The Last 3 Months in AI: The Innovations That Are Changing the Technology Landscape

Three months. That's all it took for the AI industry to ship a new frontier model family from OpenAI, three new Claude generations from Anthropic, a fully multimodal Gemini line from Google, an agent-first pivot at Microsoft's two biggest conferences, a Qwen-trained model from Meta, a new Grok from xAI, and enough Alibaba downloads to overtake Llama as the world's most-used open model family.

If you stopped paying close attention for a quarter, the landscape underneath you has genuinely moved. This is the recap — what actually shipped, why it matters, and what it means for the skills you should be building right now.

What Changed in the Last Three Months?

The headline shift isn't any single model — it's the direction every major lab moved in at the same time:

Every company covered in this post shipped something that fits at least one of those four shifts in the last quarter alone.

Company-by-Company: What Actually Shipped

OpenAI — GPT-5.6 and GPT-Live

OpenAI publicly released the GPT-5.6 family on July 9, 2026, following an earlier limited rollout to government partners for safety evaluation. It ships as three variants:

OpenAI is calling Sol 54% more token-efficient on coding tasks than its predecessor, and positions GPT-5.6 as its strongest cybersecurity model yet — achieving frontier-level performance with meaningfully fewer tokens per task. A day before, OpenAI also introduced GPT-Live, a new generation of voice models built for real-time conversational use.

Anthropic — Fable 5, Sonnet 5, and Opus 5

Anthropic had the busiest quarter of any lab covered here, shipping three model generations back to back:

• Fable 5 (June 9) — the introduction of Anthropic's Mythos-class models, its most capable reasoning tier
• Sonnet 5 (June 30) — the balanced, general-purpose model for day-to-day coding and knowledge work
• Opus 5 (July 24) — positioned as reaching near-Fable-5 capability at roughly half the cost, with a 1-million-token context window, up to 128K output tokens, adaptive thinking enabled by default, and a five-level effort setting for tuning speed against depth

For a deeper look at Anthropic's broader tool ecosystem — Claude Code, Model Context Protocol, Computer Use — see Anthropic in 2026: Everything Enterprise Teams Need to Know.

Google — Gemini 3.5, Gemini Omni, and Computer Use

At Google I/O 2026, Google introduced two new model lines: Gemini Omni, built to generate across any input starting with video, and Gemini 3.5, the first Gemini family combining frontier reasoning with real-world action. Gemini 3.5 Flash outperforms the earlier Gemini 3.1 Pro on several coding and agentic benchmarks, while running faster and cheaper.

Through June, Google extended this further:

• Gemini Omni Flash reached public preview via API — a natively multimodal model for building dynamic, custom video workflows
• Computer use was integrated directly into Gemini 3.5 Flash, letting developers build agents that see, reason, and act across desktop, mobile, and browser environments
• Nano Banana 2 Lite launched as Google's fastest, most cost-efficient image model

Microsoft — Agentic Windows and Microsoft IQ

Microsoft used Build 2026 (June, San Francisco) to reposition Windows itself as a platform for building and running AI agents — not just AI-assisted apps. Satya Nadella's keynote centred on agentic systems: AI that acts on your behalf without direct supervision.

Key announcements:

This builds directly on the shift Microsoft signalled at Ignite 2026 Paris earlier this year, where autonomous Copilot agents moved from preview into general availability. If you missed that recap, it's covered in full in Microsoft Ignite 2026 Paris: Every Major Announcement.

Meta — Training on Qwen, and Muse Spark 1.1

The most surprising story of the quarter: Meta is reportedly using Alibaba's Qwen to help train its next model, code-named Avocado — a genuine role reversal for a company that built its reputation on open-sourcing Llama. Meta also shipped Muse Spark 1.1 on July 9, 2026, a smaller update alongside the bigger strategic shift.

xAI — Grok 4.5

xAI released Grok 4.5 on July 8, 2026, continuing its rapid release cadence and tighter integration with the X platform for real-time data grounding.

Alibaba — Qwen Overtakes Llama

Alibaba had arguably the biggest structural win of the quarter without shipping a flashy keynote. Qwen3.7-Max launched May 20, 2026 as a proprietary, API-only model via Alibaba Cloud Model Studio. More importantly, the Qwen model family crossed one billion cumulative downloads, officially overtaking Meta's Llama as the world's most widely used open-source AI model family — a milestone that says as much about where open-source AI development has shifted as any single model release does.

The Snapshot Table

Why This Matters — The Real Signal Behind the Releases

AI Is No Longer Just "Generating Text"

Gemini Omni working from video input, Computer Use agents operating a browser, GitHub Copilot's agentic browsing — these all point to the same shift: models are increasingly expected to reason across data types and take action inside software, not just produce a written response to a written question.

Agents Moved From Demo to Default

A year ago, "AI agent" mostly meant a conference demo. This quarter, Microsoft Agent Framework 1.0 became a commercially supported SDK, Ignite-era Copilot agents reached GA, and Google shipped Computer Use directly into a broadly available model tier. Autonomous, multi-step AI execution is now a procurement line item, not a research preview.

Frontier Intelligence Is Getting Cheaper, Fast

Claude Opus 5 delivering near-Fable-5 reasoning at half the cost, and GPT-5.6 Sol being 54% more token-efficient at coding, are the same story told twice: the gap between "the best model" and "an affordable model" is closing quarter over quarter. That changes the calculus for which model a team should default to for routine work.

The Open-Source Centre of Gravity Has Shifted

Qwen overtaking Llama in cumulative downloads — and Meta reportedly training its next model with Qwen's help — is a genuinely notable shift. Two years ago, Llama was the default answer to "what's the best open-weight model." That answer is no longer obvious.

Skills to Develop to Stay Relevant

The specific model names in this post will be outdated within a year. The underlying skills won't be.

For the broader argument on why fundamentals matter more than any specific tool, see The AI Revolution: Learn Skills, Not Just AI Tools.

What to Watch Next

A few threads from this quarter are worth tracking as they develop:

• Whether Meta's Qwen-assisted "Avocado" model signals a broader trend of Western labs incorporating Chinese open-weight research
• How quickly Computer Use and agentic browsing capabilities move from "impressive demo" to reliable, audited enterprise workflows
• Whether the cost-down trend (Opus 5, GPT-5.6 Sol) continues fast enough to make frontier-tier reasoning a default rather than a premium tier
• How enterprise governance frameworks adapt now that autonomous agents are GA rather than preview features

The Bottom Line

Three months, seven companies, and a landscape that looks meaningfully different than it did in May. The specific models covered here will be superseded — probably within another three months, if this pace holds. What won't change as fast: the shift toward multimodal reasoning, autonomous agents as standard tooling, and intelligence getting cheaper every quarter.

Build your skills around that direction, not around any single model name — and check back in another three months, because at this pace, this list will need an update.