GPT-6 Sol/Luna Launch: API Prices Have Been Halved

On September 22, 2026, OpenAI released GPT-6 Sol and GPT-6 Luna at the same time. Input and output token prices dropped 50% versus GPT-5.6, and this is a permanent price, not a promotional one. That same day, Anthropic released Claude Opus 5.5, and the AI model market went through an earthquake of price and performance in a single day.

1. GPT-6 Lineup and Pricing

As of September 2026, GPT-6 consists of three models.

Price Change vs GPT-5.6

Note: GPT-5.6 Sol's $4/$20 is a promotional price, guaranteed "at least" until November 21, 2026. GPT-6 Sol's $2/$10 is a permanent price.

Additional Pricing Rules

• Long context (>272K input): 2x input and cache rates and 1.5x output rate applied to the entire request
• Batch/Flex mode: 50% of the standard rate
• Fast mode: 2x the applied rate
• Cache write: 1.25x the input rate

2. Key Benchmark Comparison

OpenAI Official Figures

AutomationBench 1.0.6 (work automation)

It evaluates end-to-end workflows for sales, marketing, operations, support, finance, and HR using 47 tools.

At xhigh effort, GPT-6 Sol scores 6.3 points higher than Claude Opus 5 max while costing 91% less per task.

DeepSWE v1.1 (real-world coding)

It evaluates long-horizon software engineering tasks in real codebases.

GPT-6 Luna delivers coding performance close to Claude Opus 5 while costing 93% less per task.

OSWorld 2.0 (computer control)

It evaluates long-horizon computer-use workflows spanning everyday and professional tasks.

Independent Evaluator Figures (Artificial Analysis)

Independent evaluations differ slightly from OpenAI's official figures.

Hallucination Rate Change

Key trap: Sol's hallucination improvement is achieved not by "answering more accurately" but by "saying it does not know what it does not know." GPT-5.6 Sol answered 99.9% of questions, but GPT-6 Sol answers only 83%. This has the side effect of dropping accuracy from 59% to 54%.

Coding Agent Index

On the Artificial Analysis Coding Agent Index, GPT-6 Sol (max) improved 2 points over its predecessor while halving the cost per task. GPT-6 Luna (max), by contrast, regressed 2 points.

3. GPT-6 Astra: The Flagship Standard

GPT-6 Astra is the top model of the GPT-6 family, launched September 3.

Key Benchmarks

Pricing

• Input: $10.00/1M, cached input: $1.00/1M, output: $50.00/1M
• Above 272K tokens: entire request priced at 2x (input, cache) and 1.5x (output)
• Batch/Flex: 50% off, Fast mode: 2x

ARC-AGI-3 Caveat

The 99.9% on ARC-AGI-3 is a figure using the Provider Adapter harness; on the Standard harness it is only 62.7%. When reading benchmark figures, always check the harness conditions.

4. Comparison with Claude Opus 5.5

About 90 minutes before the GPT-6 Sol/Luna launch, Anthropic released Claude Opus 5.5.

Price Comparison

At the same price point as Claude Sonnet 5, GPT-6 Sol competes at half the price of Claude Opus 5.5.

Performance Comparison (on the same harness)

Because OpenAI and Anthropic compared their models on different harnesses, a same-harness comparison is needed.

• AutomationBench: GPT-6 Sol (xhigh) 33.2% vs Claude Opus 5 (max) 26.9% -> Sol ahead
• DeepSWE v1.1: GPT-6 Sol (max) 68.8% vs Claude Fable 5 (xhigh) 69.9% -> Fable 5 slightly ahead
• OSWorld 2.0: GPT-6 Astra 72.6% vs Claude Opus 5 70.6% -> Astra ahead

5. User Feedback and Real-World Experience

Positive Feedback

• Tone improvement: Astra's collaborative style carried over to Sol/Luna, cutting verbose explanations and unnecessary conditionals
• Fewer hallucinations: The rate of honestly saying "I do not know" when it does not know went up
• Cost revolution: Luna in particular, at 10 cents per 1M input tokens, lowered the barrier to large-scale batch work

Negative Feedback

• Slower time to first token (TTFT): Output speed rose 44% thanks to the advanced reasoning computation structure, but the first-token wait time grew noticeably
• Benchmark competition: In some evaluations, Claude Opus 5.5 (AutomationBench 40.0%) exceeded GPT-6 Sol's peak score
• The shadow of the hallucination improvement: Reducing hallucination by refusing to answer comes with an accuracy drop

6. Real-World Deployment Guide

When to Use GPT-6 Astra

• Computer environment control automation (browser, spreadsheets, CRM, etc.)
• High-difficulty scientific research and mathematical proofs
• Cybersecurity penetration testing and vulnerability analysis
• Projects where "you need the best result and cost is not the issue"

When to Use GPT-6 Sol

• Complex multi-file debugging and refactoring
• Verifying API-guideline compliance
• Sophisticated agent automation (best on AutomationBench)
• Work where "strict alignment between what is verified and what is not" matters
• Cases where you need performance similar to Claude Opus 5.5 at half the cost

When to Use GPT-6 Luna

• Large-volume text classification and information extraction
• Repetitive tagging and labeling work
• Batch-processing tens of thousands of items
• A slave-agent backend running in an infinite loop with no cost risk
• Large-scale distributed processing where "even max effort carries no cost burden"

Cost Simulation

For 100M input tokens + 20M output tokens per month (no cache):

With an 80% cache hit, the Sol cost is $256 (output tokens account for most of the cost).

7. Cautions and Limits

1. The 272K long-context cliff: If input tokens exceed 272K by even one, the whole request is re-priced at 2x/1.5x. For Sol, 270K input costs $0.64, 275K input costs $1.25 — nearly 2x.
2. The hallucination trap: Sol's hallucination improvement relies on refusing to answer. "Saying it does not know" is both a strength and a cause of lower accuracy.
3. TTFT slowdown: More reasoning tokens increase the first-response wait time. For latency-sensitive real-time applications, a reasoning effort: none setting is needed.
4. GPT-5.6 Sol price risk: The current $4/$20 promotional price may rise after November 21. Migrating to GPT-6 Sol is favorable in the long run.
5. Benchmark harness differences: Because OpenAI and Anthropic compared their models on different harnesses, always check same-harness figures for a direct comparison.

8. GPT-5.6 -> GPT-6 Migration Checklist

1. Check the model ID: gpt-5.6-sol is not gpt-6-sol. If a benchmark table says only "Sol," always check the model ID.
2. Compare prices: GPT-5.6 Sol's promotional price ($4/$20) may change after November 21. GPT-6 Sol ($2/$10) is a permanent price.
3. Check the long-context cliff: Prepare for the cost surge on requests over 272K.
4. Test reasoning effort: Sol/Luna can be set to none, so use it when latency optimization is needed.
5. Adjust expectations on hallucination: Hallucination dropped, but because that comes from refusing to answer, the "rate of confident answers" did not necessarily rise.

Conclusion

The core of GPT-6 Sol and Luna is "similar performance at half the cost." If GPT-6 Astra provides the highest intelligence, Sol and Luna serve to popularize it. In particular, Luna's 10 cents per 1M tokens has made the cost of running a coding agent, once several dollars or more, effectively close to zero.

realpath In practice, the selection criteria are clear. Astra when you need the best result, Sol when you need the optimal cost-performance balance, Luna when you need large batches and slave agents. The combination of these three models currently provides the strongest cost-performance curve on the market.