Reading note

Don’t pay for a genius to summarize a spreadsheet.

Match the model to the job. Save the expensive one for hard thinking.

Cost check

Same answer, 12× cheaper.

High-volume chores rarely need a flagship model.

BetterLLMs

The right model for your work. More capable answers, less wasted spend.

Task first

What are you trying to do?

Write, debug, or review code

You need a model that follows instructions closely and still stays affordable for many back-and-forth edits.

Best overall

Claude Sonnet 5

Current

Anthropic · version 5

claude-sonnet-5

Daily coding, writing, and planning at a lower Sonnet price

Strength: Newest Sonnet, cheaper than 4.5 / 4.6
Weakness: If a company still requires Sonnet 4.6, you cannot swap silently
Context and pricing notes: Excellent memory for most projects.
How versions differ: $2 to read / $10 to write per million tokens — less than Sonnet 4.6’s $3 / $15.

Typical coding chat: $0.04

Reading what you send: about $2 each time it reads a million tokens.

Writing the reply: about $10 for a million tokens it types back. Replies cost more than reading, because generating text is harder.

Cost MediumSpeed FastContext Very long

Best value

DeepSeek V4.1 Flash

Current

DeepSeek · version V4.1 Flash

deepseek-flash

Budget coding, tool use, and visual inputs

Strength: Thinking and non-thinking modes with a 1M-token context
Weakness: Peak rates depend on time of day; test reliability on your tasks
Context and pricing notes: 1M-token context; use the deepseek-flash API alias.
How versions differ: Peak uncached rates: $0.30 / $1.20. Off-peak rates are half.

Typical coding chat: < $0.01

Reading what you send: about $0.3 each time it reads a million tokens.

Writing the reply: about $1.2 for a million tokens it types back. Replies cost more than reading, because generating text is harder.

Cost LowSpeed FastContext Very long

Cheapest usable

GPT-6 Luna

Current

OpenAI · version Luna

gpt-6-luna

Focused tasks, simple edits, and high-volume processing

Strength: Lowest base rates in the GPT-6 lineup
Weakness: Validate quality before using it for complex reasoning
Context and pricing notes: 1.05M-token context; up to 128K output tokens.
How versions differ: $0.10 / $0.50 per million tokens, below GPT-4o mini's base rates.

Typical coding chat: less than a cent

Reading what you send: about $0.1 each time it reads a million tokens.

Writing the reply: about $0.5 for a million tokens it types back. Replies cost more than reading, because generating text is harder.

Cost LowSpeed FastContext Very long

What this means for cost

Claude Sonnet 5 costs more than DeepSeek V4.1 Flash. For the same coding session, that’s about $0.04 vs < $0.01.
GPT-6 Luna is the budget pick when the work is simple.

Rule of thumb: use the cheapest usable model until the answer quality starts to cost you time. Then step up one level.

Context switch

Thinking of changing models?

Switching mid-chat can drop details and change the bill. Pick two models.

You are onSwitching to

Extra cost

Lower bill: a typical coding session goes from about $0.04 to less than a cent.

Risk of losing context

Low

Recommendation

Switch

You keep similar memory and pay less. Paste a short recap of the goal when you switch.