Skip to guide
RepoSpend
RepoSpend Guide

The AI Coding Price War:
Claude 5.5 vs OpenAI GPT-6

Lower token prices are only part of the story. Compare today’s models with their predecessors, including the newly released Sonnet 5.5, and see what these prices do—and don’t—tell you about coding costs.

Built for developersClear prices, real context
Local by designAI coding data stays on your machine
Usage over hypeUnderstand cost by repository

Models (specific models)

Individual model names, capabilities, and API pricing.

GPT-6 AstraGPT-6 SolGPT-6 LunaClaude Opus 5.5Claude Sonnet 5.5

Prior reference models

Claude Opus 5Claude Sonnet 5GPT-5.6 Sol

Products / subscriptions (how you use them)

Interfaces and plans that bundle access to models.

ChatGPTCodexClaudeClaude Code

API token rates are not subscription prices. Plan limits and included usage follow each product’s own terms.

Independent explainer based on published provider information. The release sequence does not establish that either company caused the other’s price changes.

Recent model releases

Key announcements behind this guide · 2026

  1. Claude Sonnet 5Anthropic
  2. Claude Opus 5Anthropic
  3. GPT-6 AstraOpenAI
  4. GPT-6 Sol + LunaOpenAI
  5. Claude Opus 5.5Anthropic
  6. Claude Sonnet 5.5Anthropic
  7. Guide updatedRepoSpend

Published Standard short-context API rates

USD per 1M tokens · as of September 28, 2026

Current models: USD per million input and output tokens, Standard short-context API rates as of September 28, 2026
ModelInput / 1MOutput / 1M
GPT-6 Astra$10$50
GPT-6 Sol$2$10
GPT-6 Luna$0.10$0.50
Claude Opus 5.5$4$20
Claude Sonnet 5.5New · Sep 28$2$10
GPT-6 Astra
Input / 1M
$10
Output / 1M
$50
GPT-6 Sol
Input / 1M
$2
Output / 1M
$10
GPT-6 Luna
Input / 1M
$0.10
Output / 1M
$0.50
Claude Opus 5.5
Input / 1M
$4
Output / 1M
$20
Claude Sonnet 5.5New · Sep 28
Input / 1M
$2
Output / 1M
$10

Selected prior reference rates

Earlier models for comparison

Prior reference models: USD per million input and output tokens, Standard short-context API rates as of September 28, 2026
ModelInput / 1MOutput / 1M
Claude Opus 5$5$25
Claude Sonnet 5$2$10
GPT-5.6 Sol$4$20
Claude Opus 5
Input / 1M
$5
Output / 1M
$25
Claude Sonnet 5
Input / 1M
$2
Output / 1M
$10
GPT-5.6 Sol
Input / 1M
$4
Output / 1M
$20

GPT-5.6 Sol uses its current promotional rate, available at least through November 21, 2026. These are reference models, not a record of every historical price.

Caching, longer context, processing mode, regional pricing, and subscriptions can change real cost. Sources: OpenAI pricing, Anthropic pricing, and GPT-5.6 Sol.

What actually got cheaper?

Two token-price reductions—and a different kind of efficiency claim.

Claude Opus 5 → Opus 5.5

$5 / $25 → $4 / $20

20% lower token price

Both Standard input and output rates fell by one fifth.

Anthropic announcement

Claude Sonnet 5 → Sonnet 5.5

$2 / $10 → $2 / $10

Same token price

Anthropic reports up to 30% lower task cost and over 30% faster output in its testing. Workload and effort level matter.

Anthropic’s testing claim

GPT-5.6 Sol → GPT-6 Sol

$4 / $20 → $2 / $10

50% lower token price

Compared with GPT-5.6 Sol’s current promotional Standard rate.

OpenAI announcement

How wide is the GPT-6 price range?

Luna’s input and output rates are 99% below Astra’s. This compares model tiers; it is not a historical price cut or a claim of equal capability.

Astra$10 / $50input / output per 1M
Sol$2 / $10input / output per 1M
Luna$0.10 / $0.50input / output per 1M

Price per token is not cost per successful coding task

The useful comparison includes every attempt needed to finish the work.

Sonnet 5 → Sonnet 5.5 makes the distinction clear. The $2 / $10 token price stayed the same; Anthropic attributes its task-cost improvement to needing fewer tokens and steps. Vendor testing is context, not a guarantee for your repository.

  1. Count every attempt

    Retries, failed runs, and repeated context all consume tokens.

  2. Include thinking & tools

    Effort settings and tool loops can change the total work per task.

  3. Account for caching

    Cache hits and cache writes carry different rates and TTLs.

  4. Verify the outcome

    Use your tests and review to decide whether the task actually succeeded.

Task comparison = total API-equivalent cost across attempts ÷ independently verified successful tasks

RepoSpend shows local usage and API-equivalent estimates. It does not benchmark success, judge model quality, or measure your subscription bill.

Sources & methodology

Prices are a dated editorial snapshot of published Standard short-context API rates. Percentage reductions compare both input and output rates. No benchmark winner or provider billing promise is implied.

Recheck provider documentation before making a purchase or changing a budget; promotions and rates can change.