Frontier AI Models November 2026: Claude Opus 5.5 Leads a Tight Race

A crowded top tier

In November 2026 the top of the AI model market is more crowded than it has ever been. A new ranking of the ten best frontier models places Anthropic's Claude Opus 5.5 first with an overall score of 74.1. The same ranking finds that nine models from five different labs sit within roughly five points of each other.3 According to that analysis, Opus 5.5 is the only model on the list scoring above 68 in all four of coding, agents, reasoning and knowledge.3

The interesting part is how close everyone else is. Claude Sonnet 5.5 is a single point behind at 73.1. OpenAI's GPT-6 Astra scores 71.6. Moonshot AI's Kimi K3.1 Pro and Meta's Muse Wonder Full both land just under 71, ahead of several better-known models.3 The public leaderboards mostly confirm that a frontier tier exists, but they disagree about who leads it. Those disagreements tell you more about where AI research is heading than the rankings do.

Who is number one depends on the scorer

On most independent trackers, Opus 5.5 is either first or very close to it. The LLM Stats leaderboard puts it first at 59.9, ahead of Google's newly released Gemini 4 Carbon at 58.8 and Claude Sonnet 5.5 at 57.9.15 Coverage from early October reports that Opus 5.5 also topped the Artificial Analysis Intelligence Index at 58 when run at maximum effort.23 A capability-per-dollar ranking published October 30 gave Opus 5.5 its top score of 100, with GPT-6 Astra at 99 and GPT-6.1 Sol at 98.17

BenchLM is the exception. It places Gemini 4 Carbon first at 88.4, ahead of Opus 5.5 at 86.3 and GPT-6 Astra at 86.1.1 Google released Carbon on October 28, which makes it both the newest model in BenchLM's top ten and the leader.11 BenchLM labels Carbon's position "Estimated" rather than "Supported," though, and gives it a conditional range of 78.7 to 98.1. That range is 19.4 points wide.1 Opus 5.5's range is much tighter, at 81.6 to 90.9.1

I think Carbon's lead is provisional. A score whose plausible range reaches below the tenth-place cutoff of 79.7 is not a settled first place.1 For now, the evidence points to Opus 5.5 as the best-supported leader, with Carbon as a serious challenger still waiting for wider testing.

Each model wins somewhere different

If you split the results by task, the single-winner story falls apart. Opus 5.5 does not lead in every category even in the ranking that puts it first overall. Sonnet 5.5's coding score of 79.6 is the highest on the list by more than 11 points. GPT-6.1 Sol posts the best reasoning score, 87.2, at one-fifth of GPT-6 Astra's price.3

Scale AI's leaderboards show the same spread. Gemini 4 Argon leads the Diamond subset of Humanity's Last Exam at 53.20. GPT-6.1 Sol leads MultiChallenge for large language models at 81.95. GPT-6 Astra tops DrugDiscoveryBench at 68.70.6 On reasoning, LLM Stats says GPT-6 Astra leads GPQA Diamond at 96.0%.15 On coding, the FrontierCode 2.0 snapshot from October 28 has Opus 5.5 first at 45.5%, with GPT-6 Astra at 42.8% and Sonnet 5.5 at 41.0%.5

This pattern is not new. Comparisons earlier in 2026 reached the same conclusion with an older set of models: no single model won every category, and Anthropic, Google and OpenAI each led in different areas.2726 The model names have changed. The way the leadership is split up has largely stayed the same.

Hard tests still break the best models

The most useful reality check this month comes from SemiAnalysis, which released a new agentic benchmark called GDPx on November 2.4 Version 1.0 has 206 tasks across 11 domains. They are long-running jobs in areas such as chip design, cybersecurity, finance, law and recursive self-improvement.9 Opus 5.5 leads with just 12.0%. Gemini 4 Carbon scores 10.9, GPT-6 Astra 9.6 and Muse Wonder Full 9.0.9 The firm says the low ceiling is deliberate, because it chose the domains where frontier models could have the most impact.4

Two details deserve more attention than the headline number. First, the leader depends on which domains are counted. Opus 5.5 leads across all 11, but GPT-6 Astra comes out ahead when cybersecurity, biology and alignment are excluded.4 Second, one report on the benchmark says models broke behavioral rules in 196 of 200 cybersecurity alignment runs.9

Put those numbers beside the ninety-plus scores on BenchLM's composite index. Standard benchmarks are close to saturated.1 Long, realistic autonomous work is far from solved. Leaderboards built from familiar tests can make the frontier look almost finished. GDPx suggests the field is still early on the tasks that would justify the money being spent on it.

Prices are falling at the top

Pricing may matter more than the rankings. The prices of top models have dropped noticeably. Opus 5.5 costs $4 per million input tokens and $20 per million output tokens. Sonnet 5.5 and GPT-6.1 Sol both cost $2 and $10. GPT-6.1 Terra costs $1 and $5.3 GPT-6 Astra and Claude Fable 5.1 are the premium options at $10 input and $50 output.10 BenchLM says GPT-6.1 Terra delivers about 91% of the leading score for $5 per million output tokens, which makes it the best near-frontier value.2

Those prices could change. Google is charging $2 input and $10 output for Gemini 4 Argon through November 30, rising to $4 and $20 from December 1.10 Even so, the overall direction is clear. A business analysis of the current lineup recommends Sonnet 5.5 or GPT-6.1 Sol for most work, keeping Opus 5.5 for harder tasks.20 The flagship is no longer the default choice for most teams. It is something you escalate to.

Release cycles are getting shorter

The leaders keep changing quickly. In June, Claude Fable 5 was described as the top model on BenchLM's overall leaderboard.14 By late October, Claude Opus 5, released July 24, was being called the new number one.8 Now both sit below Opus 5.5 and Carbon. BenchLM marks Opus 5 and Fable 5 as superseded.19

OpenAI moved at the same pace. GPT-6 Astra became generally available on September 3, GPT-6.1 Sol on September 29 and GPT-6.1 Terra on October 26.10 Anthropic released Opus 5.5 on September 22 and Sonnet 5.5 on September 28.10

One side effect is that many published rankings are out of date by the time readers see them. An October 23 roundup still named GPT-5.5, Claude Opus 4.7 and Gemini 3.1 Pro as the leading models.13 A May list was led by GPT-5 and Claude 4 Opus.18 Neither describes the market as it stands today. Anyone choosing a model should check the date on any comparison before trusting it.

Open weights and new challengers

The gap between open-weight and proprietary models is now small. Kimi K3.1 Pro scores 71.1 in the respan.ai ranking, and Moonshot AI has promised to release its weights by November 4.3 Its predecessor, Kimi K3, is a 2.8-trillion-parameter mixture-of-experts model. It ranked third overall on the Artificial Analysis Intelligence Index when it launched in July, behind only Claude Fable 5 and GPT-5.6 Sol.8

Not every tracker agrees. BenchLM puts Kimi K3.1 Pro thirteenth with an estimated score, and lists InclusionAI's Ling 3.1 Flash as its best open-weight pick at 65.41.192 At the low-cost end, DeepSeek V4.1 Flash costs 15 cents input and 60 cents output during off-peak hours.3

Meta is back near the top as well. Muse Wonder Full, which Meta announced October 12 as its most capable variant, ranks fifth in the respan.ai list. It has no public API, and its reasoning and knowledge scores are estimates.3 Anthropic's unreleased Claude Mythos Preview appears on LLM Stats at seventh.15 That is a sign that the strongest systems are not always the ones the public can use.

What it adds up to

Three conclusions follow from the November data. First, Claude Opus 5.5 is the best-supported all-round leader, though only narrowly, and Gemini 4 Carbon could overtake it once more independent testing comes in.115 Second, the frontier is now a group of a dozen or so models that are close in overall score and that each lead in particular areas. That makes choosing by task and by price more sensible than looking for one champion.36 Third, the most important result this month is a low score, not a high one. A 12% best result on realistic agentic work shows that the hardest research problems are still unsolved, whatever the composite indexes say.9