In July, we named GPT-5.6 Sol the top frontier model. It held the spot for about ten weeks. In September, OpenAI shipped GPT-6 Astra on the 3rd, Google shipped Gemini 3.8 Flash on the 2nd, and on the 22nd Anthropic released Claude Opus 5.5 the same day OpenAI answered with two cheaper GPT-6 models, Sol and Luna. It was the most crowded month in AI this year, and the rankings changed.
Our verdict: Claude Opus 5.5 is the best AI model you can use right now. It's #1 on the leading independent intelligence index, it leads the hardest public coding benchmark, and it costs 40% of GPT-6 Astra's price with no surcharge for long prompts. Astra is still the specialist pick for science and computer automation. GPT-6 Sol is the value workhorse. Fable 5.1, Anthropic's previous flagship, has been undercut by its own sibling.
The Short Version
Best overall: Claude Opus 5.5. It's #1 on the Artificial Analysis index (58), scores 89.9% on SWE-bench Pro and costs $4/$20 per million tokens.
Best for science and automation: GPT-6 Astra, at $10/$50, rising to $20/$75 past 272K input tokens.
Best value workhorse: Claude Sonnet 5.5 (new Sep 28), at $2/$10. It nearly matches Opus 5.5 on agentic and office work. GPT-6 Sol is the OpenAI pick at the same price.
Cheapest model worth using: GPT-6 Luna, $0.10/$0.50. See our budget model verdict.
The Scoreboard
| Model | Released | Price per 1M tokens (in / out) | Context | AA Intelligence Index | SWE-bench Pro |
|---|---|---|---|---|---|
| Claude Opus 5.5 | Sep 22 | $4 / $20 | 1M, no surcharge | 58 (#1 of 212) | 89.9% |
| GPT-6 Astra | Sep 3 | $10 / $50 ($20 / $75 over 272K) | Surcharge past 272K | 53 | — |
| Claude Fable 5.1 | Summer | Premium tier | — | 53 | 81.2% |
| Claude Sonnet 5.5 | Sep 28 | $2 / $10 | 1M, no surcharge | Near Opus 5.5 on agentic tests | — |
| GPT-6 Sol | Sep 22 | $2 / $10 | — | Below Opus 5.5 | — |
| Gemini 3.8 Flash | Sep 2 | $0.75 / $3.75 (promo) | — | Fast tier | — |
How to read this: AI labs publish benchmark tables that rarely overlap. Only seven tests appear in both Anthropic's and OpenAI's launch materials; Opus 5.5 leads four, Astra leads two, and one isn't comparable. Independent scores also run a few points below Anthropic's own. We weight the independent Artificial Analysis index most heavily and treat vendor numbers as directional.
What the Price Gap Means in Dollars
Per-token prices are hard to picture, so here's a real workload: a team sending 50 million input tokens and 10 million output tokens a month, a busy internal coding assistant.
- Claude Opus 5.5: $200 input + $200 output = $400 a month
- GPT-6 Astra: $500 + $500 = $1,000 a month, more if any request passes 272K tokens
- Claude Sonnet 5.5 or GPT-6 Sol: $100 + $100 = $200 a month
Astra has to be substantially better at your specific task to justify 2.5 times the bill. For most coding and knowledge work, it isn't.
Claude Opus 5.5: The New Default
Opus 5.5 is built for long, multi-step work: agentic coding, research and tasks that run for hours. Its 89.9% on SWE-bench Pro is well ahead of Fable 5.1 (81.2%) and Mythos 5 (80.3%). It holds a 1M-token context window with no long-prompt penalty, and Anthropic cut Opus pricing about 20% from the previous generation.
The catch: at maximum effort, Opus 5.5 thinks for a long time and burns tokens to do it. Independent testers have seen a single hard agentic task cost around $13. Run it at default effort and save max effort for problems that need it.
Claude Sonnet 5.5: Opus-Level Work for Half the Price
Anthropic followed Opus 5.5 with Sonnet 5.5 on September 28, and for most teams it's the more important launch. At $2/$10 per million tokens it costs half as much as Opus 5.5 and exactly the same as GPT-6 Sol. It also comes close to Opus on the tests that matter for everyday work. It scores 70.6% on Terminal-Bench 4.0, ahead of Opus 5.5's 66.4%. On Artificial Analysis's GDPval-AA office-work test it lands at 1,844 Elo, two points behind Opus and far ahead of GPT-6 Sol's 1,487. It's also more than 30% faster than Sonnet 5, and Anthropic says it uses fewer tokens, cutting cost per task by up to 30%.
Where Opus still wins: the hardest long-running coding and research jobs. Opus leads on FrontierCode (54.4% vs 52.1%) and holds the top overall index score. The practical setup is Sonnet 5.5 as the default, with Opus 5.5 for the problems Sonnet can't crack.
GPT-6 Astra: Still the Specialist
Astra came first and aimed at coding, research, computer use and professional multi-step work. It still leads on science-heavy reasoning and automating software through a screen. But at $10/$50, and with the whole request repriced to $20/$75 once input passes 272,000 tokens, it's the most expensive way to get frontier quality. That surcharge alone rules it out for long-document work.
GPT-6 Sol: OpenAI's Most Important Launch
Sol matters more to most OpenAI customers than Astra does. At $2/$10, it roughly halves the cost of the GPT-5.6 generation. It trails Opus 5.5 in independent testing, but it's the obvious default for everyday production work on OpenAI. If you're still paying GPT-5.6 prices, switch today.
Fable 5.1 and Gemini
Fable 5.1 is still a beautiful writer, and some teams will keep it for brand voice and long-form drafting. As a default model, though, Opus 5.5 now matches or beats it for less. Gemini 3.8 Flash isn't competing at this tier; it's Google's fast, cheap model and one of the better budget options, covered in our cheapest AI model verdict.
Which Model for Which Job
- Agentic coding and hard engineering: Claude Opus 5.5 at default effort.
- Scientific research and computer-use automation: test GPT-6 Astra, and keep prompts under 272K tokens.
- Everyday production work, bug fixes, documents and spreadsheets: Claude Sonnet 5.5, or GPT-6 Sol if you're on OpenAI.
- High-volume extraction, tagging and chat: GPT-6 Luna or Gemini 3.8 Flash.
- Data that can't leave your servers: the open-weight models in our non-cloud LLM guide.
The Verdict
Claude Opus 5.5 is the best AI model of September 2026. It has the top independent score, the top coding score and a price 60% below its closest rival. GPT-6 Astra earns its premium only for science and automation. Claude Sonnet 5.5 is the best value for everyday work, and GPT-6 Sol is the smart default for OpenAI shops. This table won't last long, so route by task, not by brand loyalty.
FAQ
What is the best AI model in September 2026?
Claude Opus 5.5. Artificial Analysis ranks it #1 of 212 models on its Intelligence Index with a score of 58, five points ahead of GPT-6 Astra and Claude Fable 5.1, and it leads SWE-bench Pro at 89.9%. GPT-6 Astra still wins some science and automation tasks.
How much does Claude Opus 5.5 cost?
$4 per million input tokens and $20 per million output tokens, with a 1M-token context window and no long-context surcharge. That's 40% of GPT-6 Astra's $10/$50 list price.
Is GPT-6 Astra worth the higher price?
For most teams, no. Astra costs $10/$50 per million tokens and $20/$75 once a request passes 272,000 input tokens. It's worth testing for science-heavy research and computer-use automation, but Opus 5.5 is cheaper and ahead on most shared benchmarks.
What happened to Claude Fable 5?
It was updated to Fable 5.1, which scores 81.2% on SWE-bench Pro and 53 on the Artificial Analysis index. Claude Opus 5.5 now matches or beats it on most tasks at a lower price, so most teams should default to Opus 5.5.
Is Claude Sonnet 5.5 better than Opus 5.5?
For most everyday work it's nearly as good for half the price. Sonnet 5.5, released September 28, 2026 at $2/$10 per million tokens, beats Opus 5.5 on Terminal-Bench 4.0 (70.6% vs 66.4%) and nearly ties it on office-work tests. Opus 5.5 still leads on the hardest coding and on the overall intelligence index.
What is GPT-6 Sol?
OpenAI's lower-cost GPT-6 model, released September 22 at $2 per million input tokens and $10 per million output tokens. It roughly halves the cost of the GPT-5.6 generation, though independent tests place it behind Opus 5.5.