| Service Type | Setup Fee Range | Monthly Retainer | Avg Margin | Best For |
|---|---|---|---|---|
| 💬 Chatbot / Assistant | $1,500–$5,000 | $500–$1,500/mo | 65–75% | SMBs, e-commerce, service cos |
| 📧 Email Automation | $2,000–$6,000 | $750–$2,000/mo | 60–72% | Coaches, SaaS, agencies |
| 🎯 Lead Generation Bot | $3,000–$8,000 | $1,000–$3,000/mo | 55–70% | Real estate, insurance, finance |
| ✍️ Content Automation | $2,500–$7,500 | $800–$2,500/mo | 65–80% | Content creators, media, blogs |
| 🏢 Full Office Automation | $8,000–$35,000 | $2,500–$7,500/mo | 45–65% | Mid-market, growing teams |
| ⚙️ Custom AI Agent | $5,000–$25,000 | $1,500–$5,000/mo | 50–70% | Tech cos, SaaS, operations |
| 📱 Social Media Automation | $1,500–$4,500 | $600–$1,800/mo | 70–82% | Brands, coaches, ecommerce |
* Ranges reflect 2026 US market rates. Final pricing depends on complexity, client size, and your experience level. Model strategy affects margins more than list prices: open-weight stacks (Qwen 3.8 Max at $2/$6 per 1M tokens, Kimi K3, GLM-5.2) cut the compute line vs. paid frontier APIs — but monitor pending final terms: Qwen 3.8 Max revenue share with Alibaba required for qualifying model-as-a-service use (rates TBD, effective date TBD), mirroring Moonshot's Kimi K3 license, which triggers a commercial agreement above $20M in annual sales (up to 30% revenue share). ⚠ DeepSeek V4 (V4-Flash $0.14/$0.28, V4-Pro $0.435/$0.87 per 1M tokens) currently lists the lowest rates on this page, but those rates are PROVISIONAL — DeepSeek announced (Aug 6, 2026) a significant API price increase with no new schedule published yet; verify api-docs.deepseek.com/quick_start/pricing before quoting. GPT-5.6 Sol (Aug 6, 2026) now powers both Instant and deep reasoning for ChatGPT Plus/Pro — one consistent model with a reasoning-effort slider; GPT-5.6 Luna is the new default for Free/Go users (unlimited text chats rolling out this week/next week). OpenAI has not published per-token API pricing for the new models, so the GPT-5.6 Sol option in this calculator is a marked estimate — verify current rates at OpenAI pricing before quoting. Agent Plugins 1.0.0 (Aug 6, 2026) adds a portability lever: one plugin package serves Codex, ChatGPT, Cursor, GitHub Copilot, Kiro, and VS Code, so setup fees carry a build-once discount and plugin packaging & distribution is priced as its own deliverable (see the Portability selector and Agent Plugins section ↓). New (Aug 2026): agent payment rails like Cloudflare Wallets (announced Aug 4, 2026) add a creator-set spending cap at the wallet layer — the same rails that make the capped vs uncapped spend comparison in this calculator worth running before you quote a retainer.
Agent workflows rarely run clean the first time. On Aug 5, 2026, levelsio (Pieter Levels) reported burning $500 per Gauntlet Loop run — an AI-coding method that fans out subagents and loops until "utterly perfect" — then corrected it to $900 total with 95% of generated code removed. Measured baselines are ~$0.06 per request and "a few dollars per task"; failure modes (retry storms, subagent fan-out, silent misconfiguration) turn that into $500 loops and $2,000 overnight bills. Use this estimator to model what retries actually add to your spend.
On Aug 4, 2026 Cloudflare announced Cloudflare Wallets and cloudflare.pay — the buyer side of agentic commerce. An Account Wallet (human-owned) funds Virtual Wallets (agent-owned) with guardrails: a spending cap / allowance, an approved merchant allow-list, and a maximum transaction size. Agents that hit a limit must request a manual override from a human — they cannot approve escalations themselves, and the cap is enforced at the wallet's API layer, not inside the model's system prompt, so prompt injection can't lift it. Cloudflare's own example: give every employee a $100/week budget for AI inference by provisioning one Account Wallet and Virtual Wallets per employee with that rule. Fees are undisclosed so far — this estimator defaults to 0% and lets you set your own. For agencies, the takeaway is cap-constrained planning: with wallet-layer limits, worst-case spend per agent becomes a known number (the cap), so cost forecasts get a hard ceiling instead of a tail.
Google Cloud's new managed model routing (API Gateway, Public Preview since Aug 3, 2026) accepts your existing OpenAI-compatible chat requests, inspects the model name in each payload, and routes the call to a cheaper foundation model — with no client-side code changes. This estimator shows the potential token-cost savings from routing simple traffic to Gemini Flash-Lite instead of paying Flash/Pro rates for everything.
On August 3, 2026 Google Cloud added managed model routing to API Gateway (Public Preview). It accepts OpenAI-compatible chat requests, transcodes them in-flight, and dispatches them to Gemini, Anthropic Claude, or OpenAI models hosted in Vertex AI Model Garden. Google positions it as a managed replacement for self-hosted proxies like LiteLLM — no proxy server to host, scale, or maintain.
Routing is driven by the model name in each request payload. You define a router with a default model plus rules mapping client model strings to cheaper backends — unmatched traffic falls back to the default. Example: send all traffic to Flash, set the default to Flash-Lite, and route only complex/agentic requests to Flash. Google's own examples use google/gemini-3.5-flash-lite, google/gemini-2.5-pro, anthropic/claude-opus-4-7, and openai/gpt-oss-120b-maas.
Using Google's published list prices: a content agency sending 50M input + 10M output tokens/mo to Flash at $165/mo could route 80% to Flash-Lite and drop to ~$65/mo — ≈ $100/mo (~61%) saved. A multi-tier client setup on 2.5 Pro at $212.50/mo with 70% budget-tier traffic could drop to ~$100.85/mo — ≈ $111.65/mo (~53%) saved. A 5% fallback-traffic leak onto Flash-Lite instead of Flash saves ~$16/mo on that slice alone. Token volumes and split percentages are assumptions; substitute your own usage.
- Public Preview: text-only, name-based routing to MaaS models in Model Garden; request-side streaming, gRPC, WebSockets, Gemini Live, VPC-SC, and Private Service Connect unsupported.
- One-way mode: you cannot retrofit routing onto an existing gateway or remove it — switching requires a new API config + gateway.
- Single-host constraint: all models in one router must share the same hostname (global or one regional endpoint).
- Pricing gap: no model-routing-specific fee was found in the reviewed sources; confirm your exact model versions and region before quoting a client.
- No per-request observability yet: routing decisions aren't attributed per request in logs during preview.
- API Gateway — Overview of model routing (Google Cloud docs)
- API Gateway — Configure model routing (Google Cloud docs)
- Google Developers Blog — A unified API for AI model routing
- Vertex AI — Generative AI pricing (Gemini token rates)
- API Gateway pricing (per-call tiers)
- Google Cloud release notes (Aug 3, 2026)
- TLDR AI — Aug 5, 2026 issue
- API Gateway quotas and limits
2026-08-08: Added DeepSeek V4 (V4-Flash / V4-Pro) as a selectable model strategy with a PROVISIONAL pricing flag — DeepSeek announced (Aug 6, 2026, footnote 2 of its official pricing page) that it plans to raise overall API pricing "in the near future, with a significant increase expected," and has NOT disclosed new rates, a percentage, or an effective date as of 2026-08-08. Current list prices (V4-Flash $0.14/$0.28, V4-Pro $0.435/$0.87 per 1M tokens) are treated as pre-hike/provisional: the strategy is deliberately given a smaller discount than the pure open-weight stack so the calculator does not over-state "DeepSeek is cheapest," and the output carries an explicit verify-before-quoting note. New prominent red warning banner in the Open-Weight section, FAQ item + FAQPage schema entry, meta tags extended, sources linked (DeepSeek pricing page live 2026-08-08, TNW, SCMP, Dataconomy/Bloomberg, TechNode). Open-weight self-host escape hatch noted (MIT weights).
2026-08-08: Added a monitor pending final terms alert for Qwen 3.8 Max: Alibaba (Reuters, Aug 7) plans a revenue-share requirement for qualifying model-as-a-service use (rates TBD, effective date TBD) — the cost model no longer presents Qwen 3.8 Max as free/unrestricted at scale. Kimi K3 precedent verified at >$20M annual sales for a commercial agreement (up to 30% revenue share). New FAQ item + FAQPage schema entry; meta tags extended. Source: Reuters Aug 7, 2026.
2026-08-06: Added Qwen 3.8 Max to the open-weight model strategy (GA Aug 2–3, 2026; 2.4T-param MoE, ~95B active; $2/$6 per 1M tokens; 1M context; open weights promised ~Aug 10). Real-world usage section added with the Aug 6, 2026 45-project field report — self-reported, no artifacts; the "destroyed Fable 5" claim was walked back by the tester (Qwen strong on fast/multimodal builds; Fable 5 on huge long-running projects). Transparent cost-variability note added: API list prices are public, but self-hosted cost is hardware/quantization-dependent and not yet knowable until weights drop.
2026-08-06: Added GPT-5.6 Sol as a selectable model strategy (frontier, Instant + deep reasoning) with a clearly marked estimate and link to OpenAI's pricing page — OpenAI announced Sol now powers both Instant and deep reasoning for Plus/Pro, and GPT-5.6 Luna becomes the default for Free/Go users (unlimited text chats rolling out this week/next week). No official per-token API pricing published for either model; assumptions date/source now included in calculator output.
2026-08-05: Added Gemini API cost & model routing savings estimator and explainer (Google Cloud managed model routing, Public Preview Aug 3, 2026). Pricing sourced from Google's published Vertex AI / API Gateway list prices; scenario figures are illustrative (directional).
⚠ Monitor pending final terms
Qwen 3.8 Max revenue share with Alibaba required for qualifying model-as-a-service use; rates TBD, effective date TBD. Reuters (Aug 7, 2026) reports Alibaba plans to ask major commercial users who offer Qwen 3.8 Max as a service for a share of the revenue they generate — a first for the Qwen Max line — with plans to implement a similar measure as soon as next week; the rate is not finalized. This follows Moonshot's Kimi K3 license, which requires anyone offering the model for sale as a service with more than $20M in annual sales to work out a commercial agreement with Moonshot (reportedly up to a 30% revenue share). Until Alibaba publishes final terms, do not model Qwen 3.8 Max as free/unrestricted at scale — treat it as open-weight with a pending commercial-licensing cost layer that could change the compute line for agencies reselling model-as-a-service.
Source: Reuters — Alibaba plans to charge big users of its next open-source AI model (Aug 7, 2026)
⚠ Price increase announced — DeepSeek rates are PROVISIONAL
DeepSeek has announced it plans to raise overall API pricing "in the near future, with a significant increase expected" — but has NOT published the new rates, any percentage, or an effective date. Current list prices (V4-Flash $0.14 input / $0.28 output, V4-Pro $0.435 input / $0.87 output per 1M tokens) are the pre-hike schedule and are provisional. The warning sits as footnote (2) on DeepSeek's official pricing page, live as of 2026-08-08, and was reported Aug 6, 2026 by The Next Web, SCMP, Dataconomy (via Bloomberg), and TechNode. Until official numbers land, do not treat "DeepSeek is the cheapest option" as a stable assumption — the calculator's DeepSeek V4 model strategy carries a provisional flag, and re-verify api-docs.deepseek.com/quick_start/pricing before quoting it to a client. Self-hosting remains an escape hatch: V4 weights are MIT-licensed open weights (1.6T Pro / 284B Flash), so agencies with sustained workloads can serve them directly via vLLM-class tooling instead of paying whatever the new API rate turns out to be.
Sources: DeepSeek API pricing (footnote 2, live 2026-08-08) · TNW — DeepSeek warns of a 'significant' price rise (Aug 6, 2026) · SCMP (Aug 6, 2026) · Dataconomy/Bloomberg (Aug 6, 2026)
Open-weight models are now a real alternative to paid frontier APIs. Alibaba's Qwen 3.8 Max (GA Aug 2–3, 2026; 2.4T-parameter MoE, ~95B active, 1M-token context) prices at $2 per 1M input tokens and $6 per 1M output tokens — the cheapest open frontier-class API on this page — with open weights promised around Aug 10. Moonshot's Kimi K3 — a 2.8T-parameter open-weight mixture-of-experts model (~104B active, 1M-token context, weights live on Hugging Face since July 27, 2026) — prices at $3 per 1M input tokens and $15 per 1M output tokens, a fraction of flagship paid APIs, while scoring within a few points of Claude Fable 5 and GPT-5.6 Sol on vendor-run coding benchmarks. Zhipu's GLM-5.2 (open weights, MIT license, 1M-token context) is the strongest open-source coding model on Terminal-Bench 2.1. A rumored GLM 5.3 has not been officially confirmed as of August 2026 — build on GLM-5.2 / Kimi K3 / Qwen 3.8 Max today, not on an unannounced model.
Real-world agency usage so far: Qwen 3.8 Max's headline marketing claim — "autonomous coding over 10+ days" — is an official claim, not yet independently replicated, and its Fable 5-beating ranking has been disputed by independent benchmark testing. The most-cited hands-on test so far (Aug 6, 2026, an agency-community builder with ~172K followers) reports building 45 real projects while ignoring benchmarks — 3D racing games, RPGs, websites, a full OS, a promo video, and autonomous workflows — with mixed results: some demos were poor, several builds looked better than Fable 5, and every project took only a few hours. The same tester's follow-up explicitly walked back the "destroyed Fable 5" framing: Qwen 3.8 Max shone on fast, multimodal, image-guided builds (including a single-file premium landing page), while Claude Fable 5 stayed stronger on huge, long-running projects that need consistency across massive contexts. Treat these as first-person, self-reported results with no linked artifacts or independent replication — useful as a delivery-speed datapoint, not as a benchmark.
Cost variability for open-weight models: published API list prices (Qwen 3.8 Max $2/$6, Kimi K3 $3/$15 per 1M tokens) are real, but total cost depends heavily on how you run the model. Qwen 3.8 Max's weights are not downloadable until ~Aug 10, so self-hosting cost is not yet knowable and will vary with hardware (DGX Spark-class vs cloud GPUs), quantization (MXPF4 vs full precision), context length, and utilization. Agencies that self-host trade a variable per-token bill for fixed hardware cost — the crossover point depends on your monthly token volume. Until you benchmark your own workloads, treat self-hosted cost as a range, not a fixed number.
What this means for agencies: model strategy is now a pricing lever. The calculator's Model Strategy selector reflects it — open-weight stacks trim the compute line (and lift margins ~5 pts), frontier-only stacks carry a premium. Keep workflows model-portable across at least two providers, benchmark on your own workloads (vendor tables are not your client's workload), and treat AI spend as a managed line item, not a fixed cost. Read the full analysis of what open-weight models mean for agency margins →
Sources: Alibaba — Qwen 3.8 Max blog · QwenCloud — Qwen 3.8 Max pricing · 45-project field report (X, Aug 6 2026) · Moonshot — Kimi K3 blog · Kimi K3 API pricing · HF model card — moonshotai/Kimi-K3 · zai-org/GLM-5 · Reuters — Alibaba Qwen revenue-share plan (Aug 7, 2026)
AI coding agent pricing is not the per-seat number on the pricing page. That's the core finding of TrueFoundry's August 6, 2026 guide, and it's why this calculator treats model strategy, delivery risk, and retries as explicit cost levers rather than hidden line items. Coding agents bill through three structures, and the same team can see very different invoices under each:
| Billing model | Examples | What agencies should know |
|---|---|---|
| Flat per-seat | Cursor Pro, Windsurf Pro, Claude Pro | Fixed monthly fee per developer — predictable, but overage charges can appear when limits are exceeded, and limits often aren't published clearly. |
| Seat + credits | GitHub Copilot Pro, Pro+, Max | Lower base fee plus a monthly credit pool. A frontier model can burn credits up to 8× faster than a standard model on the same task. |
| Pay-per-token API | Claude Code (API mode), OpenAI Codex | No per-seat charge; billing follows token consumption. Under sustained high-volume agent workloads it can become the most expensive option, and teams routinely misbudget it without usage visibility. |
Six cost variables predict spend better than the headline price: usage volume (light autocomplete vs. all-day agent workflows can differ 20–50× in token consumption), model selection, billing model, team size, billing cycle (annual discounts often 15–20%), and plan tier fit. Free and entry tiers often carry quotas production teams exceed within weeks.
ROI note for agency pricing: if you resell or deliver AI-assisted work, model the raw-token line and the retry/failure line separately — "a task that costs one credit unit on a standard model might cost eight on a frontier model." This calculator's Model Strategy and Delivery Risk selectors exist exactly for that reason. Don't quote the sticker price; quote the consumption profile.
Source: TrueFoundry — AI Coding Agent Pricing: How to Choose the Right Plan (Aug 6, 2026). Full breakdown with the six-step budgeting checklist: AI Coding Agent Pricing in 2026 (findaiagency.com) →
On August 6, 2026, an open, vendor-neutral specification called Agent Plugins 1.0.0 was published for packaging Agent Skills and MCP servers into portable plugins — announced by OpenAI's developer account alongside AWS, Cursor, GitHub, and Vercel. A plugin is a directory: a minimal plugin.json manifest, a skills/ folder, an optional mcp.json, and reverse-domain client-extension namespaces. Six clients support the format at launch: Codex, ChatGPT, Cursor, GitHub Copilot, Kiro (AWS), and VS Code — with AWS's Agent Toolkit compatible (30+ curated skills) and Google's Agents CLI / Data Agent Kit already shipping it. The Technical Steering Committee comprises Amazon, Cursor, Microsoft, OpenAI, and Vercel; Google is joining as a Core Maintainer. The spec is v1.0.0 but labeled a Working Draft, and it deliberately leaves out installation, distribution, permissions, sandboxing, and trust/provenance — all flagged for future versions.
What changes for your pricing: the old cost model assumed a fixed per-platform adaptation cost — the same skill repackaged, re-configured, and re-maintained for every client platform. With one portable plugin, that per-client reimplementation cost collapses. The calculator's Portability / Agent Plugins selector models the shift three ways:
- Build-once amortization (setup discount). One package serves 2–3 compatible clients at ~15% off setup, or 4+ clients / resellable at ~30% off — instead of charging full adaptation per platform. The brief's suggested range is a ~60–90% discount off the adaptation slice per additional client; these factors apply the discount to the whole setup midpoint conservatively.
- Shared maintenance stream (retainer factor). One package + optional per-client extensions replaces N drifting forks, so retainers scale slightly down (0.97× for 2–3 clients, 0.94× for 4+).
- New revenue line (plugin packaging & distribution). Packaging the deliverable as a branded plugin bundle — and distributing it across the compatible-client market — is a priced deliverable, shown in its own result card ($750 for a 2–3 client bundle, $1,500 for a resellable plugin). Resellable plugin lines also lift margins ~3 pts.
One caution before you ship plugins: Agent Plugins v1.0.0 has no permission model, no sandboxing, no provenance verification, and no secrets handling. Agencies selling plugins should self-impose signing, code review, and least-privilege practices — and agencies buying plugins must vet them. That trust layer is itself a sellable compliance service, and the calculator's quote notes it whenever a plugin option is selected.
Client portability cuts both ways: lower switching costs are great for client trust, but they weaken "sticky" platform-based retainers — lock-in risk moves from the package format to the marketplace/install layer. Price the build once, but keep the retainer tied to ongoing value delivered, not to platform lock-in.
Sources: Google Developers Blog — Agent Plugins (Aug 6, 2026) · @OpenAIDevs announcement (X) · Agent Plugins Specification 1.0.0 · Compatible Clients · AWS Open Source Blog · Vercel blog · Read the full agency guide on Find AI Agency →
On August 4, 2026, Cloudflare announced Cloudflare Wallets and cloudflare.pay during Agents Week — the buyer side of agentic commerce, pairing with the Monetization Gateway (seller side, waitlist opened July 1, 2026). The service gives AI agents a human-readable wallet handle for paying APIs, MCP tools, content, and AI inference within limits set by the wallet's creator. Handle reservations opened on announcement day (Aug 4–5, 2026); full wallet access — onramping/offramping funds, issuing Virtual Wallets, live x402 purchases — arrives "in the coming months" with fees still undisclosed.
- Account Wallets — owned by humans/organizations using Cloudflare. Hold stablecoins, can add/remove funds, and delegate spending authority down to Virtual Wallets.
- Virtual Wallets — owned by AI agents, operate via API keys. An agent's maximum spend is capped by the limit set by the Account Wallet owner.
- Identity is optional: a Cloudflare account gets a unique web address / handle (e.g.
research.example.cloudflare.pay) that works as a stable ID when interacting with merchants — like DNS for agent identity, built on the agent's existing cryptographic keypair (Web Bot Auth).
- Spending cap / allowance (periodic) — the dominant cost variable. Cloudflare's worked example: "$100 per week budget for AI inference" per employee, or a $10 exploration budget for a cheap-to-try agent.
- Approved merchant allow-list — agents can only pay listed merchants.
- Maximum transaction size — per-purchase ceiling.
- Manual override required — agents that hit a limit request a human-approved escalation; they cannot approve it themselves. Planned anomaly controls notify admins of unusual spending (e.g. unexpectedly fast spending) for review.
- Prompt-injection immunity: caps are enforced at the wallet's API layer, not inside the model's system prompt — so a cap cannot be overridden by content the agent reads. This is the "outside-the-model enforcement standard" and the reason wallet rails matter for agencies running fleets of agents.
- Payments run on x402 — an open protocol that attaches payment instructions to HTTP requests (server returns HTTP 402 "Payment Required" with a price manifest; agent pays and retries with proof of payment). Originally introduced by Coinbase in 2025; Cloudflare co-founded the x402 Foundation.
- Micropayment economics: stablecoins (USDC on Coinbase's Base chain used as the primary example) settle in seconds at ~$0.0001 fees, making sub-cent per-call API billing viable vs credit-card interchange (1.5–3.5% + fixed fee).
- Fees, supported stablecoins/blockchains, custody partner, and onramp provider are all undisclosed — the calculator's wallet fee input defaults to 0% for exactly this reason. Treat any fee figure as an assumption until Cloudflare publishes pricing.
- Cost forecasting shifts from prediction to cap-constrained planning. Worst-case spend per agent becomes a known number (the cap) — the 2026 cost-blowup record (prompt-caching misses, retry loops, effort scaling) shows the risk is tail spend, not average spend, and wallet caps cut the tail.
- Client billing caps become programmable and auditable. Provision a Virtual Wallet per client project with an explicit cap, approved vendors, and transaction-size limits; over-limit requests route to a human for a documented override — an audit trail that doubles as a billing dispute shield.
- Reserve handles now. Handles are free to reserve today (first-come identity, like early DNS names); pricing comes later. Agencies reselling AI inference should reserve client-facing handles and watch the pricing announcement.
- Use the estimator above: enter agents × per-agent monthly spend, set a wallet fee if your rail has one, and set the creator-set cap to see capped vs uncapped spend — including the no-wallet-fee default and the exceeded-cap edge case.
Sources: Cloudflare Blog — The programmable wallet for the agentic Internet (Aug 4, 2026) · Cloudflare press release (Aug 4, 2026) · Help Net Security (Aug 5, 2026) · crypto.news (Aug 5, 2026) · TechTimes (Aug 4, 2026) — verified via parent research brief t_6e67321b.