Methodology & Primary Sources

How this guide works

Where every number on AI Toolbelt comes from, what it means, and what it leaves out.

About AI Toolbelt

A free public guide to choosing an AI app or API model. Prices, limits, and scores are spread across dozens of vendor pages and change often; this guide puts them in one place. Labs cannot pay for placement. The site has no affiliate links, sets no cookies, and runs no analytics.

Prices

Standard, uncached API rates in US dollars per million tokens, taken directly from each lab’s pricing or model page. Announced price changes are stored with their effective start date, and the site switches automatically on that day. Unknown prices are shown as “Not published”, never as zero.

Scores

The Intelligence Index is Artificial Analysis’s combined score across its independent benchmark suite. Points are not percentages. Each model’s headline score is from Artificial Analysis’s default run, usually its highest reasoning effort; the map, table, and compare page can switch to any other tested effort level.

Effort and what it costs

Reasoning effort never changes the price per token at any of the five labs. It changes how many tokens a model spends: every lab bills thinking tokens as output. For flagship models, cost per task usually rises 5× to 10× from the lowest effort setting to the highest.

Effort can matter as much as model choice. GPT-6.2 Astra and Claude Fable 5.2 share an $8 / $40 list price, but at Max effort Astra spent $4.00 per benchmark task for 62.1 points while Fable spent $26.54 for 64.6.

Speed tiers

Faster processing is the one setting that changes the per-token price: OpenAI Fast is 2× and Ultrafast is 3× on GPT-6.2 until 28 February 2027 (then 6×); Anthropic fast mode is 2×; Google and xAI Priority are 1.8× and 2×. The calculator applies them to every token rate and never combines them with batch discounts.

Pricing rules the calculator applies

  • OpenAI: requests over 272K input tokens pay 2× input and 1.5× output for the whole request.
  • Claude Haiku 5.5: prompts over 100K tokens pay $0.50 / $2.50 instead of $0.10 / $0.50, for the whole request.
  • Grok 4.7: prompts of 200K tokens or more pay $4 / $12 instead of $2 / $6, for the whole request.
  • Gemini 4 Neon, Carbon, Hydrogen and Argon: 65,536 output tokens by default, up to 1,048,576 with long decode, inside a shared 1,048,576-token window.

Apps, plans & plain ratings

The consumer picker matches each everyday job to the features it requires: looking things up needs web search, learning and work documents need file uploads, and making images needs image generation. For your budget it shows the richest plan each app offers at or under that price, ordered cheapest first rather than ranked by opinion.

On app pages, models receive a plain rating from the Intelligence Index: Top tier (60+), Strong (50+), Capable (40+), and Basic below 40. Recommended effort on model pages is the cheapest setting within 2 points of the model’s best score.

Privacy & reuse

Your compare list and theme preference stay in your browser’s local storage. Data is CC BY 4.0 and code is MIT licensed. Product names and logos (Lobe Icons, MIT) are trademarks of their owners used for identification.

Plain-language glossary (8 terms)

Every technical term used across our maps, tables, and cost calculator, defined without jargon.

Tokens

The small pieces of text a model reads and writes, roughly three-quarters of an English word each. API prices are quoted per million tokens.

Context window

How much text a model can hold at once, counting both what you send and what it writes back.

Reasoning effort

A setting for how long a model thinks before it answers. It never changes the price per token; higher settings spend more thinking tokens and time, so each task costs more.

Intelligence Index

Artificial Analysis’s combined score across a suite of independent tests. Points are not percentages.

Batch

Sending requests to be processed later, usually within a day, at a lower price where the lab offers one.

Cached input

Prompt text the lab has already processed and stored, such as a long system prompt. It is billed at a lower rate when reused.

Thinking tokens

The working-out a model writes before its answer, also called reasoning tokens. Every lab here bills them as output tokens, even when you never see them.

Speed tier

Paying more per token for faster or first-in-line processing: Fast and Ultrafast at OpenAI, fast mode at Anthropic, Priority at Google and xAI. Unlike effort, it changes the price.

Primary sources (59 of 59)

Every vendor page and independent benchmark checked on 21 Nov 2026.

SourceNotesLink
OpenAI · Terra specificationsOfficial documentationOpen
OpenAI · Sol specificationsOfficial documentationOpen
OpenAI · Astra specificationsOfficial documentationOpen
OpenAI · Luna specificationsOfficial documentationOpen
OpenAI · API pricing (Ultrafast, background)Official pricingOpen
Anthropic · API pricingOfficial pricingOpen
Anthropic · models overviewOfficial documentationOpen
Anthropic · Claude plansOfficial pricingOpen
Anthropic · Haiku 5.5 announcementVendor announcementOpen
Google · Gemini API pricingOfficial pricingOpen
Google · Neon model cardOfficial model cardOpen
Google · Hydrogen model cardOfficial model cardOpen
Google · Introducing Gemini 4 NeonVendor announcementOpen
Google · Gemini 4 Argon launch (1M output)Vendor announcementOpen
Google · Carbon model cardOfficial model cardOpen
Microsoft · Ignite model lineupVendor announcementOpen
Microsoft · Code 2 evaluation footnotesVendor evaluationOpen
Microsoft · MAI Thinking 1Vendor announcementOpen
Artificial Analysis · model leaderboardIndependent measurementOpen
Artificial Analysis · Google endpointsIndependent measurementOpen
xAI · Grok 4.7 model pageOfficial documentationOpen
Artificial Analysis · Grok 4.7Independent benchmarkOpen
Google · Gemini 4 Carbon model page (token limits)Official documentationOpen
Microsoft Azure · Foundry Models pricingOfficial pricingOpen
GitHub Docs · Models and pricing for GitHub CopilotOfficial documentationOpen
OpenAI · GPT-6 Astra specificationsOfficial documentationOpen
OpenAI · GPT-6.1 Sol specificationsOfficial documentationOpen
Google · Gemini 3.8 Flash model pageOfficial documentationOpen
Artificial Analysis · GPT-6 AstraIndependent measurementOpen
Artificial Analysis · GPT-6.1 SolIndependent measurementOpen
Artificial Analysis · Gemini 3.8 FlashIndependent measurementOpen
OpenAI · ChatGPT Work & Codex pricingOfficial pricing, surfaces, models and per-plan limits for CodexOpen
OpenAI · ChatGPT plans and pricingOfficial pricing; shows local currency by region (read from Canada, CA$), used here for plan contents, models and feature table onlyOpen
OpenAI · Introducing GPT-6.2 Astra and SolVendor announcement, 17 Nov 2026Open
OpenAI Help · About ChatGPT Pro tiersOfficial helpOpen
OpenAI Help · What is ChatGPT Go?Official help; no annual billing for Go, Plus or ProOpen
Google · Gemini subscriptions (US)Official pricing; Antigravity rate limits per Google AI planOpen
SpaceXSI (xAI) · PricingOfficial pricing; same plans as cursor.com/pricingOpen
SpaceXSI (xAI) · Grok product pageOfficial product page; feature listOpen
SpaceXSI Docs · What is The X Subscription?Official docs (updated 2026-10-12); one plan for Grok, X and Cursor, shared weekly pool, switching is optionalOpen
SpaceXSI Docs · X Premium customers moving to The X SubscriptionOfficial docs (updated 2026-11-03); X benefits by planOpen
X Help Center · About X PremiumOfficial help; X Premium is now part of The X Subscription; legacy US prices Basic $3, Premium $8, Premium+ $40 a monthOpen
Microsoft · Copilot plans for individualsOfficial pricing; monthly prices (a 30% first-period promotion was running on 2026-11-20)Open
Microsoft · Microsoft 365 for individuals (compare plans)Official pricing; yearly prices and Copilot feature tableOpen
Microsoft · Copilot homeOfficial; 'Microsoft Copilot is free with a signed-in Microsoft account'Open
NBER w34255 · How People Use ChatGPT (2025)Primary research (OpenAI economists); topic shares read from the full paper, section 4Open
OpenAI · How people are using ChatGPTVendor summary of the NBER paperOpen
Anthropic Economic Index · March 2026 report (Learning curves)Primary research; Feb 5-12, 2026 sampleOpen
Anthropic Economic Index · September 2025 reportPrimary research (Claude.ai and API usage)Open
Pew Research Center · 34% of U.S. adults have used ChatGPT (2025-06-25)Survey of 5,123 U.S. adults, Feb 24 - Mar 2, 2025Open
Reuters Institute · Generative AI and News Report 2025Survey in 6 countries, fieldwork Jun 5 - Jul 15, 2025Open
Anthropic · Claude Code product pageOfficial; surfaces and which plans include itOpen
Anthropic · Claude Code overview docsOfficial documentation; surfacesOpen
Google · Gemini CLI READMEOfficial repository; status after the move to Antigravity CLIOpen
Google · Gemini CLI quotas and pricingOfficial docs (last updated Jun 18, 2026)Open
GitHub · Copilot plansOfficial pricingOpen
GitHub Docs · Plans for GitHub CopilotOfficial documentationOpen
Cursor · PricingOfficial pricing (shared plans with Grok / X)Open
SpaceXSI (xAI) · Grok BuildOfficial product pageOpen