GPT-6 vs Claude 5.5 API Pricing: Astra, Sol, Luna vs Opus, Sonnet, Fable
Astra matches Fable on price. GPT-6.1 Sol, GPT-6 Sol, and Sonnet 5.5 are all $2/$10, but 6.1 Sol's cache reads cost half as much.
Two model families launched within a month of each other: OpenAI's GPT-6 (Astra on September 3, Sol and Luna on September 22, GPT-6.1 Sol on September 29) and Anthropic's Claude 5.5 (Opus 5.5 on September 22, Sonnet 5.5 on September 28). If you're picking one for an API workload, price is the first filter — so here is every rate side by side.
Quick answer: price per 1M tokens
| Tier | Model | Input | Output | Cache read | Context |
|---|---|---|---|---|---|
| Flagship | GPT-6 Astra | $10 | $50 | $1.00 | 1.05M |
| Flagship | Claude Fable 5.1 | $10 | $50 | $0.25 | 1M |
| Upper mid | Claude Opus 5.5 | $4 | $20 | $0.20 | 1M |
| Mid | GPT-6.1 Sol | $2 | $10 | $0.10 | 1.05M |
| Mid | GPT-6 Sol | $2 | $10 | $0.20 | 1.05M |
| Mid | Claude Sonnet 5.5 | $2 | $10 | $0.20 | 1M |
| Small | GPT-6 Luna | $0.10 | $0.50 | $0.01 | 1.05M |
| Small | Claude Haiku 5.5 | not released yet |
Standard API rates, text tokens. OpenAI rates apply to prompts up to 272K input tokens; above that OpenAI doubles input and cache rates and charges 1.5x for output on the whole request. Both vendors give 50% off for batch processing.
Three things stand out:
- The sticker prices still line up. Astra and Fable 5.1 are both $10/$50. GPT-6.1 Sol, GPT-6 Sol, and Sonnet 5.5 are all $2/$10. Cache reads break that tie: $0.10 on GPT-6.1 Sol, $0.20 on GPT-6 Sol and Sonnet 5.5.
- Opus 5.5 is the only model in between. At $4/$20 it costs twice Sonnet 5.5 and 40% of Fable 5.1, and Anthropic says it performs near Fable 5.1 on most work.
- Only OpenAI has a small tier today. Luna is 20x cheaper than the $2/$10 models. Anthropic's Haiku 5.5 is announced for "the coming weeks"; until then its cheapest current option is Haiku 4.5 at $1/$5, ten times Luna.
What a request actually costs
Per-million rates are hard to feel. Take one request with 20,000 input tokens and 5,000 output tokens, no caching:
| Model | Cost per request | Cost per 10,000 requests |
|---|---|---|
| GPT-6 Luna | $0.0045 | $45 |
| GPT-6.1 Sol | $0.09 | $900 |
| GPT-6 Sol | $0.09 | $900 |
| Claude Sonnet 5.5 | $0.09 | $900 |
| Claude Opus 5.5 | $0.18 | $1,800 |
| GPT-6 Astra | $0.45 | $4,500 |
| Claude Fable 5.1 | $0.45 | $4,500 |
Without caching, GPT-6.1 Sol costs the same as GPT-6 Sol and Sonnet 5.5. Astra and Fable 5.1 match each other too. Luna is 20x cheaper than the $2/$10 models.
Cache is where GPT-6.1 Sol separates from the other two. Same kind of agent turn, but 200,000 input tokens are cache reads, 10,000 are fresh, and the reply is 1,000 tokens:
| Model | Cost per request | Cost per 10,000 requests |
|---|---|---|
| GPT-6 Luna | $0.0035 | $35 |
| GPT-6.1 Sol | $0.05 | $500 |
| GPT-6 Sol | $0.07 | $700 |
| Claude Sonnet 5.5 | $0.07 | $700 |
| Claude Opus 5.5 | $0.10 | $1,000 |
| Claude Fable 5.1 | $0.20 | $2,000 |
| GPT-6 Astra | $0.35 | $3,500 |
On this shape GPT-6.1 Sol is about 30% cheaper than GPT-6 Sol and Sonnet 5.5, because its cache read rate is half of theirs. Output-heavy requests shrink that gap back toward zero.
Price per token is not price per task
Token price only tells you the cost if both models use the same number of tokens. They don't.
- Anthropic says Sonnet 5.5 costs up to 30% less per task than Sonnet 5 at the same per-token price, because it needs fewer tokens, and that at High effort it matches GPT-6 Sol's best FrontierCode score at about a fifth of the cost per task. Those are Anthropic's numbers.
- Artificial Analysis found the opposite at the extreme: at max effort Sonnet 5.5 wrote about 193,000 tokens per task, the most it has measured, roughly $7.60 per task. Effort setting matters more than the price sheet.
- OpenAI says GPT-6.1 Sol, at the same $2/$10 rate as GPT-6 Sol, matches Astra on DeepSWE v1.1 at about one-fifth the cost per task, and lands 2.1 points behind Astra on OSWorld 2.0 at about one-seventh the cost per task. Those are OpenAI's numbers.
- OpenAI also makes the mirror argument for Astra: at 2.5x the price of the previous generation, it says the higher per-token rate can still come out cheaper if it finishes a hard task in fewer rounds.
Both sides' claims come from the vendor, so treat them as direction, not proof. The practical rule: run your own task at two effort levels on two models before committing.
Which one to pick
- High-volume, simple steps (tagging, extraction, routing, summaries): GPT-6 Luna. There is no Claude equivalent at this price right now.
- Everyday coding and agent work: GPT-6.1 Sol or Claude Sonnet 5.5. The sticker price matches GPT-6 Sol at $2/$10, and GPT-6.1 Sol's cache reads are half as much. Keep GPT-6 Sol only if GPT-6.1 Sol is not on the endpoint you use.
- Hard, long-running work where a wrong answer is expensive: Opus 5.5 first, since it costs 60% less than Fable 5.1 or Astra. Step up to Fable 5.1 or Astra only if Opus falls short on your task.
- Cost-sensitive but not tiny: GPT-6.1 Sol or Sonnet 5.5 at lower effort settings before you consider a bigger model.
Sources
Prices come from Anthropic's Claude Sonnet 5.5 announcement and model docs, OpenAI's GPT-6 and GPT-6.1 Sol model listings, and API pricing write-ups from Respan, TreeRouter, and TrueFoundry. Cost per request is calculated from those rates. Prices change; check the vendor pricing page before you budget.