A factual survey of artificial intelligence at mid-2026: the models that lead, the agents that ship, the silicon that scales, the capital that funds it, and the people who use it.
Between May 2025 and May 2026 the artificial intelligence industry compressed roughly a decade of normal-pace technology evolution into a single year. Frontier reasoning models, treated as exotic a year earlier, became table stakes. The open-weight tier closed to within ten benchmark points of the closed frontier. Chinese laboratories went from second-tier to setting the pace on cost. Hyperscaler capital expenditure crossed three quarters of a trillion dollars. And the most consequential single product release of the period, an agentic coding tool, displaced a category that did not exist eighteen months before.
This issue documents that landscape as it stands at the end of May 2026. It is a state of the nation, not a forecast. Where the numbers exist they are cited; where they are estimated they are flagged. The intention is to give a senior reader a map they can use to navigate the next six to twelve months without having to reconstruct the present each time a vendor rolls a new model.
The structure proceeds from the model layer outward. First the closed frontier, with Chinese laboratories woven through the same comparison rather than ghettoised. Then the reasoning shift, which is the single most important architectural development of the period. Then the open-weight pillar, the agentic application layer where commercial intensity is highest, a brief survey of adjacent modalities, the silicon and capital that underwrite everything else, and finally adoption: by businesses, by consumers, and by the capital markets that price the whole thing.
The model frontier at mid-2026 is no longer a US story. Western and Chinese laboratories now ship competitive flagships within weeks of each other, separated by capability deltas small enough that the procurement decision turns on cost, licensing, and access rather than raw intelligence.
| Model | Origin | Notable strengths | API in/out | Context |
|---|---|---|---|---|
| GPT-5.5 ProOpenAI | USA | Leads abstract mathematics at 39.6% FrontierMath Tier 4. Strongest all-rounder ecosystem, deepest integration with Microsoft, broadest consumer reach. | $2.50 / $15 | 400k |
| Claude Opus 4.7Anthropic | USA | Leads SWE-Bench Verified and software engineering benchmarks. Powers most production agentic coding through Claude Code, Cursor, Windsurf. | $15 / $75 | 200k |
| Gemini 3.1 ProGoogle DeepMind | USA | Leads multimodal and reasoning with 77.1% on ARC-AGI-2, 94.3% GPQA Diamond. Native Google Search grounding for real-time facts. | $2 / $12 | 1m |
| Grok 4.3xAI | USA | Strong on coding (75% SWE-Bench), uncensored style, real-time X data grounding. Largest practical context window via Grok 4 Fast at 2 million tokens. | $2 / $15 | 2m |
| DeepSeek V4 ProDeepSeek | CHN | 1.6 trillion parameter MoE with Hybrid Attention. 80.6% SWE-Bench, 90.1% GPQA Diamond. Cache-hit pricing as low as $0.07/M reshapes API economics. | $0.30 / $1.20 | 1m |
| Qwen 3.7 MaxAlibaba | CHN | Joint top of Artificial Analysis Intelligence Index at 57. 97.1 HMMT competition math, 92.4 GPQA Diamond. The strongest multilingual flagship. | $2.50 / $7.50 | 1m |
| GLM-5.1Zhipu AI | CHN | Leads BenchLM Chinese rankings. Tops SWE-Bench Pro. Trained entirely on Huawei Ascend silicon with zero NVIDIA dependency. | MIT open | 256k |
| Kimi K2.6Moonshot | CHN | Ties GPT-5.5 on several benchmarks at 5 to 10 times cheaper. Cheapest top-ten model by GPQA Diamond at $0.95/M input. Strong agent and coding posture. | $0.95 / $3 | 200k |
| Doubao 2.0 / Hunyuan 3.0ByteDance / Tencent | CHN | Distribution rather than benchmark plays: Doubao embedded across Douyin and Volcengine, Hunyuan inside WeChat with 1.4 billion monthly users. Combined consumer reach inside China rivals OpenAI globally. | closed | varies |
Top Chinese flagships sit within 5 to 15 points of the leading Western models on standard benchmarks. DeepSeek V4 Pro leads BenchLM at 87; Kimi K2.6 follows at 84; GLM-5 and GLM-5.1 tie at 83. The Western lead on subjective output polish, particularly English prose, remains real but narrow. The procurement decision has shifted from capability to economics, licensing, and where data is processed.
Five major closed-frontier releases between February and May 2026 pushed per-token costs down sharply. Chinese models opened a 15 to 30 times cost gap on comparable workloads through aggressive cache-hit pricing. Anthropic and OpenAI retain enterprise pricing power, but the API price floor has collapsed in commodity inference. For high-volume agentic workloads, effective input cost can sit at $0.03 to $0.07 per million tokens.
The highest-volume Chinese models rank outside the global intelligence top ten. Xiaomi processes 4.21 trillion weekly tokens on OpenRouter at 21.1% market share, three times OpenAI's share on the same platform. Hunyuan reaches 1.4 billion users through WeChat. Doubao runs inside Douyin and the Cici consumer assistant. The implication: leaderboard position is a poor proxy for deployed footprint in the world's second largest AI market.
Anthropic shipped major Claude releases roughly every two weeks through 2026: Opus 4.6 in February, Sonnet 5 in March, Opus 4.7 in April. OpenAI followed with GPT-5.4 in March and GPT-5.5 in April. Google launched Gemini 3.1 Pro in February and refreshed in May. The release cadence has effectively eliminated the planning horizon over which a procurement team can pick a single best model and stick with it.
The most significant architectural development of 2025 to 2026 is not a new model family. It is the productisation of test-time compute. Models that pause, plan, and self-correct before answering have moved from research curiosity to a standard configuration parameter on every frontier API.
OpenAI's o-series, launched in early 2025 with o1 and o3-mini, popularised System 2 cognition for language models. By mid-2026 every frontier laboratory ships an explicit thinking mode. Anthropic exposes a budget_tokens parameter on Claude Opus 4.7 Thinking that lets developers spend more inference compute for harder problems. Google's Gemini 3.1 Pro and OpenAI's GPT-5.5 Pro both offer reasoning effort controls. DeepSeek-V3.2 ships with reasoning on by default behind a sparse-attention indexer that holds the cost down.
The shift matters because it inverts the old scaling story. Training-time scaling required ten times the compute for measurable capability gains. Inference-time scaling trades latency for accuracy in close to a log-linear curve. A small model thinking for longer can match a large model thinking briefly on hard tasks, which is exactly the trade reasoning APIs now expose to the developer.
The other half of the reasoning story is distillation. DeepSeek R1, released in January 2025, demonstrated that reasoning patterns could be distilled from a 671 billion parameter teacher into 1.5 to 70 billion parameter students under permissive licences. The distilled R1 family on Qwen and Llama bases now ships under MIT terms. R1-Distill-Qwen-32B outperforms o1-mini on multiple benchmarks at a fraction of the operating cost.
For enterprise builders this changes the economics of reasoning. The capability is no longer locked behind an API paywall. A single A100 or RTX 5090 class machine can host a reasoning-tuned 8 to 32 billion parameter model that solves problems which required frontier API calls a year earlier. This is the central reason the open-weight tier has closed the gap to the closed frontier so quickly.
The reasoning revolution is not a model. It is a knob. Every frontier laboratory now ships fast mode and think mode behind the same API, and the second is becoming the default for any task that is not bound by latency.
Frontier Intelligence · May 2026Open-weight models in May 2026 sit in a fundamentally different commercial position to a year earlier. Six laboratories now ship competitive frontier-class open weights. The question for an enterprise buyer is no longer whether open is good enough; it is which licence permits the deployment they want, and which lab is most likely to ship a successor within their planning horizon.
The single most underreported development in the open-weight tier is hardware. GLM-5 was trained entirely on Huawei Ascend silicon with no NVIDIA dependency. For Chinese state enterprises and sovereign buyers facing US export controls on H100 and Blackwell hardware, this changes the procurement calculation more than any benchmark point. The forced innovation under sanctions has produced models that the rest of the world can also run on cheaper infrastructure.
Apache 2.0 and MIT remain the gold standard. Qwen, Mistral, DeepSeek, GLM, and Gemma all ship under unrestricted commercial terms. Llama 4 alone retains a 700 million monthly active user cap that excludes the largest deployments and introduces EU restrictions. For enterprises building durable products on open weights, the legal review now matters more than the benchmark sheet. Counterintuitively, this favours the Chinese ecosystem for many global buyers.
If 2024 was the year of the chatbot and 2025 was the year of reasoning, 2026 is unambiguously the year of the agent. Coding agents in particular have crossed from demo to production at unusual speed, with two companies passing a billion dollars in annualised revenue inside their first eighteen months of general availability. The pattern is replicating into adjacent enterprise functions.
The fastest-growing developer tool in software history by adoption curve. Runs in the terminal, edits files in place, and crossed one million weekly active developers within six months of general availability. Powers Anthropic's enterprise software engineering revenue and drives a meaningful share of the company's $30 billion annualised run-rate. Uses five and a half times fewer tokens per task than competitors in independent tests.
The IDE incumbent for AI-native development. $29.3 billion valuation, over one hundred thousand enterprise users, Cursor Cloud Agents support parallel task execution against the same repository. Wins decisively on inline tab-completion latency at three-second time-to-first-token. Standard pairing with Claude Code as the team default in many engineering organisations.
The fully autonomous archetype. Submit a scoped engineering task, Devin works for minutes to hours, returns a pull request for review. Enterprise adoption growing at 40% month on month. Nubank case study reports 8 to 12 times efficiency gains in production. Cognition acquired the Windsurf IDE for $250 million in late 2025, consolidating two of the four major coding agent platforms.
The distribution play. Now operates as a true agent with Copilot Workspace for issue-to-PR workflows, routing across Claude, GPT, and Gemini depending on the task. Sits inside the developer pipeline where the work already happens. The most likely default for buyers who want a single procurement contract spanning autocomplete and agentic work.
Agentic workflows are replicating fastest in functions adjacent to code: customer support resolution, security operations, legal contract review, and financial close. The common signature is a long-horizon task with verifiable outcomes and tolerance for human review. Functions that lack one of those three (open-ended creative work, high-stakes irreversible decisions, conversation with full ambiguity) remain stubbornly chatbot-shaped.
The Stanford Enterprise AI Playbook documents a clean pattern: 77% of enterprise API usage is full automation versus 45% on Claude.ai consumer surfaces. Businesses are deploying agents differently to individuals, with structured human oversight on a smaller percentage of decisions rather than every decision.
The model layer has consolidated around a small number of providers. The agent layer is consolidating around an even smaller number of protocols. Model Context Protocol (MCP), originated by Anthropic, has become the de facto standard for connecting agents to tools and data sources across all four major coding agents. The number of production MCP servers crossed several thousand in early 2026, with developer infrastructure, files, search, and databases as the most deployed categories.
The next moat is not the model. It is the catalogue of tools an agent can reach and the durability of the protocol that connects them.
Outside of the text and reasoning frontier, three other modalities crossed commercial thresholds in the period. They warrant brief coverage rather than deep treatment because the centre of commercial gravity in 2026 remains text, code, and agents.
Google's Nano Banana Pro, refreshed with Gemini 3.5 on 19 May 2026, leads on photoreal portraits and product imagery. Midjourney v8 retains the stylised art crown. ByteDance's Seedream 5.0 and Zhipu's CogView match Western models on photorealism and dominate on Chinese typography. There is no single best image model in May 2026; there are five strong ones, each better at a specific class of output.
OpenAI formally discontinued Sora 2 on 26 April 2026. Google Veo 3.1 is now the leading frontier video model. Kling 3.5 is the runner-up on fast iteration; Runway Gen-4 leads cinematic control. ByteDance's Seedance 2.0 produces multi-shot film sequences in 60 seconds from text prompts. Video generation is now genuinely production-grade for short-form social and previsualisation, still short of feature-film quality.
OpenAI's Realtime API and Google's live audio in Gemini both crossed sub-300ms end-to-end latency through 2025 and 2026. The result is conversational voice agents that no longer have the half-second pause that gave them away. Voice has slipped into customer service deployments, drive-through ordering, and accessibility tooling without much fanfare. Suno and ElevenLabs lead on music and synthetic speech respectively, both at production scale.
Foundation models for physical action are present at every major laboratory but commercially nascent. Figure, 1X, and Sanctuary in humanoids; Tesla Optimus and Boston Dynamics in industrial; Physical Intelligence and Skild as foundation-model layers. Real revenue remains scarce. Embodied AI is more likely to become a major 2027 to 2028 story than a 2026 one.
Underneath every model and every agent sits an unprecedented capital programme. NVIDIA's data centre revenue alone hit $62.3 billion in a single quarter. Combined hyperscaler capex on AI infrastructure crosses $700 billion in 2026. The largest single AI project, OpenAI's Stargate, is committed to 10 gigawatts and $500 billion of buildout over three years. By the description of its own architect: a once-in-a-generation infrastructure cycle.
| Organisation | 2026 capex | Notable commitment |
|---|---|---|
| Alphabet | $150b+ | Largest single hyperscaler programme. 3.5GW Google TPU capacity contracted with Anthropic alone, on top of internal Gemini training and Workspace deployment. |
| Microsoft | ~$180b | $250b Azure contract with OpenAI under revised October 2025 partnership terms. Maia custom silicon scaling, plus Blackwell and Rubin deployments. |
| Amazon | ~$170b | $50b investment in OpenAI alongside $38b seven-year AWS commitment for Trainium compute. Anthropic remains anchor tenant on Bedrock. |
| Meta | ~$100b | $100b AMD partnership, taking an equity stake in exchange for GPU supply. Llama 4 training, Meta AI consumer deployment across 1 billion users. |
| OpenAI / Stargate | $500b | 10GW planned across seven US sites. 0.3GW already operational in Abilene Texas. Oracle and SoftBank as construction partners. Wisconsin, Ohio, New Mexico, Michigan sites under active construction. |
| NVIDIA | ~$500b backlog | FY26 data centre revenue $215.9b annual. $500b committed orders for Blackwell and Rubin. Vera Rubin samples shipping; production racks in H2 2026. Now TSMC's largest customer. |
NVIDIA holds approximately 80% of the AI accelerator market. The Blackwell platform peaked in 2025 deployment; Vera Rubin samples shipped to customers in early 2026 with production racks coming in the second half of the year. Individual Rubin GPUs draw 2,300 watts, almost double Blackwell, with NVL72 racks pulling 120 to 130 kilowatts and demanding a fundamental rethink of data centre power and cooling.
Custom silicon is the second story. Google TPU v7 is the most mature. Amazon Trainium 2 is shipping at scale through the OpenAI commitment. Microsoft Maia is internal. Meta's MTIA programme continues alongside the AMD deal. OpenAI is developing its own inference chip with Broadcom, codenamed Titan, on TSMC 3nm with mass production targeted for H2 2026. Huawei Ascend anchors Chinese sovereign deployments and trained GLM-5 end to end.
Capital is no longer the binding constraint. Electricity is. OpenAI wrote to the US Office of Science and Technology Policy describing an electron gap between the US and China, with China adding 429 gigawatts of new power capacity in 2024 against 51 gigawatts in the United States. The Stargate buildout has gone to natural gas plants on three sites to bypass interconnect queues.
Closed-loop liquid cooling is standard on six of seven Stargate sites. HBM (high-bandwidth memory) supply is the other shortage. TrendForce projected HBM demand growing more than 100% year on year in 2025; 2026 demand stacks on top of that base. Buyers not already in NVIDIA's 2026 procurement pipeline are now looking at 2027 delivery dates for Vera Rubin systems.
The Stanford 2026 AI Index, released in April, anchors the adoption picture. The headline finding is that consumer adoption has outpaced personal computers and internet diffusion. The enterprise picture is similar in surface coverage but starkly different on depth: most organisations have deployed AI, very few have scaled it.
88% of enterprises now use AI in at least one business function, up from 55% in 2024. 72% have a workload in production. But fewer than 10% describe their AI deployment as scaled across the organisation. The most-cited risk has flipped from cybersecurity to inaccuracy, with 74% of respondents now ranking hallucination as their top AI concern, ahead of regulation and privacy.
The implementation pattern that works: structured human oversight. McKinsey reports 65% of AI high performers have defined human-in-the-loop processes against 23% of other organisations. The average enterprise now runs 4.2 AI models in production, more than double the 2023 baseline. Coding, marketing content, and customer service are the three functions where production deployment is most mature.
ChatGPT's grip on the consumer market has loosened. Web traffic share fell from 77% to 57% in twelve months while Gemini climbed from 6% to 25% and Claude from 1.4% to 6%. SimilarWeb describes this as the second-largest market share rebalancing in any consumer technology category in the last decade, behind only iOS versus Android. ChatGPT still leads in absolute scale, with 900 million weekly active users globally and a brand position no competitor has matched.
A note on comparability: Meta AI and AI Overviews figures reflect embedded surface exposure rather than active intent-driven use, and cannot be directly compared with ChatGPT's WAU or Claude's MAU. The pattern across all measures is the same: ChatGPT still leads, Gemini gained the most, Claude is the smallest by users but largest by revenue per user.
In April 2026 Anthropic crossed $30 billion in annualised run-rate revenue, ahead of OpenAI's $25 billion. The number is consequential. The trajectory that produced it, 30 times growth in fifteen months, is the structural shift. A company that was effectively pre-revenue in early 2024 now out-earns most of the Fortune 500 by run-rate.
$87m run-rate in January 2024. $1bn by December 2024. $9bn by year-end 2025. $14bn in February 2026. $19bn in March. $30bn in April. The $14 billion to $30 billion step in roughly eight weeks is hard to read in traditional software terms. Driven by enterprise: 80% of revenue from business customers, more than 1,000 organisations spending over $1 million annually, available across all three major cloud platforms.
February 2026 Series G of $30 billion at $380 billion post-money, led by GIC and Coatue. Sequoia, Fidelity, Temasek, ICONIQ participating. Reportedly considering a further $50 billion round at $900 billion that would make Anthropic the most valuable private AI company in the world. IPO targeted October 2026, potentially raising more than $60 billion. Spending roughly 4x less on training than OpenAI projects through 2030.
OpenAI confirmed $2 billion in monthly revenue alongside its $122 billion raise at $852 billion valuation. 900 million weekly active ChatGPT users. 93% of Fortune 500 with some adoption. 7 million enterprise seats, 9x year-on-year. 1.5 million business clients globally, over 1 million outside the US.
The cost structure is correspondingly large. OpenAI burning approximately $17 billion in cash this year against a projected $14 billion 2026 loss. Over $1 trillion committed to infrastructure across the next several years. Positive free cash flow not projected until 2029. $500 billion in disclosed cloud capacity across Microsoft Azure ($250bn), AWS ($38bn + $50bn equity), Google, and Oracle. The largest single corporate compute commitment in history.
A year ago the consensus was that OpenAI had an insurmountable lead. The brand, the user base, the head start. The $30 billion number says the consensus was wrong. The investors funding both companies are not confused about the cost structure either.
Frontier Intelligence · May 2026Five observations from the data, offered as context rather than forecast.
Capability convergence at the top means buyers will spend the next phase optimising for cost per query, latency per task, licensing terms, and data jurisdiction rather than chasing the next leaderboard point. Procurement teams that have not built a multi-model routing layer are leaving 30 to 60% of inference savings on the table.
For non-frontier workloads the question is not whether to consider open weights. It is which open-weight family. The closed frontier retains a meaningful lead on the very hardest reasoning and creative tasks. Everything else, increasingly, can be served at a fraction of the price by a Qwen or DeepSeek deployment.
The fastest revenue growth and the most concentrated commercial intensity is in the agentic application layer, particularly in coding. The pattern will replicate into customer support, security, legal, and back office through 2026 and into 2027. The model providers know it. Anthropic's revenue mix already reflects it.
Capital can be raised. Models can be trained. Power cannot be added quickly. The relative pace of US versus Chinese electricity capacity buildout is now a frontline AI policy question, not a footnote. Three Stargate sites operating on natural gas tells you everything about the binding constraint.
Western buyers who chose to ignore the Chinese frontier in 2024 were defensibly cautious. The same choice in 2026 is increasingly hard to justify on capability or cost grounds. The remaining concerns are jurisdictional (where the data sits), regulatory (sanctions exposure), and brand (which logos a customer can stomach in a vendor list). Those are real constraints, not capability ones.
Anthropic IPO, OpenAI IPO, Rubin shipping in production, Stargate sites coming online, DeepSeek V5, Llama 5, Gemini 4, GPT-6. None of these are speculative; all are expected before May 2027. The single planning assumption a senior reader should make is that the volatility of model and capability releases will increase, not decrease.
Primary numerical sources include the Stanford 2026 AI Index Report (April), McKinsey Global AI Survey, IDC Worldwide AI Spending Guide, US Census Bureau Business Trends and Outlook Survey, the Anthropic Economic Index, NVIDIA Q4 FY26 financial filings, Epoch AI Stargate analysis, Artificial Analysis Intelligence Index, BenchLM Chinese model leaderboard, SimilarWeb consumer AI traffic data, OpenRouter token volume data, Sacra company estimates, and disclosure from OpenAI and Anthropic on revenue and valuation. Benchmark scores reflect publicly verified leaderboard positions at the time of writing. Model pricing reflects each provider's public price list and is verified against billing samples where available. Forward statements (Rubin production timing, IPO timing, V5 and Llama 5 expected releases) reflect publicly disclosed roadmaps and are flagged as such. Numbers cited as estimates are derived from analyst aggregation rather than primary disclosure. All figures current to May 2026. All data points are sourced from public filings, vendor releases, news reports, and published research. No input from any employer's internal analysis or non-public material has been used in the preparation of this report.