Choosing a model in late 2026 comes down to two questions: does it run on someone else's servers or yours, and what does it cost per million tokens or per GPU. Every price, parameter count and licence below was read from the vendor's own pricing page, model card or docs on October 26, 2026; where a vendor published no figure, we say so.

Frontier models you rent through an API

Anthropic

Four current models, all with a 1M context window and 128K max output; prices checked October 26, 2026:

  • Claude Fable 5.1: $10 / $50 per million input/output tokens. The top tier, "for demanding reasoning and long-horizon agentic work", and the slowest of the four.
  • Claude Opus 5.5: $4 / $20. Anthropic's "daily driver for agentic coding and enterprise work", and where it tells you to start for most workloads.
  • Claude Sonnet 5.5: $2 / $10. "The best combination of speed and intelligence."
  • Claude Haiku 5.5: $0.10 / $0.50 for prompts up to 100,000 tokens, $0.50 / $2.50 above that. Built for classification, extraction and routing.

Batch requests are 50 percent off.

OpenAI

Three current flagship-tier models; prices cover prompts up to 272K tokens.

  • GPT-6 Astra: $10 / $50. "Our most capable model for the most demanding work." 1,050,000-token context, 128K output, text and image input.
  • GPT-6.1 Sol: $2 / $10. "Near-Astra performance for complex work at a lower cost."
  • GPT-6 Luna: $0.10 / $0.50. "Our most efficient model for focused, high-volume tasks."

Batch pricing halves these: Astra $5 / $25, Sol $1 / $5.

Google

  • Gemini 4 Argon: $2 / $10 through November 30, 2026, then $4 / $20 from December 1, 2026. Google's "flagship reasoning model with a unified 1M context window".
  • Gemini 3.8 Flash: $0.75 / $3.75 through December 31, 2026, then $1.50 / $7.50. "Engineered for long-horizon software engineering."
  • Gemini 3.5 Flash-Lite: $0.30 / $2.50, the high-throughput option.

A budget built on the October 2026 Argon price doubles in five weeks.

xAI and Mistral

  • Grok 4.7: $2 / $6 for requests under 200K prompt tokens ($4 / $12 above), 500K context. Grok 4.3: $1.25 / $2.50 with a 1M context.
  • Mistral Large 4, announced October 6, 2026: about 1 trillion parameters, 52 billion active, natively multimodal, $1.36 / $4.18. Mistral promises to "release the weights by the end of the month"; on October 26, 2026 its Hugging Face organisation listed no Large 4 repository.
  • Mistral Medium 3.5: $1.5 / $7.5, 256K context, weights on Hugging Face under a modified MIT licence. Mistral Small 4: $0.15 / $0.60, 119B total / 6B active, Apache 2.0.
Model Maker Price (input / output per 1M tokens) Best for
Claude Fable 5.1 Anthropic $10 / $50 Hardest reasoning, long agent runs
Claude Opus 5.5 Anthropic $4 / $20 Agentic coding, general default
Claude Sonnet 5.5 Anthropic $2 / $10 Fast coding and agents
Claude Haiku 5.5 Anthropic $0.10 / $0.50 (prompts up to 100K) Classification, extraction, routing
GPT-6 Astra OpenAI $10 / $50 Top-tier reasoning
GPT-6.1 Sol OpenAI $2 / $10 Mid-tier general and coding
GPT-6 Luna OpenAI $0.10 / $0.50 High-volume cheap tasks
Gemini 4 Argon Google $2 / $10 to Nov 30, 2026, then $4 / $20 Multimodal reasoning, 1M context
Gemini 3.8 Flash Google $0.75 / $3.75 to Dec 31, 2026, then $1.50 / $7.50 Software engineering on a budget
Grok 4.7 xAI $2 / $6 (under 200K prompt) General use, 500K context
Mistral Large 4 Mistral $1.36 / $4.18 Multimodal, European vendor
Mistral Small 4 Mistral $0.15 / $0.60 Cheap, Apache 2.0 weights

What the leaderboards say

On October 26, 2026 the Artificial Analysis Intelligence Index had Claude Opus 5.5 at max effort on top with 58, then Claude Sonnet 5.5 (max) 56, Opus 5.5 (xhigh) 56, Meta's Muse Wonder Full (max) 56 and Moonshot's Kimi K3.1 Pro 55. GPT-6 Astra (max), Gemini 4 Argon (high) and Claude Fable 5.1 (max) all scored 53; GPT-6.1 Sol (max) scored 52.

Three caveats. Opus 5.5 spans 51 at medium effort to 58 at max, so "best" depends on how much thinking you pay for. Fable 5.1, which Anthropic positions above Opus 5.5, scores below it here, so the index is not measuring the long-horizon work it is sold for. And evaluators disagree; LMArena would not load for us on October 26, 2026, so we do not quote it.

Open-weight models you can host yourself

Figures come from each model's Hugging Face card or the vendor's docs.

  • MiMo-V2.6-Pro (Xiaomi, published as MiMo-V2.6-Pro-RL): sparse MoE, 1.02T total / 42B active, 1M context, MIT, text, image, video and audio input. The card's vLLM example uses tensor-parallel 8, so plan on an 8-GPU node.
  • Kimi K3 (Moonshot AI): 2.8T total / 104B active, 896 experts with 16 active per token, 1,048,576-token context, text and image input, custom Kimi K3 License. At 4 bits per weight that is roughly 1.4 TB: a multi-node job.
  • GLM-5.3 (Z.ai): 753B parameters (active count not stated on the card), 1M context and 128K max output per Z.ai's docs, text only, custom glm-5.3 licence. GLM-5.3-Flash: 320B / 18B active, MIT, with image input; the card gives no context length.
  • DeepSeek V4.1-Flash: 552B backbone (the safetensors metadata reads 763B), 8B active during prefill and 16B during decode, 1M context, MIT, image and text input. The earlier V4-Flash is 284B / 13B active, 1M context, MIT, text only.
  • NVIDIA Nemotron 3.5 Bolt 74B-A9B-24GB: 74B / 9B active, 256K context, OpenMDW-1.1, text, image and video input. Released on Hugging Face October 15, 2026 with native 2-bit quantization-aware weights "to fit in 24GB of memory". The only model here for a single consumer card.
  • Mistral Small 4: 119B / 6B active, 256K context, Apache 2.0.
Model Maker Size / licence Context Best for
MiMo-V2.6-Pro Xiaomi 1.02T / 42B active, MIT 1M Top open-weight quality, multimodal
Kimi K3 Moonshot AI 2.8T / 104B active, Kimi K3 License 1M Cluster-scale capacity
GLM-5.3 Z.ai 753B, glm-5.3 licence 1M Coding and agents on a multi-GPU node
GLM-5.3-Flash Z.ai 320B / 18B active, MIT Not stated Cheaper GLM with image input
DeepSeek V4.1-Flash DeepSeek 552B / 16B active, MIT 1M Efficient MoE with image input
DeepSeek V4-Flash DeepSeek 284B / 13B active, MIT 1M Text-only, FP4/FP8
Nemotron 3.5 Bolt 74B-A9B NVIDIA 74B / 9B active, OpenMDW-1.1 256K One 24 GB GPU, vision
Mistral Small 4 Mistral 119B / 6B active, Apache 2.0 256K Permissive licence, small node

Artificial Analysis' open-weights ranking the same day (Intelligence Index v4.3.2): MiMo-V2.6-Pro 46, DeepSeek V4.1 Flash (max) 45, GLM-5.3 (max) 45, GLM-5.3-Flash 42, Nemotron 3.5 Super VL 26, so the best open model trails Opus 5.5 at max by 12 points.

Which one to pick

Coding and agents

Rent Claude Opus 5.5 at $4 / $20, or Sonnet 5.5 at half that when latency matters more. GPT-6.1 Sol is OpenAI's tier at the same $2 / $10. Gemini 3.8 Flash is the bargain while its $0.75 / $3.75 promotion lasts. Self-hosted, GLM-5.3, MiMo-V2.6-Pro and DeepSeek V4.1-Flash are the serious options; all three need a multi-GPU node.

Cheap bulk and classification

Haiku 5.5 and GPT-6 Luna both list $0.10 / $0.50; keep Haiku prompts under 100,000 tokens or the price quintuples. Gemini 3.5 Flash-Lite ($0.30 / $2.50) and Mistral Small 4 ($0.15 / $0.60) come next. Batch endpoints halve Anthropic and OpenAI prices. If the data cannot leave your rack, Nemotron 3.5 Bolt on one 24 GB card or GLM-5.3-Flash on a bigger node.

Multimodal

Google covers the most input types on the API side: Gemini 3.5 Flash-Lite is priced at $0.30 for text, image, video and audio, with Gemini 4 Argon as the reasoning tier above it. Current Claude and GPT-6 models take text and image only; Mistral Large 4 is described as natively multimodal. Among open weights, MiMo-V2.6-Pro takes text, image, video and audio; Nemotron 3.5 Bolt takes image and video.

Self-hosting: one GPU or a cluster

At 4-bit quantization, budget roughly 24 GB of VRAM for a 30B-class model and 48 GB for a 70B-class dense model, plus KV cache that grows with context length. Nemotron Bolt is the exception, pre-packed for 24 GB. MoE models run fast once loaded, but you still store every expert: GLM-5.3-Flash (320B), DeepSeek V4-Flash (284B) and MiMo (1.02T) each need a multi-GPU node; Kimi K3 needs several. Add GPU hours, power and the engineer who keeps vLLM or SGLang current. See our hardware guide for sizing details.

Licence gotchas for commercial use

  • MIT (MiMo, DeepSeek, GLM-5.3-Flash), Apache 2.0 (Mistral Small 4) and OpenMDW-1.1 (Nemotron; "permissive by default" per openmdw.ai) are safe for ordinary commercial use.
  • Kimi K3 License: if you run a model-as-a-service business with over $20 million revenue in any 12 months you "must enter into a separate agreement with Moonshot AI" first, and products over 100 million monthly users or $20 million monthly revenue must show "Kimi K3" in the interface. Internal use is exempt.
  • GLM-5.3 licence: model-as-a-service operators with combined revenue over $10 billion "must pass Z.AI's security review" before commercial use. Everyone else gets MIT-style terms.
  • Mistral Medium 3.5 is under a modified MIT licence; Mistral Large 4's licence was not published with the announcement.
  • Both custom licences target reselling the model as a service; embedding it inside your own product is excluded from their thresholds.

Sources