Which frontier AI labs are actually leading the model race?

Is there even one AI model race anymore?

There is no single AI model race anymore: Anthropic currently leads the strongest independent general-intelligence ranking, Google wins several important agent and multimodal tests, and OpenAI still leads parts of science, computer use and large-scale deployment.

That split has become hard to ignore. Artificial Analysis currently gives Claude Opus 5.5 at maximum effort an Intelligence Index score of 58. Gemini 4 Argon and GPT-6 Astra both score 53, while GPT-6.1 Sol reaches 52. Yet Argon scores 78% on Artificial Analysis' AutomationBench-AA, ahead of Opus at 70% and Astra at 68%. OpenAI then flips the picture on other hard tasks: Astra reaches 31% on GDP.pdf, compared with 26% for Opus and 22% for Argon.

Stanford's 2026 AI Index had already picked up the same pattern earlier in the year. Its March snapshot put Anthropic at 1,503 Arena Elo, xAI at 1,495, Google at 1,494 and OpenAI at 1,481. Only 25 points separated four companies. Stanford also found that Humanity's Last Exam performance had jumped about 30 percentage points in a year, which tells us how quickly even deliberately difficult benchmarks are losing their shelf life.

The release cycle is shortening too. OpenAI followed GPT-6 Astra with cheaper GPT-6 models and then GPT-6.1 Sol. Anthropic moved through several Claude 5 generations before Opus and Sonnet 5.5. Google returned to the top group with Gemini 4 Argon. Meta released several Muse Spark revisions within months. A lab can therefore fall two places on a leaderboard while its own model remains unchanged.

For this article, “leading the model race” has to mean more than holding first place on one benchmark. We care about who is repeatedly producing frontier models, where those models are genuinely ahead, and how widely that advantage extends across useful tasks.

Which AI model is the smartest right now?

Claude Opus 5.5 is currently the strongest general-purpose frontier model we can identify from independent testing, and Anthropic's lead is large enough to take seriously.

Artificial Analysis gives maximum-effort Opus 5.5 an Intelligence Index score of 58. Claude Sonnet 5.5 reaches 56, while Gemini 4 Argon and GPT-6 Astra sit at 53 and GPT-6.1 Sol at 52. Having two Anthropic models above every tested OpenAI and Google model is more convincing than one lucky benchmark result.

The underlying tests point in the same direction. Opus scores 67% on SciCode versus 62% for Argon and 56% for Astra. Humanity's Last Exam comes out at 61% for Opus, 57% for Argon and 55% for Astra. On AA-Briefcase, which focuses on economically useful professional work, Opus scores 1,807 against Astra's 1,569 and Argon's roughly 1,490. It also leads the three on Artificial Analysis' long-context evaluation.

Google still has a major exception. Argon's 78% on AutomationBench-AA is well ahead of Opus at 70%. OpenAI also beats Opus on GDP.pdf, a document-heavy professional benchmark. The current Claude lead is broad, but a company automating workflows or working heavily with complex documents could rationally pick another model.

There is also an important cost wrinkle. Artificial Analysis lists Opus 5.5 at $4 per million input tokens and $20 per million output tokens. Argon costs $2 and $10 under its current pricing, while GPT-6 Astra costs $10 and $50. So Opus leads on raw intelligence without being the most expensive model, although maximum reasoning can still make individual Opus tasks costly.

As of now, Claude Opus 5.5 deserves the title of smartest general-purpose model.

Model	Artificial Analysis Intelligence Index	Input / output price per 1M tokens	Where it stands out
Claude Opus 5.5 Max	58	$4 / $20	General reasoning, professional work, science, long context
Claude Sonnet 5.5 Max	56	Lower-tier Claude pricing	High-end capability with faster deployment economics
GPT-6 Astra Max	53	$10 / $50	Science, computer use, document work
Gemini 4 Argon High	53	$2 / $10 currently	Agents, automation, multimodal work
GPT-6.1 Sol Max	52	$2 / $10	Near-Astra capability at far lower cost
Muse Spark 1.3 Max	48	About $1.25 / $4.25	Speed, multimodality, long context
Grok 4.7 Xhigh	46	$2 / $6	Professional work, speed, relatively low price

If you want more recent data on this point, please see our latest frontier AI labs market report.

Has Anthropic actually overtaken OpenAI?

Anthropic has moved ahead of OpenAI on current general model quality, especially in coding and professional knowledge work, although OpenAI still has enough major wins to keep this a close race.

The cleanest comparison comes from Artificial Analysis. Maximum-effort Claude Opus 5.5 scores 58 on its Intelligence Index against 53 for GPT-6 Astra. Opus also leads Astra 67% to 56% on SciCode, 61% to 55% on Humanity's Last Exam, 60% to 59% on Terminal-Bench 4.0 and 85% to 81% on AA-LCR.

The gap becomes more interesting when we look beyond benchmarks. Anthropic said earlier this year that Claude Code had passed $2.5 billion in annualized revenue, more than doubling since the start of the year. Weekly active Claude Code users had also doubled, business subscriptions had quadrupled, and enterprise customers already accounted for more than half of Claude Code revenue.

Anthropic also said that more than 500 customers were spending at least $1 million a year with the company, up from roughly a dozen two years earlier. Eight of the Fortune 10 had become Claude customers. Those numbers tell us that Claude's strength is surviving contact with real companies that can easily pay for competing models.

OpenAI still wins enough hard evaluations to block any claim that Anthropic has opened an overwhelming gap. Astra is stronger on Artificial Analysis' GDP.pdf benchmark, and OpenAI reports extremely high performance on FrontierMath, computer use, cybersecurity and scientific terminal tasks. GPT-6.1 Sol has also brought much of Astra's capability into a far cheaper product.

So yes, Anthropic has overtaken OpenAI in the current general-purpose model ranking. The lead is real, but still small enough to reverse quickly.

Is Google DeepMind actually leading the AI race overall?

Google DeepMind has the strongest case for being the most complete frontier AI lab today, even though Gemini 4 Argon currently trails Claude Opus 5.5 on overall model intelligence.

Argon immediately returned Google to the top language-model group. Artificial Analysis scores it at 53, level with GPT-6 Astra. More importantly, Argon is exceptionally good at tasks where models have to execute work. It reaches 78% on AutomationBench-AA, eight points above maximum-effort Opus 5.5 and ten points above Astra.

Google's own evaluations show a similar pattern. Argon reaches 51.3% on AutomationBench, 65.4% on Vals Finance Agent and 39.5% on Agent's Last Exam. It scores 77.9% on DeepSWE and 91.9% on Vibe Code Bench. Its long-context performance is also unusually strong: Google reports 99.7% on GraphWalks up to 128,000 tokens and 84.2% between 256,000 and one million tokens.

The larger Google advantage comes from everything surrounding Gemini. DeepMind is working near the frontier in video, audio, images, robotics, world models, biology, weather forecasting and scientific discovery. Veo covers video. Google's newer image models handle generation and editing. Gemini Robotics connects the model stack to physical machines. Genie explores interactive world models. Scientific systems extend into genomics, proteins, mathematics and weather.

Google also builds its own TPUs and can deploy AI through Google Cloud, Search, Workspace, Android and Chrome. The same company controls a large part of the research, compute infrastructure and distribution chain.

Argon's limited rollout currently weakens the case slightly. Google has been more cautious about broad access than OpenAI and Anthropic, which means Argon has accumulated less public real-world evidence.

If we rank only current language-model intelligence, Anthropic wins. Across language, agents, multimodality, science and physical AI, Google DeepMind looks stronger overall.

If you want more recent data on this point, please see our latest frontier AI labs market report.

Does OpenAI still lead anything important?

OpenAI still leads several difficult parts of frontier AI, especially science, computer use and deployment at huge scale, while its overall model-quality lead has disappeared.

GPT-6 Astra remains extremely strong on the hardest evaluations. OpenAI reports 98% on FrontierMath Tier 4 and 99.9% on ARC-AGI-3. Its computer-use and cybersecurity results are also among the best currently published. Artificial Analysis gives Astra 31% on GDP.pdf, ahead of Opus 5.5 at 26% and Argon at 22%.

OpenAI's more important move lately may be GPT-6.1 Sol. It costs $2 per million input tokens and $10 per million output tokens, compared with Astra's $10 and $50. Artificial Analysis gives Sol a maximum-effort Intelligence Index score of 52, only one point behind Astra's 53.

That is an unusually small capability gap for a fivefold difference in standard token pricing. OpenAI's own computer-use evaluation also puts GPT-6.1 Sol within a few percentage points of Astra on OSWorld 2.0. For many companies, the cheaper model will therefore make more economic sense even when Astra remains technically better.

The same shift is happening across the market. Gemini 4 Argon also scores 53 on Artificial Analysis while costing $2 per million input tokens and $10 per million output tokens. Claude Opus 5.5 costs $4 and $20, making the current independent intelligence leader cheaper than Astra. Grok 4.7 is cheaper again at $2 and $6, while DeepSeek goes much lower.

Once several models can already solve a task reliably, those cost differences become decisive. Companies can spend the savings on retries, verification, more context or several agents working in parallel.

OpenAI also has a deployment base no frontier startup can currently match. The company has reported more than one billion weekly ChatGPT users, millions of business customers and enormous API usage. Codex has become another large distribution channel for its models inside software development.

That scale gives OpenAI constant exposure to consumer questions, enterprise workflows, coding, search, document creation, computer use and API applications.

OpenAI currently looks weaker than Anthropic on broad independent model quality and narrower than Google across frontier research areas. It still has enough technical, economic and commercial advantages to remain firmly inside the leading three.

Who is winning the AI coding race right now?

Anthropic is winning the AI coding race today because Claude combines top-tier benchmark results with unusually strong real-world adoption among developers.

Claude Opus 5.5 scores 60% on Artificial Analysis' Terminal-Bench 4.0, narrowly ahead of GPT-6 Astra at 59% and Gemini 4 Argon at 57%. Anthropic also performs strongly on professional-work and long-context evaluations, both of which matter when an agent has to understand a large repository and keep working through a complicated task.

The result changes on individual coding tests. Google's published numbers put Argon at 77.9% on DeepSWE, ahead of Astra at 74.1% and Opus at roughly 74%. On FrontierSWE v2, OpenAI reports Astra at 65.5%, Opus around 62% and Argon at 55%. Anthropic then moves back ahead on PostTrainBench.

That back-and-forth tells us the top models are close. Claude Code gives us a better way to break the tie.

Anthropic's $2.5 billion-plus annualized Claude Code revenue means developers are spending serious money on Claude even though OpenAI, Google, xAI, DeepSeek and specialized coding products are all competing for the same work. Anthropic also said weekly Claude Code usage doubled during the first part of the year and enterprise customers now generate more than half of its revenue.

A separate analysis cited by Anthropic estimated that Claude Code was producing about 4% of public GitHub commits worldwide, twice the level from only a month earlier. Any estimate of global GitHub activity deserves some caution, but the direction fits the revenue and user-growth data.

Claude seems especially strong when coding turns into a multi-step job: reading a repository, figuring out what needs changing, editing several files, using tools, testing the result and repairing mistakes.

Across the full picture of benchmarks and actual developer usage, Anthropic currently has the clearest coding lead.

If you want more recent data on this point, please see our latest frontier AI labs market report.

Who is winning AI agents and computer use?

Google DeepMind currently looks strongest in general workflow agents, while OpenAI leads some computer-use tasks and Anthropic remains particularly strong when agents spend a long time inside technical environments.

Gemini 4 Argon's clearest advantage is automation. Artificial Analysis gives Argon 78% on AutomationBench-AA, compared with 70% for Opus 5.5 and 68% for Astra. Google's own AutomationBench setup also puts Argon comfortably ahead of both.

Professional agent tests reinforce that result. Google reports Argon at 65.4% on Vals Finance Agent and 39.5% on Agent's Last Exam. These tasks involve carrying out multi-step work rather than simply knowing the right answer, which makes the results more relevant to actual office automation.

OpenAI remains extremely strong once the model has to operate a computer. GPT-6.1 Sol reaches 71.4% on OpenAI's offline OSWorld 2.0 evaluation at maximum reasoning effort, only slightly below Astra. Google reports 69.2% for Argon on its comparable OSWorld subset.

Anthropic's strength shows up most clearly in coding agents and technical work. Opus leads Terminal-Bench 4.0 among the three models in Artificial Analysis and Claude Code's commercial growth suggests that users are willing to let these agents perform increasingly large chunks of real software work.

We should be careful with exact cross-company agent rankings because tools, browser setup, reasoning budget, retry rules, context management and the agent harness can all move the score.

Google leads general workflow automation, OpenAI is especially strong at computer interaction, and Anthropic has the best case for long-running technical agents.

Agent task	Current leader or strongest case	What we see
General workflow automation	Gemini 4 Argon	78% on AutomationBench-AA
Finance agents	Gemini 4 Argon	Strong lead on Google's Vals Finance Agent comparison
Agentic coding	Claude Opus 5.5	Top-tier Terminal-Bench plus strong Claude Code adoption
Repository engineering	Gemini / OpenAI / Anthropic	Winner changes by benchmark
Computer use	GPT-6 Astra / GPT-6.1 Sol	OpenAI remains particularly strong on OSWorld-style tasks
Scientific terminal work	GPT-6 Astra	OpenAI leads several hard tool-using science evaluations
Long technical workflows	Claude Opus 5.5	Anthropic repeatedly performs well when work extends across many steps

Who is best at science and hard AI reasoning right now?

OpenAI has the strongest case on extreme mathematics and some tool-using science tasks, while Anthropic currently performs better across several broader independent science and reasoning benchmarks.

GPT-6 Astra's most striking published number is 98% on FrontierMath Tier 4. OpenAI also reports 99.9% on ARC-AGI-3 and very high performance on GPQA Diamond. These are difficult evaluations designed to probe areas where earlier models struggled badly.

OpenAI has also been pushing models into real research rather than keeping them inside exam-style benchmarks. The company has published work involving AI assistance on open mathematical problems and formal proofs. Those examples need more scrutiny than a standard benchmark score because the quality of a research contribution can take time to establish, but they show where OpenAI is aiming.

Anthropic still looks stronger on some broad independent measures. Artificial Analysis gives Opus 5.5 67% on SciCode against 56% for Astra and 62% for Argon. Opus also reaches 61% on Humanity's Last Exam, compared with 55% for Astra and 57% for Argon.

Google adds another dimension. DeepMind's scientific work stretches beyond Gemini into specialized systems for biology, weather, genomics and mathematics.

For the hardest general-purpose reasoning tasks, OpenAI and Anthropic are fighting for the lead. Across the broader scientific AI portfolio, Google remains just as important.

Who has the strongest multimodal AI stack?

Google DeepMind currently has the strongest overall multimodal AI stack because it is competitive across text, images, audio, video, long context, robotics and world models at the same time.

Gemini 4 Argon already performs well before we even leave the core model. Google reports 71.6% on Chartography, compared with 71.0% for GPT-6 Astra and 66.3% for Opus 5.5. On LVBench, Argon reaches 91.7%, ahead of Astra at 87.5% and Opus at 83.7%.

The rest of Google's model portfolio makes the gap more obvious. Veo handles video generation. Google's image systems cover generation and editing. Gemini Audio expands the stack into speech and real-time audio. Gemini Robotics moves into physical interaction. Genie explores interactive world generation.

Those areas feed different types of information back into the research program. Video forces models to understand change over time. Robotics forces them to deal with physical space and actions. Audio introduces timing, voice and conversational dynamics that text cannot reproduce.

Meta is now a serious competitor on breadth. Muse Spark 1.3 accepts text, images and video, supports a very large context window and runs much faster than many premium reasoning models. Meta also has image, video, voice and wearable-device programs.

OpenAI is strong across vision, voice, images and video as well. Anthropic's frontier position is much more concentrated around language, coding, agents and knowledge work.

For multimodal AI specifically, Google's lead looks clearer than its lead in pure language-model intelligence.

Are xAI and Meta actually in the top AI tier yet?

xAI and Meta are serious frontier competitors, but neither currently belongs in the same model-quality tier as Anthropic, Google DeepMind and OpenAI.

Grok 4.7 is the stronger case for xAI. Artificial Analysis gives the Xhigh version an Intelligence Index score of 46, with roughly 92 output tokens per second and API pricing of $2 per million input tokens and $6 per million output tokens.

The difficulty is the gap at the top. Opus 5.5 scores 58, Argon and Astra score 53, and GPT-6.1 Sol scores 52. Grok therefore sits six points behind even the cheapest model in that leading group and 12 points behind Opus.

Grok remains good on professional-work evaluations and offers attractive speed and pricing. Its weaker results on some long-horizon coding tasks make it harder to call xAI an overall leader today. Artificial Analysis has Grok 4.7 far below Opus, Astra and Argon on Terminal-Bench 4.0.

Meta has made a more dramatic comeback. Muse Spark 1.3 scores 48 at maximum effort and Artificial Analysis measures output speed around 190 to 200 tokens per second. Earlier Llama generations were nowhere close to this position.

Muse Spark still has clear weaknesses. Its agentic coding scores trail the top three, and Meta has not yet produced several consecutive generations at this new level.

Both companies have reasons to take seriously. xAI has moved extremely quickly and is spending heavily on compute. Meta combines vast infrastructure with Facebook, Instagram, WhatsApp, Messenger and wearable devices that can distribute new AI features immediately.

For now, Anthropic, Google and OpenAI remain a tier above both.

If you want more recent data on this point, please see our latest frontier AI labs market report.

Are Chinese AI labs now genuinely at the frontier?

Chinese AI labs are genuinely operating near the global frontier today, especially in open models, efficiency and developer usage, although the strongest US models still hold the top positions in general intelligence.

Stanford's 2026 AI Index found that the US-China model-quality gap had shrunk to only 2.7% in its March comparison. Chinese and American models had already traded places at the top several times since early 2025.

The latest generation still shows strong Chinese models even after Anthropic, Google and OpenAI moved ahead again. Alibaba's Qwen family, Moonshot's Kimi, Z.ai's GLM models, Xiaomi's MiMo and DeepSeek all sit in or near the broader frontier group.

DeepSeek is particularly important because its approach attacks the economics of the race. Recent DeepSeek models use mixture-of-experts architectures that activate only a fraction of their total parameters for each token. That allows DeepSeek to offer strong reasoning and coding performance at prices far below premium Western models.

Developer usage gives us another useful view. In OpenRouter's latest complete week, DeepSeek handled 21.8% of text requests, Google 21.5%, OpenAI 18.2%, Z.ai 6.6%, Qwen 6.2%, Anthropic 2.5% and Xiaomi 2.4%.

OpenRouter represents one routing platform rather than the entire AI market, so those percentages should not be read as global market share. Still, Chinese providers collectively handling such a large share of activity on a platform heavily used by developers is difficult to dismiss.

The strongest closed model currently comes from Anthropic, but Chinese labs have already erased the idea that the frontier itself is exclusively American.

Who is winning the open-model race?

Chinese labs currently dominate the open-weight AI race, with Xiaomi, Z.ai, Moonshot, DeepSeek and Alibaba doing more to set the pace than the companies that led open AI a few years ago.

Artificial Analysis currently puts Xiaomi's MiMo-V2.6-Pro around the top of its open-weight intelligence rankings, with Z.ai's GLM-5.3 and Moonshot's Kimi K3 close behind. DeepSeek sits somewhat lower on raw intelligence but offers an unusually aggressive combination of cost and speed.

Kimi K3 shows how far these systems have moved. It uses a very large mixture-of-experts architecture, supports roughly one million tokens of context and reaches an Artificial Analysis Intelligence Index score in the mid-40s. That puts an open-weight system surprisingly close to several proprietary frontier models.

DeepSeek attacks a different part of the problem. V4.1 Flash activates only a small share of its total parameter count for each output token and is priced around $0.30 per million input tokens and $1.20 per million output tokens. Independent providers have served it at more than 200 output tokens per second.

Alibaba's Qwen ecosystem adds breadth. Qwen models now cover general reasoning, coding, multimodal work and many smaller deployment sizes. Their widespread availability also makes them easy for developers to fine-tune or run through third-party providers.

Meta still matters because of its research history, distribution and willingness to release some open weights, although its strongest new Muse models have moved away from the fully open strategy associated with Llama. Mistral remains relevant in Europe and for companies looking for a Western-controlled alternative.

Open models still trail the very best closed models, but the gap is narrow enough that low cost, self-hosting and customization can outweigh it for many workloads.

Open-weight model family	Current position	Main advantage
Xiaomi MiMo	At or near the top of current open-weight rankings	Very high intelligence
Z.ai GLM	Frontier open-weight group	Strong reasoning and agent performance
Moonshot Kimi K3	Frontier open-weight group	Huge model, long context
DeepSeek V4 family	Slightly below peak open-weight intelligence	Exceptional price and inference efficiency
Alibaba Qwen	Broad frontier ecosystem	Large range of models and strong developer adoption
Meta open models	No longer leading the absolute open frontier	Ecosystem, research depth and distribution
Mistral	Competitive Western alternative	Enterprise control and European positioning

Can we still trust AI benchmarks enough to rank these labs?

AI benchmarks can still tell us which labs belong near the frontier, but they are increasingly unreliable for declaring a permanent number-one model.

Stanford's 2026 AI Index found error rates as high as 42% in some widely used benchmark datasets. The report also showed how quickly models are catching tests that were supposed to remain difficult: Humanity's Last Exam performance improved by about 30 percentage points in only one year.

Agent evaluations add another problem. The surrounding setup now affects results almost as much as the model itself. The same model can perform differently depending on available tools, context management, reasoning effort, retry rules and the software harness controlling it.

That helps explain why apparently comparable numbers can disagree. Anthropic, OpenAI, Google and independent evaluators sometimes report different Terminal-Bench or computer-use results for the same model family. The evaluations may use different releases, settings or agent systems.

Company benchmark selections also deserve skepticism. Google naturally publishes many tasks where Gemini performs well. OpenAI emphasizes difficult science, computer use and cybersecurity. Anthropic shows plenty of coding and knowledge-work evaluations.

Independent composites such as Artificial Analysis reduce that problem by evaluating models under one methodology, although somebody still has to decide which tests belong in the index and how much each one matters.

The safest reading is to look for patterns that survive several independent evaluations, real customer usage and different types of task. Using that standard, the top group is much clearer than the exact order inside it.

Which frontier AI labs are actually leading right now?

Anthropic, Google DeepMind and OpenAI are clearly leading the frontier AI race right now, with Anthropic first on current general model intelligence, Google strongest across the widest technical range, and OpenAI still exceptional in science, computer use and mass deployment.

Anthropic has the cleanest claim to number one if we force ourselves to choose one model company today. Claude Opus 5.5 leads Artificial Analysis at 58, five points ahead of Gemini 4 Argon and GPT-6 Astra. Claude Sonnet 5.5 is also above both at maximum effort. Claude's coding position is backed by a real business with more than $2.5 billion in annualized Claude Code revenue, which makes the current Anthropic lead much harder to dismiss as benchmark noise.

Google DeepMind may have the strongest long-term position across the whole field. Gemini 4 Argon already matches Astra on Artificial Analysis and leads AutomationBench-AA. Google then adds world-class video, image, audio, robotics, world-model and scientific AI teams, plus its own TPUs and one of the largest software distribution networks in the world.

OpenAI has lost the automatic presumption of technical leadership it enjoyed during earlier parts of the generative-AI boom. It remains extremely close. Astra leads several very hard science and computer-use tasks, while GPT-6.1 Sol gets within one Artificial Analysis point of Astra at one-fifth of the standard token price. More than a billion weekly ChatGPT users give OpenAI a deployment and feedback loop that Anthropic cannot currently reproduce.

Meta and xAI come next. Meta has made the bigger technical jump with Muse Spark, while xAI has remained competitive with Grok and built enormous compute capacity very quickly. Their current models still sit several points below the leading group on broad independent intelligence tests.

Chinese labs deserve their own place in the ranking because they are strongest in a different part of the market. Xiaomi, Z.ai, Moonshot, DeepSeek and Alibaba are setting much of the pace in open weights, efficiency and low-cost inference.

The hierarchy is fairly clear today. Anthropic has the best general-purpose model. Google DeepMind has the broadest frontier capability. OpenAI combines top-level research with unmatched model distribution. Meta and xAI are credible challengers. Chinese labs already control a large part of the open and efficiency frontier.

Those three, Anthropic, Google DeepMind and OpenAI, are the real leaders of the model race today. Anthropic has the small current edge.

If you want more recent data on this point, please see our latest frontier AI labs market report.