Small AI models compared: cost and quality for routine work

Many routine tasks in companies can be handled with small AI models: sorting emails, checking production orders, classifying actions on a website or monitoring a status around the clock. In terms of list prices, the providers are close together. The models differ markedly in how often they guess, how reliably they use tools and how many reasoning tokens they consume. The figures come from the providers’ price lists, from independent benchmarks and from our own hands-on test in Hawk IVO Force, as of 20 October 2026.

Costs are shown here as a cost index instead of dollars. List prices change frequently, and the providers’ prices are not Hawk IVO Force prices. The index shows how expensive a model is relative to the others and remains easy to read even when prices change.

Which models are compared

OpenAI, Anthropic, Google and Amazon do not state sizes for their small models. Here, small means the cheapest current tier of a model family, designed for speed and volume.

• GPT-6 Luna from OpenAI, available since 22 September 2026: the smallest tier of the GPT-6 family alongside Sol and Astra. According to OpenAI, its predecessor GPT-5.6 Luna corresponded to the former nano class. Available directly from OpenAI and via Microsoft Azure and Amazon Bedrock.
• Claude Haiku 5.5 from Anthropic, available since 7 October 2026: successor to Claude Haiku 4.5, available via the Claude Platform, Amazon Web Services, Google Cloud and Microsoft Azure.
• Gemini 3.5 Flash-Lite from Google, available since 21 July 2026: the smallest current Gemini tier. For comparison, Gemini 3.8 Flash is included as the next larger tier.
• Amazon Nova 2 Lite and Amazon Nova Micro, both only on Amazon Bedrock: Nova 2 Lite since December 2025, Nova Micro as a text-only model since December 2024.
• Chinese models: DeepSeek V4.1 Flash, Qwen3.8 Flash from Alibaba, GLM-5.3-Flash from Z.ai and MiniMax M3.1 Flash. Kimi K2.6 from Moonshot AI costs several times as much and serves as a comparison in the hands-on test.

Costs compared: the cost index

The cost index prices the standard task at the providers’ list prices, and the cost level gives a rough summary of it.

Model	Cost index ↓, lower is better	Cost level ↓, lower is better	Surcharge for long prompts
Amazon Nova Micro	24	1	none known
GPT-6 Luna	76	2	above 272,000 tokens: input ×2, output ×1.5
Claude Haiku 5.5	76	2	above 100,000 tokens: input ×5, output ×5
Qwen3.8 Flash	94	2	none known
GLM-5.3-Flash	97	2	none known
MiniMax M3.1 Flash	103	2	none known
DeepSeek V4.1 Flash	207	3	none
Gemini 3.5 Flash-Lite	297	3	none
Amazon Nova 2 Lite	297	3	none known
Gemini 3.8 Flash	569	4	none, double base price from January 1, 2027
Claude Sonnet 5.5 (for comparison)	1,517	5	none
Claude Opus 5.5 (for comparison)	3,034	5	none

Amazon Nova Micro costs a quarter of the median, GPT-6 Luna and Claude Haiku 5.5 around three quarters and Gemini 3.8 Flash almost six times as much. The large models Claude Sonnet 5.5 and Claude Opus 5.5 cost 15 to 30 times the median.

The price jump for Claude Haiku 5.5

With Claude Haiku 5.5, the length of each individual request determines the rate. Above 100,000 tokens in the prompt, the entire request costs five times as much. Cached parts of the prompt count towards the length. Anyone who loads an entire order backlog into a single prompt therefore pays the high rate, while anyone who splits it into sections below 100,000 tokens pays the low one. GPT-6 Luna has a similar threshold at 272,000 tokens with a smaller surcharge. Gemini Flash and Flash-Lite have no such threshold, and neither do the larger Claude models from generation 4.6 onwards.

• Gemini Flash and Flash-Lite
• GPT-6 Luna
• Claude Haiku 5.5

Price factor on the input base price ↓

Prompt length in tokens

The surcharge applies to the entire request. For GPT-6 Luna, output rises to 1.5 times. Source: OpenAI, Anthropic, Google, as of October 20, 2026.

Chart data

Model	Surcharge for long prompts
GPT-6 Luna	above 272,000 tokens: input ×2, output ×1.5
Claude Haiku 5.5	above 100,000 tokens: input ×5, output ×5
Gemini 3.5 Flash-Lite	none
Gemini 3.8 Flash	none, double base price from January 1, 2027
Amazon Nova 2 Lite	none known
Amazon Nova Micro	none known
DeepSeek V4.1 Flash	none
Qwen3.8 Flash	none known
GLM-5.3-Flash	none known
MiniMax M3.1 Flash	none known

Same price in all clouds, surcharge for data residency

Across the major clouds, prices are close together. With global processing, GPT-6 Luna costs the same on Microsoft Azure and Amazon Bedrock as it does directly from OpenAI. Microsoft Foundry bills Claude models at Anthropic’s list prices. Amazon Bedrock and Google Cloud have their own price lists, and for their regional endpoints Anthropic states a surcharge of 10% compared with the global ones. What mainly costs more is processing in a fixed region: GPT-6 Luna in the Azure EU Data Zone around 20% more, Gemini in the Google Cloud EU multi-region around 10% more and Nova 2 Lite via the EU route in Frankfurt around 40% more than globally in the US.

What saves or costs extra

• Batch: If you do not need results immediately, you pay half in batch mode with OpenAI, Anthropic, Google and Amazon. DeepSeek halves its price outside its peak hours, although on weekdays these cover a large part of the German morning.
• Caching: Recurring prompt parts such as instructions, categories and examples can be cached. When read from the cache, they cost around one tenth of the input price with OpenAI, Anthropic and Google, and one quarter with Amazon Nova.
• Tokens are not the same size everywhere: Claude Haiku 5.5 counts the same text with around 30% more tokens than its predecessor. Equal prices per token therefore do not mean exactly equal costs per email.
• Model changes: Google deprecated Gemini 3.7 Flash on 8 October 2026, just under eight weeks after general availability, and redirects requests to Gemini 3.8 Flash. Its price doubles on 1 January 2027. Prices and model versions therefore belong in every operational plan.

Reasoning tokens: the invisible cost driver

Almost all current small models can reason before answering. These reasoning tokens do not appear in the answer, but they are billed as output, and depending on the model, output costs three to eight times as much as input. How much a model reasons can usually be configured:

• GPT-6 Luna: from none to max, default medium
• Claude Haiku 5.5: from low to max, default medium
• Gemini 3.5 Flash-Lite: from minimal to high, default minimal
• Amazon Nova 2 Lite: off or low to high, default off
• DeepSeek V4.1 Flash: with or without reasoning, default with
• GLM-5.3-Flash: reasoning cannot be switched off

The difference is considerable. For the full Artificial Analysis benchmark set, GPT-6 Luna cost just over six times as much at the highest level as without reasoning, and Claude Haiku 5.5 almost ten times as much at the highest level as at the low level. In our hands-on test, Kimi K2.6 generated 3,100 to 7,300 output tokens for the same tasks, whereas Gemini 3.7 Flash generated 360 to 580. Waiting time also increases: without reasoning, GPT-6 Luna delivers the first answer token after around 0.8 seconds, at the highest level after around 90 seconds.

For classification and extraction, the lowest level is usually sufficient. For Gemini 3.5 Flash-Lite, Google explicitly recommends minimal for classification, routing and JSON extraction, but medium or high for sub-agents with multi-step tool use, because otherwise the model ends tool chains too early. The most reliable way to find the right value is a test with your own data.

• GPT-6 Luna
• Claude Haiku 5.5
• other small models

Intelligence Index ↑

Cost index of the full test run, logarithmic ↓

Lines connect the lowest and highest measured reasoning level of the same model. The cost index includes the reasoning tokens actually used. Median of the points shown = 100. No test-run costs are available for Amazon Nova. Source: Artificial Analysis, as of October 20, 2026.

Chart data

Model	Reasoning level	Cost index of the test run ↓, lower is better	Intelligence Index ↑, higher is better
GPT-6 Luna	none	6.0	18.5
Claude Haiku 5.5	low	11	29.5
GPT-6 Luna	max	38	38.1
MiniMax M3.1 Flash	default	89	42.7
GLM-5.3-Flash	max	96	41.8
Claude Haiku 5.5	max	104	43.4
Qwen3.8 Flash	default	114	39.8
Gemini 3.5 Flash-Lite	default	115	22.2
DeepSeek V4.1 Flash	max	140	45.1
Gemini 3.8 Flash	high	510	40.9

Hallucination: knowledge questions or working from source material

How much a model hallucinates depends on the task. Two questions need to be kept apart: Does the model guess when it does not know something? And does it stick to the information it is given?

Knowledge questions without a source

With AA-Omniscience, Artificial Analysis measures how often a model gives a wrong answer to specialist questions instead of admitting that it does not know the answer. The hallucination rate is the share of wrong answers among all questions not answered correctly. Lower is better.

Model	Answered correctly in % ↑, higher is better	Hallucination rate in % ↓, lower is better
GLM-5.3-Flash	28	28
Gemini 3.5 Flash-Lite	30	34
Claude Haiku 5.5	33 to 36	40 to 45
Qwen3.8 Flash	25	45
Gemini 3.8 Flash	55	55
Amazon Nova Micro	10	65
DeepSeek V4.1 Flash	43	67
GPT-6 Luna	32 to 44	77 to 85
Amazon Nova 2 Lite	14 to 19	86 to 91
MiniMax M3.1 Flash	56	93

Hallucination rate in percent ↓

Share of wrong answers among all questions not answered correctly (AA-Omniscience). The light part of a bar shows the range across reasoning levels. Source: Artificial Analysis, AA-Omniscience, as of October 20, 2026.

GLM-5.3-Flash, Gemini 3.5 Flash-Lite and Claude Haiku 5.5 admit comparatively often that they do not know an answer. GPT-6 Luna, Nova 2 Lite and MiniMax M3.1 Flash, on the other hand, mostly guess. MiniMax answers the most questions correctly, but for the rest it is almost always wrong with an invented answer.

Working with provided data

For routine work, the second question usually matters more, because the email, order or ticket is available to the model. Here the picture looks different. In the grounding test of the FACTS leaderboard by Google DeepMind, which checks whether answers are supported by a provided document, GPT-6 Luna achieves 81 out of 100 points, Gemini 3.8 Flash 73 and Gemini 3.5 Flash-Lite 71. Other providers are missing from this leaderboard, which Google runs itself. In Vectara’s summarization test, Amazon Nova 2 Lite and Nova Micro add something that is not in the source text in only around 5% of summaries, even though they guess particularly often on knowledge questions. Kimi K2.6 reaches around 11% there. The other current models in the comparison have not yet been measured.

Small models should therefore work from source material, not from memory. If they are given the data, are explicitly allowed to answer “unknown” or null and are checked against a fixed schema, the risk of invented information can be limited well. In our hands-on test, no model invented a missing value.

Tool use: tools across multiple steps

For agents, what matters is whether a model calls tools correctly across multiple steps. Independent measurements for the current small models are still incomplete. With AutomationBench-AA, Artificial Analysis tests whether a model completes realistic business processes across several simulated applications such as email, spreadsheets, CRM and a ticket system, and with Terminal-Bench 4.0 it tests work in a command line. Most measurements were taken at a high or the highest reasoning level. The values are percentages of tasks solved.

Model	AutomationBench-AA in % ↑, higher is better	Terminal-Bench 4.0 in % ↑, higher is better
MiniMax M3.1 Flash	73	26
DeepSeek V4.1 Flash	67	34
GLM-5.3-Flash	60	33
Gemini 3.8 Flash	60	20
Qwen3.8 Flash	56	25
GPT-6 Luna	53	13
Claude Haiku 5.5	35	33
Gemini 3.5 Flash-Lite	25	1
Amazon Nova 2 Lite	not measured	not measured

AutomationBench-AA, tasks solved in percent ↑

Realistic workflows across several simulated applications such as email, spreadsheets, CRM and ticketing, mostly measured at a high or the highest reasoning level. Amazon Nova 2 Lite was not measured. Source: Artificial Analysis, AutomationBench-AA, as of October 20, 2026.

Vendor figures add to the picture. Anthropic reports around 72% for Claude Haiku 5.5 in the computer use test OSWorld 2.1, and around 49% for GPT-6 Luna. Amazon states 60% for Nova 2 Lite in the Berkeley Function Calling Leaderboard, while the leaderboard itself lists 27% and only 2% for multi-step calls. The Chinese Flash models solve tool tasks well, but they consume many reasoning tokens to do so. A technical detail for developers: with GPT-6 Luna, tool calls with reasoning only work via the Responses API, and via Chat Completions only without reasoning.

JSON and classification

For routine work, a model should answer in a fixed format. Several providers enforce a JSON schema while the answer is being generated:

• GPT-6 Luna: With a strict JSON schema, the answer always complies with the schema according to OpenAI, except in the case of a refusal or a truncated answer.
• Claude Haiku 5.5: Structured Outputs also enforce the schema, on Amazon Bedrock via the bedrock-runtime endpoint.
• Gemini 3.5 Flash-Lite and 3.8 Flash: JSON schema is supported, but Google recommends always checking the values. Very large schemas may be rejected.
• Amazon Nova 2 Lite and Nova Micro: The Bedrock feature Structured Outputs is not available. The workaround is a tool schema with at most two levels.
• DeepSeek, Qwen and GLM: JSON output is supported according to the providers. There is no information for MiniMax M3.1 Flash.

A valid schema does not guarantee correct values. There is no independent comparison test of the current small models for classification. Accuracy depends heavily on your own categories and can only be measured with your own examples.

Hands-on test in Hawk IVO Force

On 20 October 2026, we tested four models in Hawk IVO Force with two typical routine tasks. Each model ran in its own agent without knowledge access and without tools, with the same short instruction. Each task ran twice.

• Email triage: twelve German-language emails from a manufacturing company, including a complaint, two phishing emails, an out-of-office reply and automatic status messages. The models were asked for the category, the priority, whether a reply is needed, and the order number, only if it appears verbatim. Item, invoice and ticket numbers served as traps. Category, need for a reply and order number were scored, 72 values in total.
• Production orders: four orders from an ERP export and a note from production planning. The models were asked for fields, risk and justification as JSON. Four pieces of information were deliberately missing, including a delivery date, a machine and a capacity in hours. One order could not be completed on time because the material only arrives after the delivery date.

Model	Email triage	Production orders	Output tokens per run ↓
GPT-6 Luna	no errors	no errors	760 to 1,330
Gemini 3.7 Flash	no errors	no errors, once wrote “Gehaeuse” instead of “Gehäuse”, without the umlaut	360 to 580
Kimi K2.6	no errors	violated the schema once, risk not assessed	3,100 to 7,300
GPT-4.1 mini, older generation	7 of 72 values wrong	no errors	300 to 480

GPT-4.1 mini would have replied to both phishing emails and taken invoice and ticket numbers as the order number. The current generation no longer makes these mistakes. In one run, Kimi K2.6 also applied the instruction to set missing information to null to the risk assessment. Per email, Gemini Flash cost around five times as much as GPT-6 Luna at list prices in the test, and Kimi K2.6 around thirty times as much because of its many reasoning tokens.

The test is a sample with two runs, not a benchmark. Claude Haiku 5.5 and Amazon Nova were not available in our test environment. Gemini 3.7 Flash has since been deprecated. Google redirects the requests to Gemini 3.8 Flash.

What routine work costs per month

Four typical use cases with these assumptions:

• Office: 80 emails per working day on 22 working days, including status messages. Per email, 3,000 input tokens for instructions, categories, examples and the email with its thread, plus 500 output tokens for classification and a short draft reply.
• Production: 500 open production orders, reassessed in each of the three shifts, on 30 days. Per assessment, 4,000 input tokens for the order, routing, material and capacity, plus 600 output tokens for risk, bottleneck and recommendation.
• Website: 1,500 visits per day with an agent that observes actions, classifies them and gives a recommendation. Per visit, six classifications and one recommendation, together 16,000 input tokens and 1,000 output tokens.
• Continuous task: An agent checks a status every five minutes around the clock, such as incoming tickets, machine messages or delivery status. Per check, 5,000 input tokens and 300 output tokens.

The cost volume shows how much model cost a use case causes per month relative to the office case, measured by the median of the small models:

Use case	Runs per month	Tokens per run, input / output	Cost volume, office = 1 ↓, lower is better
Office	1,760	3,000 / 500	×1
Production	45,000	4,000 / 600	×33
Website	45,000	16,000 / 1,000	×104
Continuous task	8,640	5,000 / 300	×6

At 80 emails a day, model costs play practically no role, even with a large model such as Claude Opus 5.5. What is expensive here is a mistake, such as an overlooked complaint or a phishing email that gets answered. Reliability therefore matters more than price. In production, on the website and for continuous tasks, by contrast, volumes grow quickly. Between a small and a large model there is then a factor of 20 to 40, between GPT-6 Luna and Gemini 3.8 Flash a factor of around 7. More reasoning tokens increase output costs. Caching reduces input costs, especially on the website, where the page context stays the same for every request.

A single prompt with the entire order backlog is the most expensive option. With Claude Haiku 5.5, a request with 300,000 tokens costs five times as much as the same volume split into sections below the threshold. At the same time, the hit rate drops with very long contexts: according to Google, Gemini 3.5 Flash-Lite reaches 72% at 128,000 tokens in the long-context test MRCR and 21% at one million tokens. An assessment per order or per group is better.

Small models as sub-agents

Small models are particularly well suited as sub-agents. A larger model plans the task, breaks it down and passes clearly defined subtasks to small models, such as reading documents, extracting data, classifying entries or summarizing search results. Each sub-agent works with a small, fresh context and returns a concise result, which the larger model checks and combines. Anthropic explicitly positions Claude Haiku 5.5 for this role.

This pays off in several ways:

• Cost: With Claude Haiku 5.5, an input token costs one fortieth of what Claude Opus 5.5 charges. Reading, which consumes the most tokens, therefore becomes inexpensive.
• Speed: Many sub-agents work in parallel.
• Overview: The planning model only sees condensed results instead of thousands of pages of raw material.
• Price thresholds: Small subtasks stay below the 100,000-token threshold of Claude Haiku 5.5.

For this to work, each subtask needs a fixed output format, ideally a JSON schema. The sub-agent receives only the data and tools its subtask requires. Multi-step tool work calls for a medium reasoning level, pure classification for the lowest. Before anything is changed in connected systems, the planning model or a person checks the result, as described in Introducing agents with approvals.

In Hawk IVO Force, the model is a setting of the individual agent. This means an agent for planning and coordination can work with a large model, while agents for triage, extraction or monitoring use a small one.

Chinese models: inexpensive, open, with data protection questions

The Chinese Flash models are inexpensive and solve tool tasks well. GLM-5.3-Flash guesses the least often of all the models compared. DeepSeek V4.1 Flash and GLM-5.3-Flash are available as open weights under the MIT license, Qwen3.8 in several open variants. For companies in Germany, what matters most is where the data is processed:

• According to its privacy policy, DeepSeek processes and stores data from its own API in China. DeepSeek V4.1 Flash is also available in the Microsoft Azure EU Data Zone, with a surcharge of around 20% over the global price.
• Alibaba offers two scopes in Frankfurt. Only the EU scope keeps processing in the region. In Frankfurt, Qwen3.8 Flash is currently listed only in the global scope.
• According to its own information, Z.ai usually processes data in Singapore. MiniMax stores data in the US.
• Open weights can be self-hosted. With several hundred billion parameters, this is demanding for DeepSeek and GLM. The smaller Qwen3.8-27B under the Apache 2.0 license is more realistic.

The US institute NIST examined older DeepSeek and Kimi versions and found an alignment with positions of the Chinese Communist Party and a higher susceptibility to manipulation. These characteristics lie in the model itself and remain when it is self-hosted. No such review exists for the current versions. When classifying into fixed categories, this is likely to matter less than with freely worded texts. The model card of GLM-5.3-Flash does not mention German, so tests with German texts are particularly important.

Which model for which task

Prices, measurements and the hands-on test lead to this assessment:

• Email triage and classification with fixed categories: GPT-6 Luna or Claude Haiku 5.5 with a low reasoning level. Both cost equally little. GPT-6 Luna stuck closely to the source material in the hands-on test and in the grounding test, while Claude Haiku 5.5 guesses less often when it lacks knowledge.
• Checking production orders and extracting from documents: the same models, with one assessment per order or group instead of a very long prompt, with a JSON schema and null allowed.
• Website agent with real-time response: a model without or with minimal reasoning, such as GPT-6 Luna with the first token after around 0.8 seconds or Nova 2 Lite without reasoning after around 1.1 seconds. At Artificial Analysis, Gemini 3.5 Flash-Lite needed around 8 seconds to the first token.
• Continuous tasks and sub-agents with tools: Gemini 3.8 Flash, GLM-5.3-Flash or DeepSeek V4.1 Flash for multi-step business processes, Claude Haiku 5.5 for command line work and computer use, each with a medium reasoning level.
• Very large volumes of very simple texts: with a cost index of 24, Amazon Nova Micro is the cheapest model in the comparison, but it knows little and processes at most 128,000 tokens. It is only suitable with a fixed schema and after your own test.
• Personal data: choose processing in the EU, such as the EU Data Zone in Azure, the EU multi-region in Google Cloud or an EU route in Amazon Bedrock, and plan for a surcharge of 10 to around 40%.

Test with your own data before deciding

In this market, model versions and prices change every few weeks. Before making a decision, a short test with 50 to 100 of your own examples is worthwhile. It shows accuracy, format adherence and actual token consumption per item. We help you define a suitable task, compare models with your data and set up agents in Hawk IVO Force. Get in touch if you have such a task in mind.

List prices and calculation

List prices and calculation (as of October 20, 2026)

Public list prices of the providers in the standard tier with global processing, in US dollars per million tokens. These are not Hawk IVO Force prices. Prices change frequently, and the linked price pages of the providers are authoritative.

List prices in US dollars per million tokens

Model	Input	Output	Surcharge for long prompts
GPT-6 Luna	0.10	0.50	above 272,000 tokens: input ×2, output ×1.5
Claude Haiku 5.5	0.10	0.50	above 100,000 tokens: input ×5, output ×5
Gemini 3.5 Flash-Lite	0.30	2.50	none
Gemini 3.8 Flash	0.75	3.75	none, double base price from January 1, 2027
Amazon Nova 2 Lite	0.30	2.50	none known
Amazon Nova Micro	0.035	0.14	none known
DeepSeek V4.1 Flash	0.30	1.20	none
Qwen3.8 Flash	0.15	0.47	none known
GLM-5.3-Flash	0.15	0.50	none known
MiniMax M3.1 Flash	0.15	0.60	none known
Claude Sonnet 5.5 (for comparison)	2.00	10.00	none
Claude Opus 5.5 (for comparison)	4.00	20.00	none

With global processing, the models usually cost the same on Microsoft Azure, Amazon Bedrock and Google Cloud as directly from the provider. Regional processing, for example in the EU, costs 10 to around 40 percent more.

Model costs in US dollars per month for the use cases in this article, rounded, without caching, batch or data residency

Model	Office	Production	Website	Continuous task
GPT-6 Luna	0.97	32	95	6
Claude Haiku 5.5	0.97	32	95	6
Gemini 3.5 Flash-Lite	4	120	330	19
Gemini 3.8 Flash	7	240	710	42
Amazon Nova 2 Lite	4	120	330	19
Amazon Nova Micro	0.31	10	32	2
DeepSeek V4.1 Flash	3	86	270	16
Qwen3.8 Flash	1	40	130	8
GLM-5.3-Flash	1	41	130	8
MiniMax M3.1 Flash	1	43	140	8
Claude Sonnet 5.5 (for comparison)	19	630	1,890	110
Claude Opus 5.5 (for comparison)	39	1,260	3,780	220

Provider price pages

• developers.openai.com (opens in a new tab) (openai)
• platform.claude.com (opens in a new tab) (anthropic)
• ai.google.dev (opens in a new tab) (google)
• aws.amazon.com (opens in a new tab) (aws)
• api-docs.deepseek.com (opens in a new tab) (deepseek)
• www.alibabacloud.com (opens in a new tab) (alibaba)
• docs.z.ai (opens in a new tab) (zai)
• platform.minimax.io (opens in a new tab) (minimax)