Every major LLM provider now sells a small, low-cost model tier, but the names, prices, context windows and pricing rules differ so much that a direct comparison is hard to assemble. This page puts the efficient tier from nine providers side by side, using identical fields for each: API identifier, release date, context window, maximum output, input types, licence and price per million tokens. It covers hosted API models such as Claude Haiku, GPT Luna, Gemini Flash-Lite and Muse Ember, and open-weight models that can also be self-hosted, such as Gemma 4 and Mistral Small 4. It is written for developers, architects and buyers who need to cost high-volume workloads such as classification, extraction, routing and sub-agent tasks. Providers are listed alphabetically; the order is never a ranking, and the page draws no verdicts.
Last checked: 2026-11-15.
This is an independent reference page from NHI Mgmt Group based on public provider documentation; it is not sponsored or endorsed by any provider. Prices and model line-ups change often, so confirm against each provider's current pricing page before committing a budget.
Key takeaways
- Input prices span a wide band. On the pages checked, list input prices for the models below run from $0.10 per million tokens (Claude Haiku 5.5 up to 100K prompt tokens, GPT-6 Luna, Muse Ember up to 200K input tokens) to $0.30 (Gemini 3.5 Flash-Lite, DeepSeek flash at peak hours).
- Pricing rules matter as much as list price. Haiku 5.5 and Muse Ember switch the whole request to a higher rate above a prompt threshold, DeepSeek halves its rates off-peak, and Alibaba shows limited-time discounts on some Qwen models.
- Context windows are large but not uniform. Haiku 5.5, Gemini 3.5 Flash-Lite, Nova 2 Lite and DeepSeek flash list about 1M tokens; Mistral Small 4 lists 256K; Gemma 4 lists 128K or 256K by size; some providers do not state a figure.
- Several low-cost models are open weight. Gemma 4 and Mistral Small 4 are published under Apache 2.0, so they can run on your own hardware as well as through an API.
- Free tiers exist but carry conditions. Google lists free-tier access for Gemini 3.5 Flash-Lite, 3.1 Flash-Lite and Gemma 4, and Z.ai lists GLM-4.7-Flash, GLM-4.5-Flash and GLM-4.6V-Flash as free; check data-use terms before sending business data to any free tier.
Which models count as small or low-cost?
There is no industry definition of a "small" model, and most providers do not publish parameter counts for their hosted models. This page therefore uses an objective rule: a model is included when its provider positions it as its efficient, fast or low-cost tier on an official page, or when it is an open-weight model the provider describes for edge, local or efficient deployment, and its facts could be confirmed on official pages on 2026-11-15. Size in parameters is listed only where the provider states it.
Provider positioning, in each provider's own words from the pages checked:
- Alibaba Qwen, qwen3.8-flash: listed in Alibaba Cloud Model Studio pricing alongside qwen3.8-max and qwen3.7-plus as the Flash tier.
- Amazon, Nova 2 Lite: Bedrock model card lists it as active, with reasoning supported but off by default.
- Anthropic, Claude Haiku 5.5: "for high-volume, latency-sensitive tasks such as classification, extraction, and routing".
- DeepSeek, deepseek-flash: the API name for DeepSeek-V4.1-Flash-1017, one of two models on DeepSeek's pricing page.
- Google, Gemini 3.5 Flash-Lite: "a low-latency, cost-effective multimodal model" for "high-throughput, low-cost execution for subagent tasks and document parsing". Gemma 4 is Google's open-weight family for phones, browsers, local hardware and servers.
- Meta, Muse Ember: "the efficient execution model in the Muse family, built for high-volume, repeated work".
- Mistral AI, Mistral Small 4: "Our powerful hybrid model unifying instruct, reasoning, and coding capabilities in a single model."
- OpenAI, GPT-6 Luna: "Our most efficient model for focused, high-volume tasks."
- Zhipu Z.ai, GLM-5.3-Flash: the Flash tier of the GLM-5.3 family on Z.ai's pricing page, alongside GLM-5.3-FlashX.
For the general trade-offs between small and large models, see Small vs Large Language Models. This page does not repeat that material; it compares the current products.
Key facts side by side
As of 2026-11-15. Providers alphabetical. "Not stated" means the provider did not state the fact on the official pages checked on that date.
| Provider | Model | API ID | Released | Context | Max output | Input types | Open weights |
|---|---|---|---|---|---|---|---|
| Alibaba Qwen | Qwen3.8-Flash | qwen3.8-flash | Not stated by Alibaba on the pages checked | Pricing tier covers up to 1M input tokens per request | Not stated by Alibaba | Not stated by Alibaba | Not stated by Alibaba for the API model |
| Amazon | Nova 2 Lite | amazon.nova-2-lite-v1:0 | 2 December 2025 | 1M | 64K | Text, image, video | No |
| Anthropic | Claude Haiku 5.5 | claude-haiku-5-5 | 7 October 2026 | 1M | 128K | Text, image | No |
| DeepSeek | DeepSeek-V4.1-Flash-1017 | deepseek-flash | Not stated by DeepSeek on the pricing page | 1M | Up to 384K | Not stated by DeepSeek on the pricing page | Not stated by DeepSeek on the pricing page |
| Gemini 3.5 Flash-Lite | gemini-3.5-flash-lite | Latest update July 2026 (model page) | 1,048,576 | 65,536 | Text, image, video, audio, PDF | No | |
| Gemma 4 (E2B, E4B, 12B, 26B A4B, 31B) | Downloadable weights | Not stated on the Gemma overview page | 128K (small sizes), 256K (medium sizes) | Not stated by Google | Text, image, video; audio on E2B, E4B and 12B | Yes, Apache 2.0 | |
| Meta | Muse Ember | muse-ember-1.0 | Not stated on the pages checked | Not stated on the pages checked | Not stated by Meta | Text, image, video, audio, PDF | Not stated by Meta |
| Mistral AI | Mistral Small 4 | mistral-small-2603 | 16 March 2026 | 256K | Not stated by Mistral | Not stated on the model card | Yes, Apache 2.0 (119B total, 6.5B active) |
| OpenAI | GPT-6 Luna | gpt-6-luna | 22 September 2026 | 1.05M | 128K | Not stated on the models page | No |
| Zhipu Z.ai | GLM-5.3-Flash | GLM-5.3-Flash | Not stated by Z.ai on the pricing page | Not stated by Z.ai on the pricing page | Not stated by Z.ai | Not stated on the pricing page | Not stated on the pricing page |
Knowledge cutoffs stated on the pages checked: Claude Haiku 5.5, June 2026 (reliable knowledge cutoff); GPT-6 Luna, 18 May 2026; Nova 2 Lite, October 2025. The other providers did not state a cutoff for these models on the pages checked. For what a cutoff means in practice, see LLM Knowledge Cutoff Explained.
How much do small and low-cost LLMs cost?
As of 2026-11-15. USD per million tokens, standard (non-batch) list prices from each provider's official pricing page. Providers alphabetical.
| Provider | Model | Input | Cached input | Output | Batch or discount rule |
|---|---|---|---|---|---|
| Alibaba Qwen | qwen3.8-flash | $0.15 | Not stated in the section read | $0.47 | International (Singapore) list price, tier 0 to 1M input tokens |
| Amazon | Nova 2 Lite | Not stated on the pages checked | Not stated on the pages checked | Not stated on the pages checked | Model card refers to the Amazon Bedrock pricing page; Standard, Priority and Flex tiers supported |
| Anthropic | Claude Haiku 5.5 (up to 100K prompt tokens) | $0.10 | $0.01 | $0.50 | Batch $0.05 input, $0.25 output |
| Anthropic | Claude Haiku 5.5 (over 100K prompt tokens) | $0.50 | $0.05 | $2.50 | Batch $0.25 input, $1.25 output |
| DeepSeek | deepseek-flash (peak) | $0.30 | $0.006 | $1.20 | Peak 01:00 to 04:00 and 06:00 to 10:00 UTC, Monday to Friday |
| DeepSeek | deepseek-flash (off-peak) | $0.15 | $0.003 | $0.60 | All other times, including weekends and Chinese public holidays |
| Gemini 3.5 Flash-Lite | $0.30 | $0.03 plus $1.00 per 1M tokens per hour storage | $2.50 | Batch $0.15 input, $1.25 output; free tier available | |
| Gemini 3.1 Flash-Lite | $0.25 (audio $0.50) | $0.025 (audio $0.05) | $1.50 | Batch $0.125 input, $0.75 output; free tier available | |
| Gemma 4 | No paid tier listed | No paid tier listed | No paid tier listed | Free of charge on the Gemini API free tier; self-hosting costs are your own | |
| Meta | Muse Ember (up to 200K input tokens) | $0.10 | $0.004 | $0.50 | Standard mode only |
| Meta | Muse Ember (over 200K input tokens) | $0.30 | $0.012 | $1.50 | Whole request billed at the higher rate |
| Mistral AI | Mistral Small 4 | See rule | Not stated on the model card | See rule | The model card lists $0.15 and $0.60 per million tokens without labelling which is input and which is output |
| OpenAI | GPT-6 Luna | $0.10 | $0.01 | $0.50 | Standard tier price |
| Zhipu Z.ai | GLM-5.3-Flash | $0.15 | $0.03 | $0.50 | GLM-5.3-FlashX is $0.37, $0.075 and $1.25 |
Free models listed by Z.ai on the same page: GLM-4.7-Flash, GLM-4.5-Flash and GLM-4.6V-Flash (vision) are free for input, cached input and output. For how tokens, caching and batch discounts work, see LLM Pricing Explained, and for a wider dated price table see the LLM Pricing Tracker.
Pricing rules that change the real cost
List prices are only comparable once the billing rules are applied to your own traffic. The rules below come from the same official pages.
| Rule | Where it applies | Effect on a bill |
|---|---|---|
| Whole-request prompt threshold | Claude Haiku 5.5 (100,000 prompt tokens); Muse Ember (200K input tokens) | A request above the threshold pays the higher rate on all its tokens, not only the part above the threshold. Anthropic counts cache reads and writes toward the prompt length. |
| Peak and off-peak hours | DeepSeek | Off-peak rates are half the peak rates; scheduling batch work off-peak halves token cost. |
| Batch discount | Anthropic (50% on input and output); Google (batch prices listed per model) | Asynchronous jobs cost about half the standard rate. |
| Cache storage charge | Google Gemini context caching | An hourly storage fee per million cached tokens is added to the cached input rate. |
| Data residency multiplier | Anthropic inference_geo "us" | A 1.1x multiplier on all token categories, including Haiku 5.5's over-100K prices. |
| Limited-time discounts | Alibaba (some Qwen models labelled "Limited-time 20% off") | Budget on the undiscounted rate if the workload outlives the offer. |
| Free tiers | Google (Flash-Lite, Gemma 4); Z.ai (older Flash models) | No token charge, but rate limits and data-use terms differ from paid tiers. |
Worked example: a job of 1 million requests, each with 2,000 input tokens and 300 output tokens, uses 2,000M input and 300M output tokens. At $0.10 input and $0.50 output that is $200 plus $150, or $350, before caching or batch discounts. At $0.30 and $2.50 it is $600 plus $750, or $1,350. The same arithmetic applied to your own prompt and output lengths is the only reliable way to compare these models, because output-heavy workloads reward a low output price and long-prompt workloads hit the threshold rules above.
What is the context window of each small model?
Most of the hosted models in this comparison list a context window of about one million tokens, which is no longer a feature reserved for flagship models. The exceptions are the open-weight models and the providers that do not state a figure.
- About 1M tokens: Claude Haiku 5.5 (1M), DeepSeek flash (1M), Gemini 3.5 Flash-Lite (1,048,576), GPT-6 Luna (1.05M), Nova 2 Lite (1M).
- 256K tokens: Mistral Small 4; Gemma 4 medium sizes (26B A4B and 31B).
- 128K tokens: Gemma 4 small sizes (E2B and E4B).
- Not stated on the pages checked: GLM-5.3-Flash, Muse Ember and Qwen3.8-Flash (Alibaba's pricing tier extends to 1M input tokens per request, which is not the same as a stated window).
Maximum output is a separate limit. GPT-6 Luna and Claude Haiku 5.5 list 128K, DeepSeek flash up to 384K, Gemini 3.5 Flash-Lite 65,536 and Nova 2 Lite 64K. A large window does not mean every token is used equally well, and long prompts may trigger the threshold pricing above. See LLM Context Windows Explained and LLM Context Windows Compared Across Providers.
Which small models are open weight?
Two of the models in this comparison are confirmed open weight on the official pages checked.
- Gemma 4 (Google): five sizes, E2B and E4B ("effective" 2B and 4B models for mobile, edge and browser use), 12B, 31B (dense) and 26B A4B (a mixture-of-experts model that activates about 4B parameters per token but loads all 26B). Google's Gemma page states the weights are open and permit responsible commercial use, and links a Gemma 4 licence that points to Apache 2.0. Every size includes a multi-token prediction draft model for faster inference.
- Mistral Small 4 (Mistral AI): 119B total parameters with 6.5B active, Apache 2.0, also available on Mistral's own API as mistral-small-2603.
Open weights change the cost model: there is no per-token fee when self-hosted, but you pay for hardware, operations and security. The hosted models from Amazon, Anthropic, Google (Gemini) and OpenAI publish no weights. For DeepSeek flash, GLM-5.3-Flash, Muse Ember and Qwen3.8-Flash, the pricing and model pages checked do not state whether the API model's weights are published; see each provider's hub page. See Open-Weight vs Closed LLMs Explained, Running LLMs Locally Explained and the Self-Hosted LLM Security Guide.
Where are these models available?
As of 2026-11-15, from the official pages checked. Providers alphabetical.
| Provider | Model | Availability stated by provider |
|---|---|---|
| Alibaba Qwen | qwen3.8-flash | Alibaba Cloud Model Studio; international (Singapore) and mainland China price lists are separate |
| Amazon | Nova 2 Lite | Amazon Bedrock, Converse and Invoke APIs; geo inference profiles for the US, EU and Japan plus global cross-region inference; in-Region inference not supported in any listed region |
| Anthropic | Claude Haiku 5.5 | Claude API, Amazon Bedrock, Claude Platform on AWS, Google Cloud and Microsoft Foundry, per Anthropic's release notes |
| DeepSeek | deepseek-flash | DeepSeek API |
| Gemini 3.5 Flash-Lite; Gemma 4 | Gemini API and Google AI Studio (paid, batch and free tiers); Gemma 4 weights for download and on the Gemini API free tier | |
| Meta | Muse Ember | Meta Model API, Standard mode only; Meta states availability can vary by country |
| Mistral AI | Mistral Small 4 | Mistral API (mistral-small-2603) and open weights |
| OpenAI | GPT-6 Luna | OpenAI API |
| Zhipu Z.ai | GLM-5.3-Flash | Z.ai API |
Benchmarks
None of the official model, pricing or overview pages checked for this comparison publishes a benchmark score for these small models, and this page does not repeat scores from secondary sources. Provider announcement posts and model cards sometimes include provider-reported figures; treat those as the provider's own testing, check the benchmark version and settings, and keep them separate from independent leaderboards such as LMArena or Artificial Analysis, which must be read with their as-of date. See How to Read LLM Benchmarks.
Data handling and security
Low-cost models are often used in high-volume pipelines that process large amounts of business data, so data terms matter as much as price.
- Free tiers: read the provider's terms for free-tier traffic before sending personal or confidential data; free and paid tiers can have different data-use rules.
- Residency: Amazon lists geographic inference profiles for Nova 2 Lite; Anthropic offers a US inference option at a 1.1x multiplier; DeepSeek and Alibaba publish region-specific price lists. Confirm the processing location against your obligations.
- Self-hosting: open-weight models keep data in your environment but move model security, patching and access control to you.
- Keys and spend: cheap tokens make abuse of a stolen key less visible, not less costly at scale. Store keys in a secrets manager, scope them and set spend limits.
See AI Data Retention and Deletion, the LLMjacking Guide and How to set spend limits on LLM API keys.
How to choose a small model: factual criteria
- Token mix: calculate cost from your own input and output lengths; output prices differ more than input prices across this group (from $0.47 to $2.50 per million on the pages checked).
- Prompt length: if prompts often exceed 100K or 200K tokens, model the threshold pricing on Haiku 5.5 and Muse Ember.
- Timing: work that can run off-peak or in batch can halve cost on DeepSeek, Anthropic and Google.
- Input types: Gemini 3.5 Flash-Lite and Muse Ember list audio, video and PDF input; Nova 2 Lite lists text, image and video; Haiku 5.5 lists text and image.
- Deployment: choose Gemma 4 or Mistral Small 4 if weights must run on your own hardware or offline.
- Platform: confirm the model is offered on the cloud and region you already use.
- Lifecycle: check deprecation dates; Amazon states Nova 2 Lite end of life no sooner than 2 December 2026 with a legacy period of at least six months.
Practitioner checklist
- Pin each workload to an explicit API ID and record provider, model and price date in your AI inventory.
- Build a cost model from real prompt and output lengths, including threshold, peak and batch rules.
- Re-check prices monthly; several providers on this page show scheduled or limited-time changes.
- Run the same evaluation set on two or three candidates before switching a production pipeline.
- Confirm data-use terms and processing region before sending business data, especially on free tiers.
- Set spend limits and alerts on every API key used for high-volume jobs.
- For open-weight models, track the licence version and apply the same access controls as any production service.
Standards and references
- Anthropic: Pricing (accessed 2026-11-15)
- Anthropic: Models overview (accessed 2026-11-15)
- Anthropic: Claude Platform release notes (accessed 2026-11-15)
- Alibaba Cloud: Model Studio model pricing (accessed 2026-11-15)
- Amazon Web Services: Amazon Nova 2 Lite model card (accessed 2026-11-15)
- DeepSeek: Models and pricing (accessed 2026-11-15)
- Google: Gemini API pricing (accessed 2026-11-15)
- Google: Gemini 3.5 Flash-Lite model page (accessed 2026-11-15)
- Google: Gemma 4 model overview (accessed 2026-11-15)
- Meta: Muse models (accessed 2026-11-15)
- Meta: Model API pricing and rate limits (accessed 2026-11-15)
- Mistral AI: Mistral Small 4 model card (accessed 2026-11-15)
- OpenAI: Models (accessed 2026-11-15)
- OpenAI: Pricing (accessed 2026-11-15)
- OpenAI: API changelog (accessed 2026-11-15)
- Z.ai: Pricing (accessed 2026-11-15)
Related NHI Mgmt Group resources: Small vs Large Language Models, LLM Pricing Tracker, Coding LLMs Compared, Frontier AI Models Compared, LLM Security Guide.