Welcome to the forefront of conversational AI as we explore the fascinating world of AI chatbots in our dedicated blog series. Discover the latest advancements, applications, and strategies that propel the evolution of chatbot technology. From enhancing customer interactions to streamlining business processes, these articles delve into the innovative ways artificial intelligence is shaping the landscape of automated conversational agents. Whether you’re a business owner, developer, or simply intrigued by the future of interactive technology, join us on this journey to unravel the transformative power and endless possibilities of AI chatbots.
Latest news
Editions
Most guides to the best AI chatbot for customer support rank software products. Fewer ask the question that decides your bill: which model sits behind the chat window. In November 2026 that choice got harder, because four credible options shipped or changed price inside ten weeks.
As of 8 November 2026, Anthropic’s newest model is Claude Opus 5.5, which a November model ranking from GeoToolbox dates to 22 September. OpenAI followed on 26 October with GPT-6.1 Terra, priced at $1 per million input tokens. Google made Gemini 4 Argon generally available on 13 October with an introductory price that ends on 30 November. Z.ai open-sourced GLM-5.3-Flash on 26 August, and it undercuts all three.
This comparison puts those four side by side for one job: answering customer messages for an Irish business. We cover specs, prices, vendor benchmarks, a cost model for 1,000 chats, migration steps and a verdict. If you want the wider flagship fight, our Gemini 4 Argon vs Claude Opus 5.5 vs GPT-6.1 Sol test covers general reasoning. Here the question is narrower and more practical: what does each model cost per resolved conversation, and where does it break?
Don't miss new tech stories on Google
Add Tech Insider once in the Google app and our stories appear in your news suggestions.
Every figure in the table below comes from the vendors’ own announcement pages or from facts we confirmed on at least two fetched pages on 8 November 2026. Where a vendor has not published a number, the cell says so rather than guessing. That matters for support work, because token limits and retention terms often decide whether legal signs off.
Two patterns jump out. First, only Argon publishes a context window figure in the sources we could verify, and 1 million tokens is large enough to paste a full help-centre export into a single prompt. Second, GLM-5.3-Flash is the only one of the four you can run on your own hardware, which changes the data-protection conversation entirely.
Model names in this market change fast, so a quick status check helps. Claude Opus 5.5 is the first model in Anthropic’s 5.5 family. Claude Sonnet 5.5 arrived on 28 October and Claude Haiku 5.5 on 7 October, which makes Haiku the cheaper, faster sibling aimed at high-volume work. Claude Sonnet 5 remains available and is the model that became the default for Free and Pro users on 30 June.
On the OpenAI side, GPT-6.1 Terra sits alongside GPT-6 Astra, GPT-6.1 Sol and GPT-6 Luna. OpenAI positions Terra around fast computer use at scale and does not call it the most capable model in the family. Google announced Gemini 4 Carbon on 28 October as a model building on Argon, but our sources give no public release date, so we treat Argon as the Google option you can buy today. Gemini 3.x models, including Gemini 3.8 Flash, are the previous generation and we leave them out of the ranking.
We skip packaged helpdesk products such as Intercom Fin or Zendesk AI here. They bundle a model with ticketing, routing and analytics. This article is for teams building on a model directly, or choosing which model a vendor should be configured to use.
A support bot lives on the API, not in a consumer app, so token prices drive the economics. Pricing data aggregators such as Maximem’s LLM API pricing comparison track these figures weekly, but the vendor pages remain the source of truth. The table lists only prices published on official pages or confirmed on two fetched pages. All prices are in US dollars per million tokens and exclude Irish VAT.
Standard Opus 5.5 pricing deserves a hedge. Anthropic states that Opus 5.5 charges 20% less per token than Opus 5 for both input and output, and says the model costs 40% less to run on typical workloads because it uses fewer tokens per task. One secondary report puts the standard rate at $4 per million input tokens and $20 per million output tokens, which we treat as reported rather than confirmed. If that holds, Opus 5.5 costs the same per token as Argon will after 30 November.
Z.ai’s own announcement describes Flash as priced at one-tenth of GLM-5.3. Secondary reports put GLM-5.3 at $1.40 input and $4.40 output, and Flash at $0.15 input and $0.50 output. Both pairs are single-source, so use them for planning only and check the Z.ai documentation before you commit. Rate cards move often, and our vision pricing comparison covers the image side of the bill. That volatility is a good reason to build your bot so the model name is a config value.
Price per million tokens means little to a support manager. So we built a simple model. Assume each conversation sends 2,000 input tokens (system prompt, a few knowledge-base snippets, the customer’s messages) and returns 400 output tokens (the reply). That is our assumption, not a measured figure, so scale it to your own transcripts. The arithmetic below uses only the confirmed prices above.
Set side by side, the listed prices show a wide spread. GPT-6.1 Terra costs $1 per million input tokens and $5 per million output tokens, while Opus 5.5 and Gemini 4 Argon after its introductory period both cost $4 and $20. GLM-5.3-Flash is reportedly $0.15 and $0.50. Treat any per-chat cost estimate built on these rates as rough. Even so, the spread is striking, and a team handling 50,000 chats a month would feel it in the bill if the reported rates hold.
Caching changes the picture most for support bots, because the system prompt and policy text repeat on every turn. Terra’s cached rate is one-twentieth of its input rate. Argon’s introductory cached rate is a 95% discount. Anthropic lists cache reads at $0.20 per million tokens. Design your prompt so the static part comes first and stays identical, then caching does the work for you.
Token cost is only one line. Multi-turn chats resend history, so a 10-message thread costs far more than ten times a single reply. Retrieval, logging, human review and the cost of a wrong answer all sit outside the table. A cheap model that escalates 15% more tickets to humans can cost more overall than a pricier one. That is why the benchmark section matters, even though no vendor publishes a support-specific score.
We found no neutral, support-specific benchmark covering all four models with a method we could verify. What exists are vendor-run tests on adjacent skills. The table collects them from Anthropic, Google, Z.ai and OpenAI, and each one is the vendor’s own claim, so read it as a signal, not a verdict.
Three cautions apply. Anthropic compares Opus 5.5 with GPT-6 Astra and GPT-5.6 Sol, not with Terra, so the FrontierCode claim says little about the Terra matchup. Z.ai compares Flash with Opus 4.8, which is no longer Anthropic’s newest model. And Google’s 68% is a security score, not a conversation score. The weekly LLM Meter from Implicator is a useful third-party tracker for how these claims hold up once developers test them.
Our advice: run your own test. Pull 200 real tickets from the last quarter, strip personal data, split them into answerable and unanswerable, and score each model on resolution, correct refusal and tone. A single afternoon of that beats any leaderboard.
Anthropic describes Opus 5.5 as performing at the level of Claude Fable 5.1 on most work while costing 40% less to run than Opus 5. The company says the saving comes from two things: a lower per-token price and fewer tokens spent per task. For a support bot, the second part matters more. Verbose models burn output tokens on pleasantries, and output is the expensive side of every rate card.
Opus 5.5 is not the model most teams should put on the front line. It is the escalation model. Anthropic built a 5.5 family so that Opus handles hard reasoning while Claude Haiku 5.5 handles volume. Haiku 5.5 is described by Anthropic as its fastest, most capable small model, built for high-volume work such as summarisation, subagents and browser use. Claude Sonnet 5.5, announced on 28 October, runs more than 30% faster than Sonnet 5 and costs up to 30% less for most work. Our Opus vs Sonnet vs Haiku 5.5 breakdown covers the tier gaps in more detail.
Three details stand out for a business handling customer data. Opus 5.5 is available with zero data retention, as earlier Opus models were. It ships with watermarking measures that Anthropic says comply with the EU AI Act. And it runs on Amazon Web Services, Google Cloud and Microsoft Azure, so an Irish company already committed to one hyperscaler can buy it through that contract. The model is also no longer available with thinking switched off, according to Anthropic, so test response latency before you promise customers a snappy chat.
One warning on plans. Claude’s consumer tiers (Free, Pro, Max, Ultra) are for people, not for production bots. Anthropic reportedly changed the data-use option for Free, Pro and Max users on 15 October, allowing opt-in model-improvement use with five-year retention for participants. Customer conversations belong on the API or an enterprise agreement, where retention terms are contractual.
OpenAI launched GPT-6.1 Terra as a fast, efficient model for high-volume professional work, with image analysis and computer use. The standard API price is $1 per million input tokens, $0.05 per million cached input tokens and $5 per million output tokens. That makes Terra the cheapest of the three Western models in our cost model by a clear margin, even before caching.
OpenAI says Terra beats GPT-5.6 Terra on MMMU-Pro, a visual reasoning test, at every reasoning setting, and nearly matches GPT-6 Astra and GPT-6.1 Sol at much lower cost. That last claim is the interesting one for support teams. If Terra gets close to the flagship on screenshot-heavy tickets, you pay a fraction for most of the quality. OpenAI also added Terra to the model family powering dots, its always-on agent product, on 3 November.
Terra’s rollout is uneven. The launch post gives an API identifier, gpt-6.1-terra. OpenAI’s ChatGPT help centre, meanwhile, says Terra rolls out in ChatGPT Work and Codex for Plus, Pro, Business, Enterprise and Edu users, subject to account, region and workspace availability. Staff-facing use through ChatGPT is straightforward. For a customer-facing bot, confirm API access on your own account before you design around it.
If you also want staff to use ChatGPT directly, OpenAI lists Pro tiers at $100, $200 and $500 a month, and reportedly prices ChatGPT Business at $125 per user per month billed monthly or $100 on an annual plan. Those seat costs are separate from the API bill. For multi-step tasks such as refunds that touch several systems, see how the models behave in our agent mode comparison.
Google’s pitch for Argon rests on scale. The model has a 1-million-token context window, adjustable thinking levels and built-in Gemini API tools. For support, that means you can load an entire policy manual, product catalogue and past-ticket digest into one prompt instead of building a retrieval pipeline. That simplicity has value, though it also means you pay for those tokens on every call unless caching is on.
The price story is the catch. Google launched Argon with an introductory rate of $2 per million input tokens and $10 per million output tokens, with cached input at 95% off. After 30 November 2026, the $4 and $20 rate applies to the publicly accessible model, as PromptsRush’s model decision matrix also flags. That doubles both sides of the bill. Any pilot you run in November will look twice as cheap as the production bill from December.
Argon reached general availability on 13 October across the Gemini API, Google AI Studio, Google Cloud, Gemini for Business, the Gemini app and AI Mode in Search, with consumer access for Google AI Pro and Google AI Ultra subscribers. On 28 October Google announced Gemini 4 Carbon, which builds on Argon. We found no public release date for Carbon, so do not plan a launch around it. For a wider capability view, see our Argon vs GPT-6 Astra vs Grok 4.7 test.
Z.ai released GLM-5.3-Flash as an open-source model with 320B total parameters and 18B active per inference. It is the first native multimodal model in the GLM-5 series. AutoClaw positions Flash as a lower-compute model for high-frequency and agent use, while GLM-5.3 remains the higher-end choice for complex, long-horizon engineering. Vendors choose their own comparisons, so verify any claim on your own tickets.
Z.ai positions Flash for high-frequency, lighter tasks and keeps the flagship GLM-5.3 for complex, long-running engineering work. That split mirrors the Haiku and Opus division on the Anthropic side, and it suits support, where most tickets are simple. One reported adoption signal: Revolut was using self-hosted GLM 5.3 as a coding model by 5 November, according to a weekly industry tracker. That is a coding use, not a support deployment, but it shows a regulated European fintech trusting the family with self-hosting.
Open weights give Irish firms an option no other model here offers: run the model inside your own EU cloud account, so customer messages never leave your control. The trade-off is operations. You need GPU capacity, monitoring and someone on call. If you instead call Z.ai’s hosted API, you are sending EU personal data to a provider outside the EU, which your data protection officer will want to review. Our EU AI Act comparison covers the labelling duties that apply to customer-facing chatbots whichever model you pick.
Our sources also say little about GLM’s English-language or Irish-language support quality. Do not assume parity with the Western models. Test it with real transcripts before any customer sees it.
Model choice sits inside a bigger shift. Industry surveys show support leaders under pressure to deploy AI quickly, and customers more open to it than many assume. We use Gartner’s published findings here because they come with sample sizes and dates. For more model coverage, browse our AI and machine learning section.
The budget picture comes from Gartner’s survey of 265 service and support leaders, published in a Gartner press release: “Seventy-seven percent of service and support leaders reported pressure from senior executives to deploy AI, while 75% reported increased AI budgets compared with the previous year.” Pressure plus budget explains the rush. It also explains why so many pilots launch on promotional pricing and then stall when the bill arrives.
These are 2024 and 2025 figures, so treat them as direction, not a measurement of November 2026. The takeaway: demand is real, and the winner is the model you can afford to run at volume without a quality collapse.
Public support-deployment case studies for these brand-new models are scarce, since all four are under three months old at most. The examples below are documented uses of each model family, and we say plainly where the use is not customer support.
None of these is a published support-bot win with ticket numbers. If a vendor offers you one, ask for the sample size, the ticket mix and who scored the answers.
The right model depends on volume, risk and data rules. These recommendations are our analysis built on the prices and positioning above. Each is a starting point for your own test.
Pick GPT-6.1 Terra or Claude Haiku 5.5. At $4 per 1,000 chats, Terra costs about $20 a month for a shop with 5,000 conversations. Put order status, returns policy and delivery times in a cached system prompt and keep a human fallback for anything with a refund attached.
Pick Claude Opus 5.5 through your existing cloud contract, or self-host GLM-5.3-Flash. Zero data retention and enterprise terms suit a bank, insurer or clinic. Self-hosting suits teams with GPU skills that need messages to stay in their own EU account. Either way, involve your data protection officer before launch.
Pick Gemini 4 Argon for its 1-million-token context, but budget at the post-30 November rate of $4 and $20. If your content changes weekly, long-context loading can replace a fragile retrieval pipeline. Test accuracy on questions buried deep in the documents, because long context does not guarantee the model uses all of it.
Pick GLM-5.3-Flash self-hosted, or Terra with aggressive caching. A retailer with 500,000 chats a month would pay about $2,000 on Terra without caching and, at the reported Flash rate, about $250. The gap pays for a lot of GPU time. Check how each model handles Irish English phrasing and, if you serve Gaeilge speakers, read our Gaeilge benchmark comparison first.
Pick GPT-6.1 Terra first and test Argon second. OpenAI’s visual reasoning claim on MMMU-Pro targets this case, and Terra’s price lets you afford image-heavy conversations. Tie the bot into your ticketing and CRM tools, and read our app integrations comparison to see which assistants connect to which services.
Pick a cheap front line (Haiku 5.5, Terra or Flash) and route hard cases to Opus 5.5. If 90% of chats stay on the cheap tier, your blended cost stays close to the cheap model while complex complaints get the strongest reasoning. Define the routing rule up front, such as sentiment, refund value or two failed answers in a row.
Whether you are leaving a Gemini 3.x model, an older GPT or Claude version, or a rules-based bot, the steps are the same. Do them in order and keep the old setup live until the new one proves itself.
The thresholds above are placeholders, so tune them against your own shadow-run data. Keep the log retention short and write it into your privacy notice. If you switch from a consumer subscription to an API contract mid-project, tell your legal team, because retention and training terms differ between the two.
The pattern is familiar from every cloud decision. You trade money against control against convenience, and no model wins all three. What changed in autumn 2026 is that the cheap option got good enough to be taken seriously, and the premium option got cheaper to run.
For most Irish businesses, start with GPT-6.1 Terra as the front line and Claude Opus 5.5 as the escalation tier. Terra’s $1 and $5 pricing keeps a 1,000-chat bill low under our assumptions, and caching lowers it further. Opus 5.5 earns its place on the hard tickets, where Anthropic’s workflow and cost-per-task results suggest it wastes fewer tokens than its predecessor.
Two cautions before you sign anything. First, every benchmark in this article is a vendor claim on a task that is not customer support, so your own 200-ticket test outranks all of them. Second, the Argon price doubling on 30 November shows that any launch-month cost comparison can expire within weeks. Build the bot so you can swap models in an afternoon, and re-run the cost model each quarter.
If you remember one thing, make it this: the best AI chatbot for customer support is the cheapest model that passes your own ticket test, with a stronger model waiting behind it for the cases that matter. Right now, on published prices, that is Terra in front of Opus 5.5, with GLM-5.3-Flash as the self-hosted challenger worth a serious trial.
Among the confirmed prices, GPT-6.1 Terra is listed at $1 per million input tokens and $5 per million output tokens. GLM-5.3-Flash is reportedly cheaper still, but its exact rate is single-source.
Yes. Google’s introductory rate of $2 input and $10 output per million tokens ends after 30 November 2026. From then the public model costs $4 and $20, which doubles both sides of the bill.
Consumer plans are meant for individuals, so use the API or an enterprise or business agreement for customer-facing bots. Retention and training terms differ between plan types, and Anthropic reportedly updated its data-use option for Free, Pro and Max users in October. Read the current terms before you route customer data anywhere.
Anthropic states that Opus 5.5, like previous Opus models, is available with zero data retention. Availability depends on your agreement, so confirm it in writing with Anthropic or your cloud provider.
That depends on how you run it. Self-hosting the open weights inside your own EU cloud account keeps messages under your control. Calling a hosted API from a provider outside the EU raises data-transfer questions for your data protection officer. This is not legal advice, so take it to your compliance team.
Google announced Gemini 4 Carbon on 28 October 2026 as a model that builds on Gemini 4 Argon. Our sources show no public release date, so plan around Argon for now.
Take 200 to 500 anonymised closed tickets, label which ones your documents could answer, and score each candidate on resolution, correct refusal, tone and escalation. Price the result at the rate you will actually pay, not a launch discount.
The EU AI Act includes transparency duties for systems that interact directly with people, and good practice is to say so plainly at the start of every chat. Get legal advice on the exact obligations and timing for your business.
Niamh Kelly is the iGaming Editor at Tech Insider, where she previously worked as a freelance fashion journalist for The Irish News for three years and honed her media skills during her time at the BBC. At Tech Insider, she leads Ireland’s coverage with hands-on experience testing consumer and business technology, delivering in-depth analysis on AI, cybersecurity, cloud computing, and hardware trends shaping the future. Kelly was featured in RSVP online as part of their “Women of Style” series and has interviewed notable figures such as Katie Price and Louise Redknapp for major beauty product launches.
Tech Insider delivers in-depth coverage of the technologies shaping the future: AI, cybersecurity, cloud computing, hardware, and the trends that matter.
