Models

Large Language Models (LLMs) differ in reasoning ability, speed, cost, context size and intended workload. There is no universally “best” model; model selection is usually a trade-off between capability, latency and cost.

The major general-purpose model families are currently OpenAI GPT, Anthropic Claude, Google Gemini and DeepSeek.

Last updated: August 2026. Model names and prices change frequently.

1. Pricing

1.1 Consumer Plans

Prices below are approximate USD/month for individual users.

Provider Tier USD / Month What It Means Usage
ChatGPT Free $0 Basic access to GPT models and tools Limited GPT-5.x usage; stricter limits on files, images and data analysis
Go $8 Low-cost paid tier More usage than Free
Plus $20 Main individual plan; advanced reasoning, files, images, Deep Research and other tools Higher model and tool limits than Free; limits can vary
Pro $100 / $200 Heavy professional usage and access to Pro models/features $100 tier ≈ 5× Plus usage; $200 tier ≈ 20× Plus usage
Claude Free $0 Basic Claude chat, coding, writing and image/text analysis Limited usage that resets periodically
Pro $20 Everyday professional use; more models, Projects and extended thinking Higher usage than Free
Max $100 / $200 Heavy Claude usage, higher output limits, Research and priority access Approximately 5× or 20× Pro usage
Gemini Free $0 Standard Gemini access Standard compute allowance; 32K context
AI Plus $7.99 Entry paid tier ≈ 2× Free usage; 128K context
AI Pro $19.99 Regular power-user tier with larger context and advanced features ≈ 4× Free usage; 1M context
AI Ultra $249.99 Highest access including Deep Think and premium generation features Approximately 5×–20× AI Pro usage; 1M context
DeepSeek Free $0 Consumer web/app access to DeepSeek models Service-dependent limits; API is separately usage-based

The tiers can broadly be interpreted as:

Free
└── Try the service
    Limited model/tool usage
    Smaller context or stricter rate limits

Standard / Plus / Pro
└── Normal individual power user
    Higher limits
    Better model selection
    Advanced reasoning and tools

Max / Ultra / Pro High-Usage
└── Heavy professional user
    Much higher compute allowance
    Priority access
    Most capable / expensive modes

1.2 Usage Limits

Consumer AI services generally do not provide a fixed number of tokens each month.

Instead they use limits based on factors such as:

For example, Gemini uses a compute-based allowance. AI Plus receives roughly the standard allowance, AI Pro , and AI Ultra substantially more. Limits refresh periodically and more expensive reasoning modes consume the allowance faster.

ChatGPT similarly applies model-specific and tool-specific limits. Free users receive limited access within a time window, while Plus and Pro increase the available compute. Exact message caps may change based on capacity.

Claude follows a similar model: Pro increases usage above Free, while Max provides approximately 5× or 20× the usage available on Pro.

Therefore, “$20/month” should not be interpreted as unlimited model inference. Consumer subscriptions purchase a usage allowance and feature access, while APIs are billed directly by tokens.

Sources: ChatGPT Free Tier, ChatGPT Plus, Claude Pricing, Gemini Usage Limits.

1.3 API Pricing

API pricing is based on how many tokens the model processes.

For example, if a model costs $2 / 1M input tokens and $10 / 1M output tokens, a request using 100,000 input tokens and generating 10,000 output tokens would cost:

Input:  100,000 × $2 / 1M  = $0.20
Output:  10,000 × $10 / 1M = $0.10
Total = $0.30

In simple terms:

Input  = what you give the model
Output = what the model gives you

Output tokens are generally more expensive because generating new tokens requires more computation than processing the input.

Typical text API price per 1 million tokens:

ProviderModelInputOutputPosition
OpenAIGPT-5.6 Sol$5.00$30.00Flagship
OpenAIGPT-5.6 Terra$2.00$12.00Balanced
OpenAIGPT-5.6 Luna$0.20$1.20High-volume / cheap
AnthropicClaude Fable 5$10.00$50.00Maximum capability
AnthropicClaude Opus 5$5.00$25.00Complex work
AnthropicClaude Sonnet 5$2.00*$10.00*Balanced
AnthropicClaude Haiku 4.5$1.00$5.00Fast / cheap
DeepSeekV4 Pro$0.435$0.87Higher capability
DeepSeekV4 Flash$0.14$0.28High-volume / cheap

* Sonnet 5 introductory pricing until 31 August 2026; standard pricing becomes $3 input / $15 output.

Google Gemini pricing varies more by model and context size; see the official pricing page rather than treating one price as representative.

Sources: OpenAI API Pricing, Claude API Pricing, Gemini API Pricing, DeepSeek API Pricing. (OpenAI Developers)


2. Purpose

Model families generally contain several sizes rather than one model.

Family Tier Profile Key Strengths Trade-off Best For
GPT-5.6SolFlagshipMaximum reasoning, coding and toolsHighest cost and latencyDifficult reasoning and coding
TerraBalancedStrong capability with practical speed and costLess capable than SolGeneral-purpose work
LunaEfficientFast and inexpensiveLower reasoning ceilingHigh-volume workloads
ClaudeFableFlagshipDeep reasoning and long-horizon agentsHighest cost and slower responsesComplex research and agents
OpusAdvancedComplex analysis, coding and engineeringMore expensive than SonnetComplex engineering
SonnetBalancedStrong coding and general toolsLess capable than OpusEveryday coding
HaikuEfficientFast and economicalLower capability ceilingReal-time and batch tasks
GeminiProFlagshipStrong reasoning, multimodal capability and long contextHigher cost and latencyComplex multimodal work
FlashBalancedFast multimodal and agentic workloadsLess capable than ProGeneral applications
Flash-LiteEfficientVery fast and low costLower reasoning ceilingClassification and extraction
DeepSeek V4ProAdvancedStrong reasoning, coding and tool useHigher cost and lower concurrencyReasoning and coding
FlashEfficientFast, inexpensive, with optional thinkingLower capability than ProCost-sensitive workloads

Model tiers are usually differentiated by training quality, reasoning budget, inference compute, context limits, tool support, latency and price—not by parameter count alone. Proprietary providers generally do not publish exact parameter counts, so this page focuses on observable behavior and practical trade-offs.

The pattern is broadly:

High Capability
      ↑
Flagship / Pro / Opus / Sol
      |
Balanced / Sonnet / Terra / Flash
      |
Mini / Haiku / Luna / Flash-Lite
      ↓
High Speed + Low Cost

Provider naming is marketing rather than a technical standard; models at similarly named tiers are not necessarily equivalent. (Claude Platform)


3. Token Usage

A token is a small unit of text processed by a model. English text averages roughly a few characters per token, although tokenisation varies between models and languages.

API cost is approximately:

Cost =
    input_tokens × input_price
  + output_tokens × output_price

There are several important token types:

TokenMeaning
InputPrompt, conversation history, documents and retrieved context
OutputTokens generated by the model
Cached inputReused prompt/context, normally significantly cheaper
Thinking / reasoningInternal reasoning compute; billing treatment varies by provider

The context window is the maximum amount of information available to the model during one interaction.

Large context is useful for:

However, a larger context window does not automatically mean better reasoning. Sending unnecessary context also increases latency and cost.


4. Performance

Model capability changes quickly, so benchmarks should be treated as indicators rather than absolute rankings.

Representative current generations:

ProviderGenerationReleaseMain Characteristic
AnthropicClaude Fable 5Jun 2026Highest-end Claude
AnthropicClaude Opus 5Jul 2026Frontier coding / knowledge work
OpenAIGPT-5.6Jul 2026Reasoning, coding and professional work
DeepSeekV4Apr–Jul 2026Strong open-weight cost/performance
GoogleGemini 3.1 ProFeb 2026Reasoning + multimodal + long context

OpenAI GPT-5.6 itself contains Sol, Terra and Luna, while Claude uses families such as Fable, Opus, Sonnet and Haiku. A version number therefore does not tell the full story — the model tier matters as well. (OpenAI)

One useful independent comparison is the Artificial Analysis Intelligence Index. As of August 2026, frontier models such as Claude Opus/Fable and GPT-5.6 occupy the upper end of its intelligence measurements, while models such as DeepSeek V4 compete particularly strongly on cost per task. (Artificial Analysis)

Common benchmark categories include:

Benchmark TypeMeasures
GPQADifficult scientific reasoning
Humanity's Last ExamBroad expert-level reasoning
SWE / coding benchmarksSoftware engineering ability
Terminal-BenchAgentic computer / terminal tasks
MMMUMultimodal reasoning
Arena / EloHuman preference between outputs

No single benchmark represents overall model quality.


5. Rule of Thumb

For most applications:

RequirementStart With
Hardest reasoningGPT-5.6 Sol / Claude Opus or Fable
Coding / agentsGPT-5.6 Sol / Claude Opus or Sonnet
General everyday workGPT-5.6 Terra / Claude Sonnet / Gemini Flash
Very large context / multimodalGemini
Cheapest API workloadsDeepSeek V4 Flash / GPT-5.6 Luna
High-volume simple tasksLuna / Haiku / Flash-Lite
Self-hosting / open weightsDeepSeek and other open-weight families

A useful selection rule is:

1. Choose the cheapest model that reliably completes the task.
2. Increase model capability when quality becomes insufficient.
3. Use reasoning effort only when the task actually needs reasoning.
4. Minimise unnecessary context.
5. Measure models against your own workload, not only public benchmarks.

For current comparisons, the most useful references are:

Contents