Six new tree rows. The loudest model-layer week of Q2 — Gemini 3.5 + Omni, Qwen3.7-Max in the global top 5, three open-weights drops that reset what 'open' means.

Registry movement

I/O + Alibaba Cloud Summit + NVIDIA Q1 FY27 + Cohere / Microsoft / Tencent open-weights all stacked into five business days. This was the densest model-layer week of the quarter. Microsoft Build (June 2-3) sits outside the window, so Microsoft's MAI silence on the platform layer is not yet a signal.

Added

gemini-3-5-flash, gemini-omni-flash, qwen-3-7-max, cohere-command-a-plus, fara-1-5-27b, hy-mt2-30b-a3b

Updated

None

Google DeepMind · 2026-05-19 · Frontier · Multimodal

Gemini 3.5 Flash

Buyers should rerun agentic-coding RFPs — a Flash-tier model now out-codes prior-gen Pro on most agent benchmarks at 4x speed.

Gemini 3.5 Flash hits Terminal-Bench 2.1 76.2%, MCP Atlas 83.6%, and OSWorld 78.4% — beating Gemini 3.1 Pro on almost every agentic eval while running at ~280 output tokens per second, the new Pareto frontier on speed-vs-intelligence. Architects optimizing for sustained agent loops (not chat) should re-baseline cost-per-task within 30 days; Flash pricing is up 3x ($1.50/$9 per Mtok) but per-task cost is often half of frontier peers because of latency and token economy.

Model registry ID

gemini-3-5-flash

Sources deepmind.google/models/gemini/flash, blog.google

Google DeepMind · 2026-05-19 · Frontier · Multimodal

Gemini Omni Flash

Marketing / ad / content-ops leaders should pilot an Omni workflow this quarter — natively multimodal video gen is now in-product, not a research demo.

Omni Flash ingests text / image / audio / video and outputs grounded 10-second video clips with conversational editing — landing simultaneously in the Gemini app, Google Flow, YouTube Shorts, and YouTube Create. The 'create anything from any input' framing is real because the model carries Gemini's world-knowledge through the generation, unlike Veo's text-to-video pipeline. API access lands in weeks; enterprises that ignore this until then will be six weeks behind on creative-ops tooling.

Model registry ID

gemini-omni-flash

Sources blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-omni

Alibaba · 2026-05-20 · Frontier · Reasoning

Qwen3.7-Max

China-sovereignty buyers have a top-5 global intelligence option in-country — add Qwen3.7-Max to closed-model bake-offs this quarter.

First Chinese model in the Artificial Analysis Intelligence Index top 5 (56.6, ahead of Gemini 3.5 Flash's 55.3); 1M context, 35-hour autonomous tool-use run demonstrated on a real engineering task. Closed-weight strategy and Alibaba Cloud-only access (with homegrown Zhenwu M890 silicon underneath) explicitly target enterprises that want frontier intelligence without US-vendor dependence. The 3-4 point gap to GPT-5.5 has closed enough that 'use Western frontier' is now a debatable position, not a default.

Model registry ID

qwen-3-7-max

Sources alibabacloud.com/blog, marktechpost.com, Artificial Analysis

Cohere · 2026-05-20 · Open Frontier · Moe

Command A+

Sovereign-AI and data-residency buyers should add Command A+ to their on-prem shortlist — Cohere just shipped the first Apache 2.0 frontier-adjacent MoE that fits on two H100s.

218B total / 25B active MoE, multimodal with citation grounding, W4A4 quantization on B200 or 2x H100 — a deployment profile that previously required either a closed API or a license-restricted open model. τ²-Bench Telecom jumped 37% to 85% generation-over-generation; AA-Omniscience Non-Hallucination #1 at 86%; 48 languages. Procurement teams running RAG or agentic workflows in regulated industries should evaluate this within the next two sprints; strongest non-Chinese open-weight frontier-adjacent candidate in the Western catalog.

Model registry ID

cohere-command-a-plus

Sources cohere.com/blog/command-a-plus, venturebeat.com

Microsoft Research · 2026-05-22 · Specialist · Multimodal

Fara1.5-27B

Teams building computer-use agents should switch to open-weight Fara1.5 before scaling — Microsoft's tiny model beats OpenAI Operator and Google Computer Use on live web tasks.

Fara1.5-27B scores 72% on Online-Mind2Web vs 58.3% for OpenAI Operator and 57.3% for Gemini 2.5 Computer Use; even the 9B variant beats them at 63.4%. Built on Qwen3.5, open weights on Azure AI Foundry. Collapses the cost structure for browser-agent deployments — operators currently paying per-task to closed providers can run an entire fleet on a single host. Microsoft's strategy reveal: in browser-agent specialty, scale is no longer the moat.

Model registry ID

fara-1-5-27b

Sources microsoft.com/en-us/research/articles/fara1-5-computer-use-agent, decrypt.co

Tencent · 2026-05-21 · Specialist · Moe

Hunyuan Hy-MT2-30B-A3B

Localization and globalization ops teams should test Hy-MT2 against incumbent MT APIs — 30B MoE running 3B active now leads open-source translation; 1.8B variant runs offline on a phone.

Family release (1.8B / 7B / 30B-A3B) supporting 33 languages + 5 Chinese dialects. The 1.8B variant fits in 440MB after 1.25-bit AngelSlim quantization — runs offline on a phone. The 7B and 30B-A3B beat DeepSeek-V4-Pro and Kimi K2.6 on fast-thinking translation. Architects shipping multilingual products should benchmark: this is the first open-weight family where on-device offline translation is genuinely production-grade, not a research curiosity.

Model registry ID

hy-mt2-30b-a3b

Sources github.com/Tencent-Hunyuan/Hy-MT2

Speed-intelligence Pareto frontier reset by Flash tier

Gemini 3.5 Flash at 278-284 output tokens per second with AA Intelligence Index 55+ has redrawn the Pareto frontier — Flash-class models now match prior-generation Pro intelligence at sub-100ms-per-token throughput. Architects budgeting for agentic loops should recompute cost-per-task assuming Flash-tier models now do what required a Pro model six months ago. Closed Pro-tier models retain a 3-5 point intelligence lead but at 4-6x the latency.

Examples

Gemini 3.5 Flash (~280 tok/sec, AA Index 55+), Command A+ (W4A4 quantized, 2x H100), Gemini 3.1 Flash-Lite (carryover anchor)

Sources artificialanalysis.ai/articles/cohere-launches-open-weights-model-command-a, deepmind.google/models/gemini/flash

Agent-first ground-up redesigns replace chat-first architectures

Every flagship model launched this week was explicitly engineered for sustained multi-step agentic workloads — long-horizon planning, parallel tool use, persistent state — not for conversational turns. Antigravity 2.0 ships as a standalone agent-first GUI replacing the IDE paradigm; Command A+ collapsed four prior single-purpose models into one unified weights. Architects assuming a chat-first interaction model in their AI stack should review their roadmaps — vendor pricing, latency tuning, and API shapes are all moving toward agents as the unit of work.

Examples

Gemini 3.5 Flash + Antigravity 2.0, Qwen3.7-Max (35h autonomous demo), Command A+ (unified four prior single-purpose models), Mistral Medium 3.5 (anchor)

Sources antigravity.google/blog/introducing-google-antigravity-2-0, cohere.com/blog/command-a-plus

Open-weight frontier shifts from license-restricted to fully Apache 2.0

Cohere's Apache 2.0 release of a 218B MoE — its first ever under that license — joins Gemma 4's Apache 2.0 shift in marking a definitive move away from license-restricted open weights at the frontier. Llama Community License-style restrictions are now the laggard, not the standard. Procurement teams that previously had to lawyer up around 700M-MAU caps and field-of-use carveouts have a fast-growing set of frontier-adjacent options without that overhead.

Examples

Command A+ (Cohere Apache 2.0), Gemma 4 (Apache 2.0), DeepSeek V4 (MIT), Mistral Medium 3.5 (modified MIT)

Sources venturebeat.com/technology/cohere-cracks-lossless-quantization-and-native-citations-with-first-full-apache-2-0-licensed-open-model-command-a

1M-token context becomes table stakes for closed frontier

Every major frontier or frontier-adjacent model shipped this month carries a 1M-token context as the default, not a premium tier. The bar moved from 256K to 1M in two quarters. Architects building RAG or long-document workflows should stop treating retrieval chunking as the default — single-shot whole-corpus inference is now economically viable for many workloads. Mid-tier models still at 128K-256K (Command A+ at 128K input) are now the exception that needs justification.

Examples

Gemini 3.5 Flash (1M), Qwen3.7-Max (1M), DeepSeek V4 Pro/Flash (1M), Claude Opus 4.7 (1M), GPT-5.5 Pro (1M)

Sources benchlm.ai/benchmarks/artificialAnalysis, model cards across labs

Specialist computer-use models challenge generalist agents on UI tasks

Fara1.5-27B's 72% Online-Mind2Web beats OpenAI's Operator (58.3%) and Gemini 2.5 Computer Use (57.3%) — purpose-built browser agents on fine-tuned small bases now outperform frontier generalists on web-task completion. Operators paying per-task on closed Operator / Computer Use APIs should run a TCO comparison against a self-hosted Fara1.5 fleet. The pattern: specialist fine-tunes on focused data beat broad-but-shallow frontier generalists on narrow real-world workflows.

Examples

Fara1.5-27B (72% Online-Mind2Web), Yutori Navigator n1 (64.7%), OpenAI Operator (58.3%), Gemini 2.5 Computer Use (57.3%)

Sources decrypt.co/368807/microsoft-fara15-open-source-ai-beats-openai-gemini, microsoft.com/en-us/research/articles/fara1-5-computer-use-agent

Artificial Analysis Intelligence Index (snapshot May 21, 2026)

Top remains GPT-5.5 (60.2); Qwen3.7-Max debuts at 56.6 — first Chinese model in global top 5, ahead of Gemini 3.5 Flash (55.3); Claude Opus 4.7 holds 57.3.

GPT-5.5 (xhigh)

60.2

Claude Opus 4.7 (Adaptive)

57.3

Gemini 3.1 Pro Preview

57.2

Qwen3.7-Max

56.6

Gemini 3.5 Flash (high)

55.3

Sources artificialanalysis.ai/articles/cohere-launches-open-weights-model-command-a

Terminal-Bench 2.1 (agentic terminal coding)

Gemini 3.5 Flash jumps to 76.2% — outscores Claude Opus 4.7 (66.1%), Gemini 3.1 Pro (70.3%), and Gemini 3 Flash (58.0%); GPT-5.5 leads at 78.2%.

GPT-5.5

78.2

Gemini 3.5 Flash

76.2

Gemini 3.1 Pro

70.3

Claude Opus 4.7

66.1

Gemini 3 Flash

58.0

Sources deepmind.google/models/gemini/flash

Online-Mind2Web (browser computer-use, 300 tasks across 136 live sites)

Open-weight Fara1.5-27B sets new SOTA at 72.0%, beating OpenAI Operator (58.3%), Gemini 2.5 Computer Use (57.3%), and Yutori Navigator n1 (64.7%).

Fara1.5-27B

72.0

Yutori Navigator n1

64.7

Fara1.5-9B

63.4

OpenAI Operator

58.3

Gemini 2.5 Computer Use

57.3

Sources microsoft.com/en-us/research/articles/fara1-5-computer-use-agent

GDPval-AA (economically valuable knowledge work, Elo)

Gemini 3.5 Flash reaches 1656 Elo — leapfrogging Gemini 3.1 Pro (1314) and Gemini 3 Flash (1204); Claude Opus 4.7 still leads enterprise reasoning at 1753.

GPT-5.5

1769

Claude Opus 4.7

1753

Claude Sonnet 4.6

1676

Gemini 3.5 Flash

1656

Gemini 3.1 Pro

1314

Sources deepmind.google/models/gemini/flash

τ²-Bench Telecom (agentic tool use)

Command A+ jumps from 37% (Command A Reasoning) to 85% — a 48-point generation-over-generation gain on open weights.

Mistral Medium 3.5

91.4

Command A+

85.0

Command A Reasoning

37.0

Sources cohere.com/blog/command-a-plus

2026-05-19 · Google

Gemini 3.5 Flash + Omni Flash + Antigravity 2.0 + Managed Agents API at I/O — three-month product cycle compressed into one keynote

Google shifted the conversation from 'model choice' to 'agent platform choice'. Procurement teams should reopen Gemini-vs-incumbent bake-offs that closed in Q1; the agent-harness lock-in (AGENTS.md, SKILL.md, managed Linux sandboxes) is now the strategic battleground, not the model itself.

Sources blog.google/innovation-and-ai/technology/developers-tools/google-io-2026-developer-highlights

2026-05-20 · NVIDIA

Q1 FY27 record $81.6B revenue (+85% YoY); Vera Rubin + Vera CPU + BlueField-4 STX announced; Vera CPUs already shipping to OpenAI, Oracle, Anthropic, xAI

Data Center revenue $75.2B (+92% YoY) confirms the buildout has not yet peaked; Vera Rubin Q3 2026 ship date with a purpose-built agentic-AI CPU reframes 2H training capacity. CFOs and infrastructure architects should assume frontier training compute remains supply-constrained through year-end and budget GPU access on Q3-availability cadence, not on price.

Sources investor.nvidia.com/news/press-release-details/2026/NVIDIA-Announces-Financial-Results-for-First-Quarter-Fiscal-2027

2026-05-19 · Anthropic

Claude Managed Agents add self-hosted sandboxes (public beta) and MCP tunnels (research preview) — agent execution moves into customer infrastructure boundary

Anthropic's response to I/O wasn't a new model — it was an enterprise security control surface. Security and platform teams should evaluate this as the new baseline for agent-deployment threat models: tool execution stays on-prem, credentials never enter agent context, only outbound traffic. Buyers that previously gated Claude agents on data-leakage risk now have a clear path forward.

Sources claude.com/fr/blog/claude-managed-agents-updates

2026-05-21 · OpenAI

Codex 'Goal mode' GA, macOS Appshots, locked-Mac computer use, browser annotations + Dell hybrid deployment partnership (May 18-19)

OpenAI did not counter-launch a model into I/O week — instead doubled down on Codex as the front door for enterprise on-prem (Dell AI Factory) and pushed agent persistence (Goal mode, locked computer use). Engineering leaders evaluating coding-agent vendors should now treat Codex as a multi-surface platform (CLI + IDE + mobile + macOS app + Windows incoming) rather than a chat feature, and re-evaluate hybrid deployment paths.

Sources help.openai.com/en/articles/6825453-chatgpt-release-notes

2026-05-20 · Alibaba

Alibaba Cloud Summit: Qwen3.7-Max + Qwen3.7-Plus-Preview + homegrown Zhenwu M890 AI chip (3x predecessor) + rebuilt full-stack agentic cloud platform

Alibaba projects model & application services ARR of ¥10B (~$1.4B) in the June quarter and ¥30B (~$4.1B) by year-end — a credible commercial AI business at scale. Buyers in APAC or with China-region data residency requirements should treat Alibaba as a primary frontier vendor evaluation this quarter, not a regional alternative.

Sources alibabacloud.com/blog/alibaba-unveils-new-ai-chip-flagship-model-and-rebuilt-cloud-stack-ai-for-agentic-era

2026-05-20 · Cohere

Command A+ shipped under Apache 2.0 — first ever fully-permissive license from Cohere on a frontier-adjacent MoE

Sovereign-AI and data-residency buyers now have a Western on-prem frontier-adjacent option without license restrictions. Strategic signal: Cohere is positioning as the Western counter to Chinese open-weights labs (DeepSeek, Moonshot, Z.ai) for regulated enterprises that cannot use Chinese weights. Procurement teams should test Command A+ vs Mistral Medium 3.5 for RAG and tool-use workloads inside the next two sprints.

Sources cohere.com/blog/command-a-plus

May 25 - 30

Anthropic round close — final terms, lead investor, full investor list

Bloomberg May 22 reported close 'as soon as next week' at $30B-plus / $900B-plus. Resets the frontier-lab valuation curve 15x in 14 months. A slip past end-May or a haircut below $700B would be the loudest re-rating signal in a year.

May 27 - 28

Samsung-union ratification vote on tentative HBM4 agreement

Binary outcome. Approval keeps the 18-day walkout off the table and HBM4 supply for Vera Rubin intact; rejection re-activates the W20 supply-risk thesis on three days' notice.

May 26 - June 13

Anthropic Opus 4.8 — leaked partner evaluations, expected I/O response

Multiple sources reporting select Anthropic partners running Opus 4.8 internal evals; Anthropic's I/O-week response was infrastructure (sandboxes, MCP tunnels), not a model. The model itself likely lands within 3 weeks based on Opus 4.6 to 4.7 cadence.

June 2 - 3

Microsoft Build 2026 (San Francisco) — MAI-line announcements, Foundry updates, Fara1.5 enterprise rollout

Build was pushed from May to June 2-3 — Microsoft has been notably quiet on its own MAI model line through I/O week. Build is where MAI frontier models, Copilot agent stack moves, and Foundry consolidation are likely to land; procurement teams running Azure / Foundry should hold model-selection decisions until after Build.

May 26 - July 15

Gemini 3.5 Pro launch — frontier intelligence rematch

Google promised Pro 'next month' (June 2026). This is the model that determines whether Google retakes the AA Intelligence Index lead from GPT-5.5 (currently 60.2) or stays on the second podium tier. Buyers planning Pro-tier procurement should hold pending Pro pricing and benchmarks.

May 26 - July 31

DeepSeek R3 / V4-Thinking — overdue per W20 watchlist

DeepSeek V4 Pro/Flash landed April 24 with strong benchmarks but explicitly without R3-style deep reasoning weights. Cohere CEO and DeepSeek itself have publicly acknowledged being 3-6 months behind US frontier — an R3 release is the catalyst that could close that gap or confirm it.

June 9 - 13

Apple WWDC 2026 — on-device model strategy, Apple Intelligence v2

Apple absent from the model-launch cycle through Q2; WWDC is the only major Apple model-disclosure window in 1H. Mobile / edge architects building on-device AI should expect significant Apple Intelligence repositioning and possibly partnered closed-model integration changes.

• First fully Saturday-cadence Pulse. Window: Mon May 18 - Sat May 23, 2026 (full ISO Mon-Sat catching the I/O / NVIDIA-earnings / Alibaba Cloud Summit / Anthropic-round news cycle).
• 6 model rows added across 6 vendors (Google DeepMind 2x, Alibaba, Cohere, Microsoft Research, Tencent) — densest model-layer week of Q2.
• Scorecard updated: closed-frontier intelligence crown unchanged but Qwen3.7-Max joins the AA Index top 5 (first Chinese model); multimodal tier consolidates around Google Omni Flash; edge / small consolidates around specialists.