.webp)
A UK CX Director we spoke to in June greenlit her first AI chatbot. Intercom Fin, deployed across 6,000 tickets per month, projected to save 40 percent of her support cost. Contract signed. Fin went live in month one. Bill in month three was 2.4x the quoted number.
The reason was straightforward. Fin bills per outcome. Her tickets were converting to Fin resolutions at 55 percent, higher than projected. Every extra resolution was billable. Her chatbot was working better than expected and it was killing her budget.
That is the state of AI chatbot cost conversations in 2026. Vendors quote what they know, buyers estimate what they hope, and month three often disagrees with both. This article is the honest read on real 2026 chatbot cost, token math with current OpenAI and Anthropic pricing, and where hidden costs land.
The Four Chatbot Categories and What They Cost in 2026
No-code builders. Chatbase, Botpress, Voiceflow, Landbot. Drag-and-drop flows plus LLM-powered responses on top of your knowledge base. Cost: £0 to £500 per month depending on volume. Best for small businesses, marketing sites, and internal tools. Not built for high-volume production customer support.
Mid-market SaaS. Intercom Fin, Zendesk AI, HubSpot AI agents. Priced per resolution (Fin at roughly $0.99 per outcome), per user, or per conversation. Monthly bills land £500 to £5,000 for most mid-market deployments. Includes native integration with the surrounding CRM and helpdesk. Best for D2C, SaaS support teams, and any business already on Intercom, Zendesk, or HubSpot.
Enterprise platforms. Ada, Decagon, Kore.ai, Salesforce Einstein Bots, Yellow.ai. Enterprise contracts with named accounts, custom integrations, dedicated CSM. Priced £5,000 to £50,000 per month at scale. Best for large operators with unusual compliance requirements or unusual integration needs SaaS cannot cover.
Custom build. Direct on OpenAI or Anthropic APIs plus your own orchestration.
Proof of concept: £15k to £40k over 6 weeks. Narrow use case, no live integrations, no compliance review.
Pilot: £30k to £90k over 3 to 4 months. Live knowledge base retrieval, one integration, evaluation harness.
Production: £90k to £300k over 5 to 9 months. Multi-integration, compliance review, monitoring, drift detection, escalation workflow.
For most operators under 500,000 monthly conversations, buying off-the-shelf beats building. Custom becomes defensible when SaaS cannot cover the workflow or above that scale where SaaS math tips.
The Real Token Cost Math: GPT-5.6 vs Claude
Current 2026 API pricing verified against the OpenAI pricing page and the Anthropic Claude pricing page in August 2026.
OpenAI GPT-5.6 family:
Sol (flagship): $5 input / $30 output per 1M tokens
Terra (balanced): $2 / $12 per 1M
Luna (cost-optimised): $0.20 / $1.20 per 1M
Anthropic Claude:
Opus 4.8: $5 / $25 per 1M
Sonnet 5: $3 / $15 per 1M
Haiku 4.5: cost-tier below Sonnet, sub-$1 input
Real per-conversation cost. Assume 1,000 tokens in and 500 tokens out per turn (typical support chatbot exchange).
Flagship (GPT-5.6 Sol or Claude Opus 4.8): $0.005 input + $0.015 output = $0.02 per turn
Mid-tier (GPT-5.6 Terra or Claude Sonnet 5): $0.002 + $0.006 = $0.008 per turn
Cheap tier (GPT-5.6 Luna): $0.0002 + $0.0006 = $0.0008 per turn
A 10-turn conversation costs $0.20 / $0.08 / $0.008 respectively. Prompt caching (available on both providers) cuts input tokens to 10 percent of standard on repeated context, taking the flagship 10-turn conversation to roughly $0.14. Batch API cuts another 50 percent on non-realtime workloads.
For 10,000 conversations per month, monthly inference cost lands £500 to £3,000 depending on model choice and caching discipline. For 1 million conversations per month, £30k to £500k. Model routing (Luna for classification, Terra for standard responses, Sol reserved for hard cases) cuts blended cost by 40 to 70 percent.
Where Total Cost of Ownership Actually Comes From
Token cost is the visible number. TCO is where the surprise lives. Six components rarely fully priced.
RAG infrastructure. Vector database (Pinecone, Weaviate, or pgvector), embedding generation, index rebuilding when knowledge base changes. £500 to £3,000 monthly at moderate volume.
Evaluation harness. Golden test set of 100 to 200 known-outcome conversations. Weekly re-runs against latest model behaviour. Build cost £5k to £15k; ongoing tooling £200 to £1,000 monthly.
Monitoring and guardrails. Real-time content filtering, PII detection, hallucination monitoring, drift detection. Tooling £500 to £3,000 monthly plus engineering time.
Integration development. CRM sync, helpdesk sync, calendar integration, product catalogue integration. Each mature integration costs £8k to £25k to build and £200 to £1,000 monthly to maintain.
Compliance review. GDPR data flows, if you handle healthcare or financial data add HIPAA or PCI review, DPA with model provider. £5k to £30k one-time; ongoing legal £200 to £1,000 monthly.
Prompt engineering and iteration. Not zero. Realistic budget for a serious chatbot in production is £2k to £8k monthly in engineering time on prompt improvements, guardrail tuning, and evaluation.
Total TCO commonly runs 50 to 150 percent above the pure token bill. Build the TCO number into your budget before you sign. See our related read on custom AI software development cost for context.
When Custom Build Beats Off-the-Shelf
Four profiles where custom earns its money.
Unusual workflow SaaS treats as edge case. Your business has a specific escalation logic, a specific tone requirement, or a specific tool the chatbot must call. Intercom and Zendesk are architected for the middle 80 percent of use cases. If you sit in the tail, custom fits.
Integrations no vendor offers. Bespoke ERPs, industry-specific CRMs, proprietary product catalogues. Custom builds these once and owns the integration.
Data control requirements. Healthcare data under HIPAA, financial data under PCI DSS, EU citizen data under strict interpretations of GDPR. Some SaaS vendors cannot meet the requirement. Custom gives you the compliance posture.
Scale above 500,000 monthly conversations. At this scale, per-outcome SaaS billing outstrips the amortised cost of a custom build. The math starts favouring build somewhere between 500k and 1M monthly conversations.
Below those profiles, off-the-shelf wins. Do not build custom for the vanity of ownership.
What We Learned Building AI Feedback in IELTSArena
IELTSArena is our AI IELTS preparation platform. The writing feedback feature is a single-shot LLM call against IELTS band descriptors. Not a chatbot, but the discipline transfers directly.
We hold a golden set of essays graded by trained IELTS examiners. Every week and after every model upgrade, we re-run our AI feedback against this set and track correlation with human grades. Twice in the last twelve months we caught meaningful drift between model versions we would not have seen from user feedback alone.
Apply the same pattern to chatbots. Build a set of 100 to 200 real conversations with known correct outcomes (resolution, escalation, tool call). Re-run your chatbot against the set weekly. When resolution rate drops on the golden set, investigate before your users do. Every chatbot programme that fails in year two skips the harness.
You can see IELTSArena in our portfolio. If you want to talk about a chatbot build against real cost bands, get an AI chatbot cost estimate.
Hidden Costs Rarely in a Quote
Five costs that surprise buyers three months in.
Knowledge base preparation. Chatbot quality depends on knowledge quality. Cleaning, deduplicating, and structuring 500 to 2,000 KB articles typically takes 3 to 8 weeks of content ops time. Rarely priced in.
RAG index rebuilding. Every time your knowledge base changes materially, the vector index needs a rebuild. Small changes are cheap; full rebuilds cost engineering time and often break edge cases. Budget monthly.
Escalation workflow to humans. Building the handoff from chatbot to human agent with context preserved (conversation history, retrieved knowledge, proposed action) is a real engineering effort. £5k to £15k that vendor quotes usually skip.
Compliance review. Especially in healthcare, finance, and regulated industries. £5k to £30k one-time plus ongoing legal review.
Prompt tuning at scale. After launch, every real production interaction reveals prompt improvements. Realistic budget is £2k to £8k monthly in engineering for the first 6 to 12 months post-launch.
Frequently Asked Questions
How much do OpenAI or Claude tokens cost per conversation?
Assuming 1,000 input plus 500 output tokens per turn, a 10-turn conversation costs $0.20 on flagship models (GPT-5.6 Sol, Claude Opus 4.8), $0.08 on mid-tier (GPT-5.6 Terra, Claude Sonnet 5), and $0.008 on cost-optimised tiers. Prompt caching cuts input tokens to 10 percent on repeated context.
Should I use a chatbot builder or a custom build?
Chatbot builders (Chatbase, Botpress, Voiceflow) at £0 to £500 monthly work for small business, marketing, and internal use. Mid-market SaaS (Intercom Fin, Zendesk AI) at £500 to £5,000 monthly works for most support and CX use cases. Custom build (£90k to £300k plus ongoing) is defensible when SaaS treats your workflow as edge case, when integrations require it, or above 500,000 monthly conversations.
What is RAG and how much does it add?
RAG (Retrieval Augmented Generation) means the chatbot retrieves relevant content from your knowledge base before generating a response. It sharply reduces hallucination. Cost: vector database plus embedding generation lands £500 to £3,000 monthly at moderate volume. Setup: 3 to 6 weeks of engineering to build cleanly.
How do I estimate a monthly chatbot bill?
Model per-turn cost times average turns per conversation times monthly conversations, then multiply by 2 to 2.5 for TCO. Example: 10,000 conversations at 10 turns each on Claude Sonnet 5 costs about £800 in tokens, roughly £2,000 all-in with RAG, monitoring, and integrations.
What is the cheapest way to test AI chatbots?
Start with a no-code builder (Chatbase or Voiceflow) at £30 to £150 monthly for 100 to 500 conversations. Prove the use case. Then move to mid-market SaaS or custom build once you understand the workflow. Skipping this step usually costs £30k to £100k in a rebuild.
The One Thing to Remember
Chatbot cost in 2026 is not one number. The token bill is roughly a third of TCO. RAG, monitoring, integrations, and prompt engineering are the other two thirds. Budget for all three or your surprise arrives in month three.
If you want a candid conversation about your chatbot against real cost bands, browse our AI development services or come to the estimate.
.webp)
.webp)