The map below sorts the AI models that matter as of October 2026 along two axes. The first is size and capability. Frontier models sit at the top of the leaderboards. Mid-tier models do most everyday work at a fraction of the price. Small models are cheap enough to run on every request or on a laptop. The second axis is how you pay. A closed model comes through an API billed per million tokens, or through a subscription in a chat app. You download open weights and run them yourself. An on-device model ships inside a phone or operating system and has no token bill at all.

Every price here was read on 30 October 2026. It comes from the vendor's own page where that page could be fetched, and otherwise from the third-party source named next to it. Several prices are promotional and have an end date, which this guide notes where it applies. The companion pricing snapshot tracks the same numbers row by row, with a source link on each.

How to read the map

Parameter counts are public only for open models. Every closed vendor (OpenAI, Anthropic, Google, xAI, Amazon, Meta's Muse line) keeps its sizes undisclosed. For closed models, the columns therefore follow the vendor's own tier names and the price. Open models publish their counts, and most large ones are mixture-of-experts. DeepSeek-V4-Pro, for example, has 1.6 trillion parameters in total but only 49 billion active per token, according to DataCamp's write-up of the release. Total size tells you how much memory the weights need. Active size tells you roughly how fast and how cheap the model is to serve.

Most vendors sell three sizes per family. Anthropic has done it since Claude 3 Opus, Sonnet and Haiku in March 2024. OpenAI launched gpt-5, gpt-5-mini and gpt-5-nano together in August 2025. The current lines follow the same shape: GPT-6 Astra, GPT-6.1 Sol and Terra, GPT-6 Luna at OpenAI; Opus, Sonnet and Haiku at Anthropic; Argon, Flash and Flash-Lite at Google. Each size gets its own column on the map. The input price gap between neighbouring columns runs from under 2x (Grok 4.7 to Grok 4.3) to 20x (Sonnet 5.5 to Haiku 5.5, GPT-6.1 Sol to GPT-6 Luna). Most of the savings in an LLM budget come from that gap.

OpenAI's August 2025 API launch put three sizes in one line, a pattern most major vendors now follow.

The most-cited independent ranking for the frontier column is the Artificial Analysis Intelligence Index. The snapshot below (index v4.3.2, read 17 October 2026) comes via DevelopersDigest and Javadex, which both republish the board. Treat the order as approximate.

Model Index score Weights
Claude Opus 5.5 (max effort) 58 closed
Claude Sonnet 5.5 56 closed
Muse Wonder Full (Meta) 56 closed
Claude Fable 5.1 53 closed
GPT-6 Astra 53 closed
Gemini 4 Argon (high) 53 closed
Muse Wonder 52 closed
GPT-6.1 Sol 52 closed
MiMo-V2.6 Pro (Xiaomi) 46 open, MIT
Grok 4.7 46 closed
GLM-5.3 (Zhipu) 45 open
Kimi K3 (Moonshot) 44 open, modified MIT
DeepSeek V4.1 Flash 39 open, MIT
Qwen3.8 27B 34 open, Apache 2.0

Two things stand out. First, the best open model (46) ties Grok 4.7 and sits 12 points below the top closed model. Second, a closed flagship and the closed mid-tier model from the same vendor are only one or two points apart on this index. GPT-6 Astra scores 53 and GPT-6.1 Sol 52 at a fifth of the price. Claude Opus 5.5 scores 58 and Sonnet 5.5 56 at half the price.

The map

AI model map, October 2026: capability tier across, access model down Frontier top of the leaderboards Mid most everyday work Small cheap, fast, local Closed API pay per token GPT-6 Astra Claude Fable 5.1 Claude Opus 5.5 Gemini 4 Argon Muse Wonder (Meta) Grok 4.7 GPT-6.1 Sol GPT-6.1 Terra Claude Sonnet 5.5 Gemini 3.8 Flash Grok 4.3 Nova 2 Pro (Amazon) GPT-6 Luna Claude Haiku 5.5 Gemini 3.5 Flash-Lite Gemini 3.1 Flash-Lite Nova 2 Lite (Amazon) Open weights run or rent DeepSeek V4-Pro Kimi K3 (Moonshot) GLM-5.2 (Zhipu) MiMo-V2.6 Pro (Xiaomi) Llama 5 Mistral Large 4 Qwen3.8 2.4T DeepSeek V4.1-Flash gpt-oss-120b Qwen3.8-27B Gemma 4 31B Command A+ (Cohere) Nemotron 3 Super MiMo-V2.6 Flash gpt-oss-20b Gemma 4 E2B / E4B Phi-4 / Phi-4-mini Ministral 3 (3B to 14B) Qwen3 0.6B to 8B Nemotron 3 Nano Llama 3.2 1B / 3B On-device no token bill none V4-Pro weights alone are about 865 GB gpt-oss-20b (16 GB) AFM 3 Core Advanced Llama 3.1 8B (4.7 GB) Mistral 7B (4.1 GB) Granite 3.2 8B (4.9 GB) laptop, 8 to 16 GB RAM Gemma 4 E2B / E4B Gemini Nano 4 AFM 3 Core (Apple, 3B) Phi-4-mini (2.5 GB) Llama 3.2 3B (2.0 GB) phone, Raspberry Pi
Capability tiers across, ways to get the model down, October 2026. Bracketed sizes are Q4 downloads from the Ollama library, except gpt-oss-20b, where 16 GB is the memory the AWS model card says it fits in.

Each row has a different bill. A closed API model is priced per million tokens, and the same model usually sits behind a consumer subscription too. Open weights are free to download under a stated licence. Apache 2.0 and MIT are the common ones; Meta, Qwen's largest model and Moonshot add custom clauses. You then rent a GPU, pay a host such as Groq, Together or OpenRouter per token, or run the weights on your own hardware. The on-device row is a subset of the open row, plus closed models that ship inside a device (Apple's AFM 3, Google's Gemini Nano 4). It covers anything that fits in the RAM of a laptop or a phone and therefore costs nothing per call.

Vendor by vendor

OpenAI

The flagship is GPT-6 Astra, at $10 per million input tokens and $50 per million output, with cached input at $1. The context window is 1.05M tokens, with 922K of input and 128K of output. Any request with more than 272K input tokens is billed at 2x on input and 1.5x on output for the whole request. Fast mode doubles the price. Launch coverage by DataNorth puts the release on 3 September 2026 and quotes OpenAI's pitch that Astra "operates software through screens and controls".

The mid tier has two models. GPT-6.1 Sol costs $2 / $10, and OpenAI's model page describes it as "balanced speed, cost, and intelligence". GPT-6.1 Terra costs $1 / $5. The small model is GPT-6 Luna at $0.10 / $0.50, which the same page calls the "fastest and most cost-effective". Previous-generation models stay on sale. GPT-5.5 costs $5 / $30, and GPT-5.6 Sol is at a promotional $4 / $20 (standard $5 / $30) until at least 21 November 2026. The older o3 costs $2 / $8 and o4-mini $1.10 / $4.40. Batch processing is half price across the line.

OpenAI also has an open pair, gpt-oss-120b and gpt-oss-20b, released in August 2025 under Apache 2.0 with a 128K context. The AWS model card says the 20B model fits in 16 GB of memory. TestingCatalog's launch report says the 120B model runs on a single 80 GB GPU. Hosted, gpt-oss-120b costs $0.15 / $0.60 on Groq and from $0.03 / $0.15 on OpenRouter, which also lists a free variant.

Anthropic

Anthropic sells closed models only. All of them have a 1M-token context and 128K output, and all except Haiku 5.5 charge a flat price across that context. Claude Fable 5.1 is the top model at $10 / $50, with cache reads at $0.25. Claude Opus 5.5 is the flagship for most work at $4 / $20 (cache read $0.20, Fast mode $8 / $40). Claude Sonnet 5.5 is the mid tier at $2 / $10. Claude Haiku 5.5 is the small model at $0.10 / $0.50 for prompts up to 100K tokens, rising to $0.50 / $2.50 above that. Message Batches cut input and output prices by 50% on every active model.

Launch coverage positions each model differently. BRIDGE and eesel describe Fable 5.1 as the "world's most advanced model for coding and knowledge work". TechNews calls Opus 5.5 "Fable 5.1-level on most tasks" and up to 40% cheaper than Opus 5. AF.NET says Haiku 5.5 targets high-volume summarisation, classification and browser use. All four shipped between 1 September and 7 October 2026, according to Hidekazu Konishi's compiled timeline of Claude releases.

Consumer plans on claude.com/pricing start with Free and Pro at $20 a month ($17 on annual billing). Max costs $100, and Team $25 or $125 per seat. Enterprise is $20 per seat on annual billing, with a 20-seat minimum.

Google

Gemini 4 Argon is the flagship, at $2 / $10 until 30 November 2026 and $4 / $20 from 1 December. Google's models page lists a combined input-plus-output context of 1,048,576 tokens. The same page calls Argon the "flagship reasoning model" for software engineering, maths and science, and multi-step agents. According to Reuters via Khaleej Times, general API and app access opened on 13 October 2026.

Gemini 3.8 Flash is the mid tier at $0.75 / $3.75 until 31 December 2026, then $1.50 / $7.50. Google calls it the "most capable Flash model for long-horizon coding, agents, enterprise workflows". The small models are Gemini 3.5 Flash-Lite at $0.30 / $2.50 ("fastest, most cost-effective 3.5 model") and Gemini 3.1 Flash-Lite at $0.25 / $1.50. Batch is 50% off, the Flex service tier matches the batch price, and Priority costs 1.8x. The Flash and Flash-Lite models have a free API tier; Argon and 3.1 Pro Preview do not. The pricing page states that free-tier content is "used to improve products".

Google's open line is Gemma 4, released on 2 April 2026 under Apache 2.0, according to The Next Web and DataNorth. It came in E2B (2.3B effective), E4B (4.5B effective), a 26B mixture-of-experts with 3.8B active, and a 31B dense model. A 12B followed in June. The same reports put the 31B third among open models on the Arena text board. Consumer plans, according to Engadget and TechCabal, are AI Plus at $7.99, AI Pro at $19.99, AI Ultra at $99.99 (cut from $250) and a $199.99 top tier.

Meta

Meta now runs two tracks, according to StartupHub. Llama 5, released on 8 April 2026, is open weights under a community licence, with a context of up to 5M tokens. Muse is the closed line, and Muse Wonder came out on 12 October 2026. According to YottaLabs, Muse Wonder costs $4 / $20 (cached $0.20) through the Meta Model API, while Muse Wonder Full is subscription only. The same source lists Muse Code plans at $5, $15, $50 and $200 a month. Older Llamas are cheap to rent. Llama 3.3 70B is $0.10 / $0.32 on OpenRouter, and Llama 3.1 8B is $0.045 / $0.384 on Cloudflare Workers AI in its fp8-fast variant.

DeepSeek

DeepSeek-V4-Pro and DeepSeek-V4.1-Flash are open under the MIT licence. According to DataCamp and Kanerika, V4-Pro is a 1.6T-parameter mixture-of-experts with 49B active and weights of roughly 865 GB; V4.1-Flash is 284B with 13B active. Both have a 1M context and 384K maximum output.

The API price depends on the time of day. V4-Pro is $1.32 / $3.96 at peak and $0.66 / $1.98 off-peak; V4.1-Flash is $0.30 / $1.20 at peak and $0.15 / $0.60 off-peak. Cache hits cost $0.044 (Pro) and $0.006 (Flash) at peak. Peak hours are 01:00 to 04:00 and 06:00 to 10:00 UTC on weekdays. There is no batch API; off-peak pricing serves as the discount. Thinking mode does not change the per-token price. Third parties host the same weights: Together at $1.32 / $3.96 and $0.30 / $1.20, Mistral's platform at $0.30 / $1.20 for Flash.

Qwen (Alibaba)

According to nxcode and LetsDataScience, Alibaba released Qwen3.5-397B-A17B in February 2026 under Apache 2.0. Two models followed in August 2026: a 27B dense Qwen3.8 under Apache 2.0 and a 2.4T-parameter Qwen3.8 with 95B active under a custom Qwen licence. The hosted Qwen3.8-Max is API only. Alibaba claims the 27B "matches a model ten times larger". The native context is 262K, extendable to 1M. Alibaba's own Qwen3.5-Plus API costs $0.40 / $2.40 up to 256K tokens. Qwen3.8-27B is the most widely hosted open model in our data; OpenRouter lists 21 providers, from $0.035 / $1.50. The small Qwen3 models (0.6B, 1.7B, 4B, 8B) are Apache 2.0.

Mistral

Mistral Large 4 went public on 27 October 2026 as an open-weight model, according to TechCrunch and Reuters. The coverage did not state its licence name or size. Mistral's pricing docs list it at $0.70 / $2.10, with cached input at $0.07. The marketing pricing page, fetched the same day, says $0.50 / $1.50, so check before quoting. Mistral Small 4 is $0.15 / $0.60. Ministral 3 at 14B, 8B and 3B costs $0.20, $0.15 and $0.10 flat for input and output. Codestral is $0.30 / $0.90. Batch halves all of them.

According to eWeek, Mistral Large 3 (December 2025) is Apache 2.0, at 675B total with 41B active. Ministral 3 is also Apache 2.0 and aims at on-device use, data privacy and intermittent connectivity. The platform's free mode allows 1 request per second, 500,000 tokens per minute and 1 billion tokens a month. Le Chat plans, according to CostBench, are Free, Pro at $14.99 and Team at $24.99 per user.

xAI

Grok is closed. The current flagship is Grok 4.7, at $2 / $6 with cached input at $0.50 and a 500K context; prompts above 200K tokens pay double. Grok 4.3 is the mid tier at $1.25 / $2.50 with a 1M context. Batch takes 20% off. According to ai-toolbox's model guide, Grok 4.7 shipped on 21 September 2026 and xAI recommends it "for code and general work". The same guide lists X plans at $8 (Plus), $30 (Premium), $100 (Super) and $200 (Ultra) a month.

Microsoft Phi

Phi is Microsoft's small open family, MIT licensed. According to frankx's analysis, the current set is Phi-4 (14B), Phi-4-mini (3.8B, 128K context, function calling), Phi-4-multimodal (5.6B, speech and audio), Phi-4-reasoning (14B) and Phi-4-reasoning-vision (15B, March 2026). There was no Phi-5 as of June 2026. Phi-4-mini is a 2.5 GB download at Q4 in the Ollama library. Microsoft hosts the models on Azure; no per-token price appeared in our data.

Amazon

Amazon Nova is closed and sold only through Bedrock. Issoh read the AWS Price List API on 18 October 2026. By that reading, Nova 2 Lite costs $0.30 / $2.50, Nova 2 Pro $1.25 / $10 and Nova 2 Omni $0.30 / $2.50, all with a 1M context. Amazon positions Lite for general text and long context, and Pro for maths and strict instruction following. Nova 2.5 Sonic arrived on 5 October 2026. AWS's pricing page confirms batch inference on Amazon models at 50% off on-demand.

Cohere

According to Pulse, Command A+ is Cohere's first fully Apache 2.0 model, released on 20 May 2026. It has 218B total parameters with 25B active and supports citations natively. It runs in 4-bit on one B200 or two H100 GPUs at 375 tokens a second. Cohere's pricing page shows hourly Model Vault rates rather than a per-token price. Trial API keys are free but not permitted for production use.

NVIDIA

Nemotron 3 comes in Nano (30B total, 3B active, December 2025), Super (100B, 10B active) and Ultra (500B, 50B active), according to The New Stack and AI Weekly. All three are open weight with a 1M context; neither report named the licence. NVIDIA claims Nano delivers 4x the throughput of Nemotron 2 Nano with up to 60% fewer reasoning tokens. Weights are on Hugging Face, Bedrock and Google Cloud.

Moonshot

According to Forkast, Kimi K3 launched on 16 July 2026, and its weights followed on 27 July under a modified MIT licence. A branding clause applies above 100M monthly users or $20M monthly revenue. The model is a 2.8T mixture-of-experts with a 1M context, priced at $3 / $15 and $0.30 cached. Kimi K3.1 Pro reached the API on 23 October 2026 at $2 / $10, with weights due 4 November, according to LetsDataScience. The same source lists coding, agents, 3D and financial research as its targets. The API is OpenAI-compatible.

Zhipu (Z.ai)

According to AI Weekly, GLM-5.2 came out on 13 June 2026, with an MIT licence promised at launch. It has about 744B total parameters with roughly 40B active, a 1M context and 131K output, and targets coding and long-horizon agents. Zhipu sells it through a Coding Plan starting around $18 a month rather than at a published per-token rate. GLM 5.3 appears on the Artificial Analysis open-weight board at 45. Mistral's platform hosts it at $1.40 / $4.40, and GLM 5.3 Flash at $0.15 / $0.50.

MiniMax

According to Artificial Analysis, MiniMax-M3, from June 2026, is the first multimodal model in the M series. It has a 1M context and costs $0.30 / $1.20 up to 512K tokens and $0.60 / $2.40 beyond. Weights were announced as coming, with the licence unstated.

Xiaomi

Xiaomi entered the open-weight board at the top. According to eesel, MiMo-V2.6 came out on 22 September 2026 under MIT, with a 1M context in every version. Pro has 1.02T total and 42B active; Flash has 309B and 15B active; the third is a 9B dense distillation of Qwen3.5-9B. Its Index score of 46 made it the top open-weight model in October 2026. OpenRouter lists Flash at $0.14 / $0.28 and Pro at $0.435 / $0.87.

Apple

Apple's models are closed but, uniquely, free at the point of use. A developer write-up of WWDC 2026 lists the third-generation Apple Foundation Models. AFM 3 Core (3B dense) and AFM 3 Core Advanced (20B sparse with 1 to 4B active) run on device. AFM 3 Cloud and Cloud Pro run on Private Cloud Compute. Developers call them through the Foundation Models framework, with no API key and no token cost. Supported hardware is iPhone 15 Pro and later, M1 and later iPads and Macs, and Vision Pro. At launch, the on-device models were not available in the EU on iPhone or iPad. Apple publishes only human-preference comparisons against its 2025 models. See Apple Intelligence and the on-device cost angle for what this does to a mobile AI bill.

Free, paid, API

You can pay for the same model in three ways: a consumer subscription (flat monthly fee, usage caps), the API (per token, no cap), or nothing (free tiers and open weights). Which is cheaper depends on volume. API calls vs subscription works through the break-even.

Consumer plans, USD per month

Product Free Entry Standard Top tiers Source
ChatGPT $0 Go $5.50 (eesel says $8) Plus $20 Pro $100 or $200 (the $200 tier paused for new sign-ups since Sep 2026); Business $25/user ($20 annual) Krater, eesel; OpenAI help centre confirms Pro at $100 and $200
Claude $0 none Pro $20 ($17 annual) Max $100; Team $25 or $125/seat; Enterprise $20/seat annual, 20-seat minimum claude.com/pricing
Gemini $0 (32K context, per TechCabal) AI Plus $7.99 (ai-toolbox lists $4.99 as of Oct 2026) AI Pro $19.99 AI Ultra $99.99; $199.99 top tier Engadget, TechCabal, ai-toolbox
Le Chat (Mistral) $0 none Pro $14.99 Team $24.99/user CostBench
Grok (X plans) not listed Plus $8 Premium $30 Super $100; Ultra $200 ai-toolbox
Perplexity $0 none Pro $20 Max $200; Enterprise Pro $40/seat; Enterprise Max $325/seat Finout
Muse Code (Meta) not listed $5 $15 $50; $200 YottaLabs

Only the Claude row and the ChatGPT Pro price come from the vendor's own page. The rest are third-party readings, so check them again before you budget on them.

Free API tiers

Provider Free API Limits on the fetched page
Google Gemini API Yes, for Flash and Flash-Lite models and 2.5 Pro Per-model limits shown only in AI Studio; no Argon or 3.1 Pro; free-tier content used to improve products
Mistral Yes, on by default 1 request/s, 500K tokens/min, 1B tokens/month per workspace
Groq Yes gpt-oss-120b and qwen3.8-27b: 30 RPM, 1K requests/day, 8K TPM, 200K tokens/day (the page does not separate Free and Developer plan limits)
OpenRouter :free models Yes 20 RPM; 50 requests/day, or 1,000/day after buying 10 credits
Cloudflare Workers AI Yes 10,000 Neurons per day
GitHub Models Yes (preview) Low tier: 15 RPM, 150 requests/day, 8K in / 4K out tokens
Cerebras Free trial gpt-oss-120b, qwen-3.8-27b: 5 RPM, 1M tokens/day
OpenAI Free usage tier $100 monthly limit in allowed geographies; no RPM printed
Anthropic, DeepSeek No free tier on the fetched pages none

The limits, exclusions and training-data terms are covered in Free LLM APIs in 2026.

Large and small

A bigger model knows more, stays reliable over long multi-step tasks and follows a complicated instruction without being told twice. It also needs more memory, answers more slowly and costs more per token. An LLM budget comes down to finding the smallest model that still passes your evaluation.

The published numbers show where the lines fall. The NVIDIA Research and Georgia Tech position paper Small Language Models are the Future of Agentic AI gives a working definition. A small model fits on a common consumer device and responds with low enough latency for a single user, "as of 2025, most models under 10bn parameters". The paper argues that most agentic work should run on such models, calling a large model only when general conversation is needed. Two vendor claims point the same way. Alibaba says a 27B dense model "matches a model ten times larger". Mistral says Ministral 3 models "match or exceed comparable models while producing far fewer tokens". Both suggest that small models are closing the gap on much larger ones.

On the leaderboard above, scores fall steeply from the best closed model (58) to the best open model (46) to a 27B open model (34). That fall is the trade. Small language models: when a small model is enough covers which jobs survive it.

What runs on a laptop

The Ollama README gives the rule of thumb: at least 8 GB of RAM for 7B models, 16 GB for 13B models and 32 GB for 33B models. The download sizes below are the Q4 quantised files from the same library.

Q4 download size in GB for six small models, from the Ollama library Download size at Q4, GB (Ollama library) Llama 3.2 3B 2.0 GB Phi-4-mini 3.8B 2.5 GB Gemma 3 4B 3.3 GB Mistral 7B 4.1 GB Llama 3.1 8B 4.7 GB Granite 3.2 8B 4.9 GB Rule of thumb: 8 GB RAM for 7B models, 16 GB for 13B, 32 GB for 33B.
Six small models and their Q4 download sizes; all six are 8B parameters or smaller.

Below the laptop, the data gets thinner but stays useful. DataNorth's Gemma 4 report says the E2B model decodes at 7.6 tokens a second on a Raspberry Pi 5, in under 1.5 GB at 2-bit. E4B needs 12 to 16 GB and runs well on Apple Silicon. Google says the edge family is up to 4x faster than the previous Gemma and uses up to 60% less battery. Above the laptop, the AWS model card puts gpt-oss-20b in 16 GB. Past that you need one or more data-centre GPUs, which is where the 120B-class and MoE models live. The sources checked had no documented tokens-per-second benchmark for M-series Macs or consumer GPUs, so test on your own hardware before planning around it. Local LLM vs cloud API compares the cost of running these yourself with renting them.

How to choose

The video below is the best general-audience walkthrough of the model ecosystem and of switching between models by task. It dates from February 2025, so the model names are a generation old, but the method still applies.

Andrej Karpathy, "How I use LLMs" (Feb 2025), a practical tour of the ecosystem and of picking a model per task.

A short list by task follows. Each entry names two or three models from the map with their price tier. Closed API prices are in USD per million input / output tokens unless marked.

Chat and general assistant work

For one person, a consumer subscription is usually cheapest. ChatGPT Plus, Claude Pro and Gemini AI Pro all cost about $20 a month. Through the API, the mid tier is the default. That means Claude Sonnet 5.5 or GPT-6.1 Sol at $2 / $10, or Gemini 3.8 Flash at $0.75 / $3.75 until the end of the year. The free ChatGPT vs Claude vs Gemini lesson in Divero School compares the three apps.

Coding and agents

Claude Opus 5.5 at $4 / $20 holds the top index score. Gemini 4 Argon costs $2 / $10 until 30 November. GPT-6 Astra is the expensive option at $10 / $50. Two open models aim squarely at coding: Kimi K3.1 Pro at $2 / $10 (weights due 4 November) and GLM-5.2 through its coding plan. Qwen3.8-27B is the cheap self-hosted choice.

Extraction and classification at volume

This work belongs in the small column. Claude Haiku 5.5 and GPT-6 Luna cost $0.10 / $0.50, Gemini 3.1 Flash-Lite $0.25 / $1.50 and DeepSeek V4.1-Flash $0.15 / $0.60 off-peak. Batch APIs halve the OpenAI, Anthropic and Gemini prices on top. The highest volumes fit gpt-oss-120b at $0.03 to $0.15 input on OpenRouter or Groq, or a small open model on your own GPUs.

Long documents

Check the context price as well as the window. Claude Sonnet 5.5 and Opus 5.5 charge a flat rate across 1M tokens. Gemini 4 Argon offers 1,048,576 tokens combined. GPT-6.1 Sol accepts 922K input tokens but doubles the input price above 272K, according to Lindy. Claude Haiku 5.5 moves to a tier five times more expensive above a 100K-token prompt. Grok 4.7 doubles above 200K. For repeated documents, prompt caching cuts input cost by 90% or more at OpenAI, Anthropic, Google, DeepSeek and Mistral, and by 75% on Grok 4.7.

On-device and privacy

Gemma 4 E4B or Phi-4-mini run on a laptop, and gpt-oss-20b fits if you have 16 GB. Ministral 3 suits cases where Apache 2.0 matters. Apple's AFM 3 is available through the Foundation Models framework on supported Apple devices. None of these has a token price.

Several models in production

A production system rarely runs on one model. Sending each request to the cheapest model that passes your tests is the single largest lever in most LLM budgets. Model routing explains how to set that up. The pricing comparison has the current numbers for every row of the map.

What to do next

Pick one task you run at volume. Run it through a model one column to the right of what you use now, with your own evaluation set. Do it before 1 December, when the Gemini 4 Argon promotional price ends. Write down the price per million tokens for each model you call today, including the cached and batch rates. If your notes lack a price for a model on this map, look at the vendor page in the sources below. For a second opinion on which column each of your workloads belongs in, we run a short audit.

Sources