seats, else +$150/mo), usage analytics, access controls, shared templates, data-training opt-out |
Credit mechanics (vendor-listed, Manus's own consumption-rules documentation, quoted by multiple independent sources verbatim): a credit covers three things at once — "LLM tokens, Virtual machines, and Third-party APIs" — and "credits are only consumed during active task processing. Completed tasks and the storage or deployment of their outputs do not consume any credits." Most sources describing real usage say failed tasks are not automatically refunded, while one source reports a refund. Manus does not clearly state a refund rule on a primary policy page, so the public evidence conflicts. Monthly plan credits do not roll over; daily refresh credits reset every 24 hours without accumulating; only purchased add-on credit packs persist, and only while the subscription stays active. Annual billing saves 17% across all paid tiers.
Public sources disagree over whether Meta acquired Manus in December 2025; the available information does not establish its current ownership. This does not affect the pricing comparison.
ChatGPT Work — vendor-listed (chatgpt.com/pricing feature table; help.openai.com articles on Codex/Work credit sharing, as established in NotCheapAI's existing Claude-Cowork-vs-ChatGPT-Work coverage)
Launched July 9, 2026. No separate price — access gated by plan (full access from Plus/$20 up; Free/Go get only "limited" desktop-app access). ChatGPT Work, Codex, ChatGPT for Excel/PowerPoint and Workspace Agents share one agentic usage allowance, which is explicitly separate from ordinary chat metering — OpenAI's own documentation states "Chat usage is metered separately from Astra usage in Work and Codex," and regular chat keeps working when the Work/Codex allowance is exhausted. OpenAI has not published a per-task dollar or credit cost for this allowance — it is described only as "variable... depending on complexity," the same framing applied to Codex generally.
Claude (unified agentic mode, formerly Cowork) — vendor-listed (claude.com/pricing, claude.com/blog/cowork-is-now-claude, as established in NotCheapAI's existing coverage)
Merged into plain Claude on September 16, 2026. Access via Pro ($17–20/mo) and Max first; Team and Free rollout still in progress at last check. Chat and Claude Code are documented as sharing one combined usage pool; whether the newly-merged agentic task capability draws from that same pool or a separate one was not stated explicitly in Anthropic's own merge announcement and remains unconfirmed. No fixed task count or dollar figure is published for any tier — usage limits are described only as multiples of each other (Pro ≥5× Free per 5-hour session; Max 5×/20× more than Pro).
2. Why Task Units Are Not Comparable Across Platforms (Critical Rule From the Brief)
This article does not convert Manus credits, ChatGPT Work's allowance, and Claude's pool into one shared unit, because no defensible conversion exists. Three reasons:
- Manus is the only one of the three with a published, fixed per-credit dollar value ($0.005, derived by arithmetic from its own price/credit ratio, flat across every Pro tier). ChatGPT Work and Claude have no equivalent figure — their "allowance" is a relative sizing (multiples of a baseline), not a currency.
- What counts as "one task" differs by design. Manus explicitly bundles LLM tokens, VM time and third-party API calls into a single credit meter; ChatGPT Work and Claude's allowances are not documented at that level of mechanical detail in either company's own materials.
- None of the three tells you the cost before you run the task. Manus shows an estimated range in its dashboard before launch, but confirms the final number only after the task completes; ChatGPT and Claude give no pre-task estimate at all in anything reviewed for this article.
3. Normalized Scenarios: 10 / 50 / 200 Agentic Tasks per Month
MODELING ASSUMPTION: "task" = a moderately complex, multi-step agentic job (roughly Manus's own "deep research / multi-step report" band, 500–900 credits, midpoint 700). This assumption exists only to give Manus a number; it cannot be applied to ChatGPT Work or Claude, which is the point.
| Volume | Manus credits needed (CALCULATION, 700/task) | Manus $ equivalent | Manus plan that fits | ChatGPT Work | Claude |
|---|---|---|---|---|---|
| 10 tasks/mo | 7,000 | $35.00 | Pro $40 (8,000 credits) — Pro $20's 4,000 falls short | $20/mo flat (Plus), subject to an unpublished allowance ceiling | $17–20/mo flat (Pro), subject to an unpublished allowance ceiling |
| 50 tasks/mo | 35,000 | $175.00 | Pro $200 (40,000 credits) — tight fit, little headroom | $20/mo flat, allowance likely stressed at this volume — unconfirmed | $17–20/mo flat, allowance likely stressed at this volume — unconfirmed |
| 200 tasks/mo | 140,000 | $700.00 | Exceeds every published self-serve tier — requires multiple accounts, heavy top-ups, or a sales conversation not researched here | $20/mo flat if the account survives the allowance ceiling; more likely requires Business/Enterprise or multiple seats — unconfirmed | $17–20/mo flat under the same caveat, or Max/Team — unconfirmed |
The honest reading of this table: Manus is the only platform where this table can be filled in with real numbers, because it is the only platform with a published price-per-unit. ChatGPT Work and Claude's cells are marked unconfirmed not because this research was incomplete, but because OpenAI and Anthropic have not published the information needed to fill them in — a finding in itself, and one worth flagging to readers rather than papering over with an invented number.
4. Scheduled Tasks, Concurrency and Cloud Execution
- Manus: up to 20 concurrent and 20 scheduled tasks on every paid tier (vendor-listed); Free is capped at 1 concurrent, 2 scheduled. Parallel execution (e.g., Wide Research spinning up sub-agents) multiplies credit consumption per multiple independent reports — running 5 parallel research threads is described as costing roughly 5× a single thread, not a discounted bundle.
- ChatGPT Work: the existing Cowork-vs-Work comparison does not confirm cloud-execution persistence for ChatGPT Work, so this remains an open point.
- Claude: cloud execution mode (inherited from Cowork) is confirmed to continue running after the client is closed, resumable from any surface — the clearest "keeps working while you're away" guarantee among the three, though how many parallel agentic tasks a single account can run is not specified in the cited information.
5. File/App Integrations
- Manus: connectors to Gmail, Google Calendar, Notion, Slack, WhatsApp, Telegram (vendor-listed-adjacent, widely and consistently reported); a Web App Builder and Design View (interactive image canvas) added in the 1.6 release; Mobile Development (iOS/Android app generation from a prompt).
- ChatGPT Work: 1,400+ connectable apps via OpenAI's plugin/connector directory (figure carried over from existing NotCheapAI research, not re-verified here).
- Claude: Connectors and Skills (carried over from the Cowork era), plus the newly-launched Claude Docs and Claude Slides for in-conversation document/deck creation.
None of these integration counts were normalized against each other — a connector count is not a cost figure, and this article does not attempt to price "value of integrations" as a dollar amount.
6. Overages, Top-Ups and Commercial/Team Tiers
- Manus: purchasable add-on credit packs exist (reported price roughly the same ~$0.005/credit as plan pricing, independent reporting, not confirmed as an official static rate); add-ons persist only while a paid subscription stays active, and go inactive on downgrade to Free. A Basic Cloud Computer add-on for persistent disk storage was reported at $10/month for 35GB (independent reporting, single source, not primary-confirmed). Team requires a 2-seat minimum at $20/seat/month with SSO free only at 30+ seats.
- ChatGPT Work: no overage or top-up price beyond the shared Work/Codex allowance is published in the coverage cited here; check current terms before budgeting.
- Claude: same — no published overage price for the agentic pool found in existing or new research.
7. Hidden Costs
- Manus's credit unpredictability is itself a cost. Multiple independent sources describe users exhausting a monthly allocation faster than expected because the same nominal task type ("deep research") can land anywhere in a 500–900+ credit band depending on how many sources the agent decides to check, how many revision loops occur, and whether execution hits errors — and a failed or looping task still spends the credits already consumed, with refund policy unconfirmed.
- Parallel/Wide Research execution scales cost linearly with sub-agent count, which is easy to trigger accidentally on a "research this broadly" style prompt.
- ChatGPT Work and Claude's opacity is the inverse problem: with no published unit price at all, a user has no way to model monthly cost in advance beyond "pay the subscription and see if I hit a wall." That is not a criticism of either product's value — it is a structural difference from Manus that this article is required to surface, not paper over.
- None of the three publishes what happens when a heavy user's usage grows month over month beyond generic "upgrade to a higher tier" guidance — no platform in this research disclosed a hard cutoff or a pay-as-you-go extension mechanism with a stated price.
8. BUILD & EARN Scenario: A Research-as-a-Service Business
MODELING ASSUMPTION: a solo operator sells finished research reports at $50 each, delivering 40 reports/month = $2,000 revenue, using Manus at the "deep research" band (midpoint 700 credits/report). Metric: platform-cost gross margin (excludes the operator's own review/editing time — not profit).
CALCULATION: 40 × 700 = 28,000 credits/month = $140.00 at $0.005/credit → fits inside the Pro $200 tier (40,000 credits) with headroom, while the Pro $40 tier would only fit much lower-volume months near its 8,000-credit ceiling. Platform-cost share of $2,000 revenue: 7.0–10.0% depending on which tier is actually purchased; model-cost gross margin: 90–93%.
The same business built on ChatGPT Work or Claude cannot be modeled this way, because neither publishes a per-task cost — the operator would need to run the actual workload for a billing cycle and observe where the account's allowance ceiling sits before any margin figure could be trusted. That gap is the clearest practical takeaway of this whole comparison for anyone trying to build a business on top of these tools today.
FAQ
How much does one Manus task cost? There's no fixed price — Manus's own credit price is a flat $0.005, but a task's credit consumption (roughly 10 to 1,500+ credits) depends on complexity, and Manus won't tell you the final number until the task finishes.
Is ChatGPT Work or Claude cheaper than Manus? Their subscriptions are cheaper on paper ($20/month vs. Manus's $20–$200 range), but neither publishes a per-task cost, so a true apples-to-apples comparison isn't possible with public information.
Does ordinary ChatGPT chat use up my Work/Codex allowance? No — OpenAI meters them separately, and regular chat keeps working after the Work/Codex allowance is exhausted.
Can I run 200 agentic tasks a month on any of these for a predictable price? Only Manus gives you a number to check against (and even then, 200 mid-complexity tasks would likely exceed the highest published self-serve tier). ChatGPT Work and Claude have no published ceiling to check against at all.