basforge

Search titles, text and the “If it doesn’t work” cases.

    RU
    Cheat sheetAILanguage models

    Models as of October 2026: families, tiers, prices and windows

    Anthropic and OpenAI have converged on one price grid: the strongest model is $10 per million input tokens and $50 per million output, the mid-tier $2 and $10, the small one $0.10 and $0.50. Google is cheaper at the flagship and pricier at the small end. Below: who’s who, what it costs and how much fits in a request. Prices change every few months, so check the vendor’s pricing page before choosing.

    How to read the prices

    Prices are in dollars per 1 million tokens (MTok), separately for input and output. Output costs 5–8 times more than input: every output token is a separate pass of the model, while input is processed in one pass. The model’s reasoning is billed as output (Match reasoning to the task). How many tokens are in a text — Tokens.

    Anthropic — Claude

    Model API ID Input Output Window
    Fable 5.1 — the strongest claude-fable-5-1 $10 $50 1M
    Opus 5.5 — the everyday flagship claude-opus-5-5 $4 $20 1M
    Sonnet 5.5 — mid-tier claude-sonnet-5-5 $2 $10 1M
    Haiku 5.5 — small and fast claude-haiku-5-5 $0.10 $0.50 1M
    • Output up to 128K tokens on all four.
    • Haiku 5.5 costs $0.10 / $0.50 for requests up to 100K tokens; longer ones are $0.50 / $2.50.
    • Cache reads are $0.20 per MTok on Opus 5.5 and Sonnet 5.5, $0.25 on Fable 5.1 (Prompt cache).
    • Batched (Batch API) — half price, with the answer within 24 hours (Non-urgent API requests go as a batch at half price).
    • Opus 5.5 can’t turn reasoning off; effort sets the depth, medium by default (Match reasoning to the task).
    • Fable 5.1 thinks longer: one request on a hard task takes minutes.

    OpenAI — GPT

    Model Input Cache Output
    gpt-6-astra — flagship $10 $1 $50
    gpt-6.1-sol — mid-tier $2 $0.10 $10
    gpt-6-luna — small $0.10 $0.01 $0.50
    • Requests longer than 272K tokens are billed at a separate, higher price.
    • The window is 1.05M tokens, of which up to 128K for output: that’s what the Azure documentation and summaries say; the figure wasn’t checked on OpenAI’s page.
    • Batched — half price: $5 / $25 for gpt-6-astra.
    • The pricing page still lists the 5.6 generation (gpt-5.6-sol — $4 / $20) — use 6 for new code.

    Google — Gemini

    Model Input Output Free tier
    Gemini 3.1 Pro (Preview) — flagship $2 up to 200K, $4 beyond $12 up to 200K, $18 beyond no
    Gemini 3.8 Flash — mid-tier $0.75 until the end of 2026, $1.50 from January 1, 2027 $3.75, from 2027 — $7.50 yes
    Gemini 3.5 Flash-Lite — small $0.30 $2.50 yes
    • The window on 3.1 Pro and 3.8 Flash is 1,048,576 tokens, output up to 65,536.
    • Flash and Flash-Lite have a free tier. Google uses free-tier data to improve its products, and human reviewers may read prompts and responses — so say the Gemini API terms. For work documents, use a paid key: paid-tier data isn’t used for improvement.
    • Batched — half price.

    Open models — downloadable weights

    They run on your own machine (A model on your own hardware) or at third-party hosts via API, usually cheaper than closed models. Versions change every two or three months; the list comes from summaries as of October 2026, and none were run. Check the version on the model’s page before choosing.

    Family Whose What’s known
    Kimi K3 Moonshot AI among the strongest open models in summaries, 1M window, MIT license
    GLM-5.x Z.ai (Zhipu) 5.2 and 5.3 are close to Kimi on agentic coding
    DeepSeek V4.1 DeepSeek the Flash version targets code and agents; DeepSeek has its own cheap API
    Qwen 3.6 / 3.8 Alibaba from small to huge; 27B is a frequent pick for your own server
    Gemma 4 Google small models for a laptop and a single GPU, Apache 2.0
    Mistral Mistral AI a European vendor; Large 3 and Large 4 in different summaries
    Llama 4 Meta Scout and Maverick, April 2025; summaries know of no newer open Llama

    “Open weights” doesn’t always mean open code and training data: the weights download, but how the model was trained isn’t published. Some licenses restrict commercial use — read the license of the specific model.

    Tiers side by side

    Tier Claude OpenAI Google Input / output
    the strongest Fable 5.1 gpt-6-astra — $10 / $50
    everyday flagship Opus 5.5 — 3.1 Pro $2–4 / $12–20
    mid-tier Sonnet 5.5 gpt-6.1-sol 3.8 Flash $0.75–2 / $3.75–10
    small Haiku 5.5 gpt-6-luna 3.5 Flash-Lite $0.10–0.30 / $0.50–2.50

    The table shows price, not quality on your tasks. Compare by running them: ten typical requests on two neighbouring tiers, and see where the cheaper one copes (Price of a solved task).

    Where to get fresh figures

    What Where
    Claude models and prices platform.claude.com/docs — the models and pricing section; the model list with windows is GET /v1/models
    OpenAI prices developers.openai.com/api/docs/pricing
    Gemini prices ai.google.dev/gemini-api/docs/pricing
    open models the model card on Hugging Face: size, license, window
    comparisons on tasks LMArena, Artificial Analysis — a general guide, not a substitute for your own run