Anthropic's Opus 5.5 Tops Artificial Analysis as OpenAI Halves Prices With GPT-6 Sol

TL;DR

On September 22 both frontier labs shipped on the same afternoon. Claude Opus 5.5 costs $4/$20 per million tokens, 20% below Opus 5, and took the top spot on the Artificial Analysis Intelligence Index at 58, five points clear of GPT-6 Astra. Roughly 90 minutes later, by TechCrunch's count, OpenAI released GPT-6 Sol and GPT-6 Luna at half the price of their 5.6 predecessors. Sol scores 48 on the same index, but a task costs about a fifth of what Opus 5.5 spends. One lab went for the crown, the other went for your invoice.

What Anthropic shipped

Opus 5.5 is the first model in Anthropic's 5.5 family, and the company's pitch is that it does Fable-level work at Opus prices. The launch post puts it this way: at default settings it "will cost 40% less than Opus 5 on typical workloads," and it "generates output more than 30% faster than Opus 5."

The per-token price cut is smaller than the headline 40%. Input and output drop 20% and cache reads drop 60%; the 40% figure is Anthropic's own measurement of typical workloads at default settings.

• Input / output: $4 and $20 per million tokens, down from $5 and $25.
• Cache reads: $0.20 per million, 60% below Opus 5. Cache writes are $5.
• Fast mode: $8 input and $40 output, for up to 2.5x the speed.
• Context: 1 million tokens, unchanged from Opus 5, per OfficeChai.

On Anthropic's own table, Opus 5.5 posts 66.4% on Terminal-Bench 4.0 against 55.8% for Fable 5.1 and 57.9% for GPT-6 Astra, and 1,846 Elo on GDPval-AA v2.1 against 1,542 for Astra. Astra keeps a narrow lead on AutomationBench, 41.4% to 40.0%. Those are vendor numbers run at max effort, so read them as the ceiling, not the median.

The subscription change

Some coverage said Anthropic scrapped five-hour caps. It did not. The primary source says Anthropic is "increasing five-hour usage limits on Pro, Max, Team, and seat-based Enterprise plans," and it gave subscribers a saved rate limit reset they can spend whenever they like. MacRumors reports that reset is usable until October 22. Bigger limits and a reset button are welcome. A missing limit would be a different story.

Anthropic says METR and Frontier Design evaluated the model before launch. It also says Claude Sonnet 5.5 and Haiku 5.5 "will follow in the coming weeks."

The independent scoreboard

Artificial Analysis ran both launches through its suite within the day, which is the closest thing to a neutral referee this week. Opus 5.5 at max effort lands at 58, first of 212 models tracked. It leads six of the ten underlying evals, including 61.4% on Humanity's Last Exam and 66.9% on SciCode. In AA's own run, it ties GPT-6 Astra on Terminal-Bench 4.0 at 59.6%, below the 66.4% in Anthropic's table.

There is a catch, and it is the same catch every Opus has had. AA measured about 119,000 output tokens per task for Opus 5.5 at max effort, against roughly 73,000 for Opus 5, 78,000 for Fable 5.1 and 27,000 for GPT-6 Astra. The per-token price fell and the tokens per answer went up. It is a cheaper taxi that takes the scenic route.

AA's four highest Opus 5.5 effort settings (max, xhigh, high and medium) all sit on its intelligence-versus-cost Pareto frontier. The low setting drops to 42. If you run Opus in production, the effort knob now matters about as much as the model name.

What OpenAI shipped

OpenAI did not try to beat Opus 5.5 on the leaderboard. Sol and Luna "build on the advances behind GPT-6 Astra," says OpenAI's developer forum announcement, and the story is price. The company attributes the cut to caching and inference improvements.

• GPT-6 Sol: $2 input, $0.20 cached, $10 output per million tokens, down from $4/$20 for GPT-5.6 Sol. Knowledge cutoff April 20, 2026.
• GPT-6 Luna: $0.10 input, $0.01 cached, $0.50 output, down from $0.20/$1.20. Knowledge cutoff May 18, 2026.
• Both: a 1,050,000-token context window with 922,000 tokens of max input, 128,000 max output, and six reasoning efforts from none to max. Cache writes bill at 1.25x the input rate, and Batch and Flex run at half price.

In ChatGPT, both models roll out to Plus, Pro, Business, Enterprise and Edu in ChatGPT Work and Codex, and Free and Go users can try Luna in the desktop app. On GitHub Copilot, Sol is on Pro+, Max, Business and Enterprise, and Luna adds the Pro plan.

Same score, smaller bill, fewer made-up answers

AA's write-up reports both Intelligence Index scores unchanged from GPT-5.6. What changed is cost: Sol at max effort costs $1.06 per index task, down from $1.99, and Luna costs $0.07, down from $0.18. On AA's Coding Agent Index, Sol gains two points to 57 at $2.99 per task, while Luna slips two points to 41.

The biggest improvement is in hallucination. On AA-Omniscience, Sol's hallucination rate falls from 92% to 60% and Luna's from 93% to 77%. That fits OpenAI's claim, relayed by TechCrunch, that Sol makes about half as many mistakes as its predecessor. The trade is knowledge work: AA measured Sol about 100 Elo lower on GDPval-AA v2.1 and Luna about 75 lower.

What to do with this

The two launches do not compete for the same job, which is why the same-day timing matters less than it looks.

• Hardest agentic work: Opus 5.5 is now the top-scoring model you can buy, and it costs less per token than Opus 5. Budget for the token usage, not the list price. At 119k output tokens per task, a $20 output rate works out to more per task than you might expect.
• High-volume coding and tool loops: Sol at $2/$10 with a 1.05M window and a large hallucination drop is the new default to test. Paying 5.6x for 10 index points is sometimes worth it and often not.
• Extraction, classification, summarization: Luna at $0.10/$0.50 is priced like a rounding error. Watch its verbosity: AA counted 51k output tokens per task, up from 41k.
• Anthropic subscribers: spend the saved rate limit reset before October 22 and check your new five-hour ceiling instead of assuming it is gone.

Anthropic's own claim is that Opus 5.5 at default effort beats GPT-6 Astra at max effort for about a fifth of the cost per task. OpenAI's reply is a cheaper model that does not need that argument. Run your own evals on your own tasks at the effort level you will ship with. Both vendors' charts assume max effort, and your finance team does not.

Key Takeaways

• Claude Opus 5.5 costs $4/$20 per million tokens and scores 58 on the Artificial Analysis Intelligence Index, five points ahead of GPT-6 Astra and Fable 5.1.
• Opus 5.5 uses about 119k output tokens per task at max effort, so its cost per task on AA's index is $5.98.
• GPT-6 Sol ($2/$10) and Luna ($0.10/$0.50) halve the 5.6 prices, with index scores unchanged at 48 and 37.
• Sol's hallucination rate on AA-Omniscience drops from 92% to 60%, though it loses about 100 Elo on GDPval-AA.
• Anthropic raised five-hour limits on Pro, Max, Team and seat-based Enterprise instead of removing them, and added a saved reset.
• Anthropic says Sonnet 5.5 and Haiku 5.5 are due in the coming weeks, which is when the cheap tiers get their head-to-head.

Sources: Anthropic: Claude Opus 5.5, OpenAI: Introducing GPT-6 Sol and Luna, OpenAI API: GPT-6 Sol, OpenAI API: GPT-6 Luna, OpenAI Developer Community announcement, Artificial Analysis: Opus 5.5, Artificial Analysis: GPT-6 Sol and Luna, Artificial Analysis: Opus 5.5 model page, Artificial Analysis: GPT-6 Sol model page, GitHub Changelog, TechCrunch: GPT-6 Sol and Luna, TechCrunch: Opus 5.5, MacRumors, OfficeChai