Tari Ibaba

This is just unbelievable.

Imagine having a model that’s:

• More than 5 times cheaper than Claude Opus — yet just as intelligent?
• Blazing fast — up to 15x faster than Claude Opus for heavy prompts and massive context?
• Open-source with open weights to top it all off?

Yeah it’s definitely no surprise how so many people have been going absolutely wild over the few days over the new MiniMax M3.

Look how closely it matches Claude’s and GPT’s capabilities! It’s even better in certain key areas:

I still can’t wrap my head around the absolutely wild feats of endurance and intelligent it displayed during testing — and has been performing ever since then.

This just truly completely shattered the limits of what we thought was possible with open-source models — or any type of model really.

1. Not the same “1 million token context”…

Don’t let AI companies fool you with their loud claims of 1 million token context.

Not all 1 million token contexts are remotely the same — not even close.

Yes they can all process massive codebases, research archives, books, or long-running projects.

But they vastly differ in the quality of processing all that massive data, how much they can actually make sense of, how much they comprehend the interconnectedness of all the data.

And speed. A big, big one.

Most AI models become painfully slow and expensive when dealing with that much information.

M3 uses an innovative new approach called MiniMax Sparse Attention (MSA) that allows it to focus on the important parts instead of wasting resources on everything at once.

This allows M3 to:

• Read huge amounts of information more than 9x faster
• Generate responses more than 15x faster
• Use a fraction of the computing power normally required

The giant context window is no longer just for fancy — it actually becomes a beast in the real world.

And those models that still use slower methods of parsing large token contexts will get absolutely left in the dust.

2. Insane long-horizon ability

MiniMax M3 shocked every with insane feats of autonomy on tasks spanning several hours.

The 12-hour research paper loop: M3 was given a complex AI research paper and spent 12 hours straight reproducing the core experiments. It managed its own context window, parsed charts, wrote code, handled 18 GitHub commits, and generated 23 experimental figures completely unaided.

Self-training models: In another test, M3 was given raw base models and spent 12 hours handling data synthesis, training, and evaluation loops on them entirely on its own.

The 24-hour CUDA optimization: It was assigned to optimize an FP8 GEMM kernel (a notoriously painful low-level GPU operation). Over 24 hours, it autonomously went through 147 benchmark submissions and nearly 2,000 tool calls, boosting hardware peak utilization from a lousy 7.6% to a highly optimized 71.3% — a 9.4x speedup with zero human help.

This is the kind of work you’d normally assign to a high-level engineer or researcher, yet here comes MiniMax M3 doing it all on its own, unbelievable.

3. Punching way above its weight

Despite competing against much larger companies with way bigger resources at their disposal, M3 is posting amazing benchmark results.

• SWE-Bench Pro: 59.0%, edging past GPT-5.5 and Gemini 3.1 Pro
• BrowseComp: 83.5, outperforming Claude Opus 4.7’s 79.3

MiniMax is competing right at the cutting edge, particularly in software engineering and autonomous research tasks.

4. Ultra-competitive pricing

It will undoubtedly be one of the biggest selling points.

All that intelligence and speed and innovation, for such a ridiculously low price.

M3 launched at roughly:

• $0.60 per million input tokens
• $2.40 per million output tokens

That makes it significantly cheaper than many of the most advanced AI models available today.

Just compare to the latest Claude Opus pricing:

• $5.00 per million input tokens
• $25.00 per million output tokens

This will be huge for startups and smaller teams.

Many AI projects fail not because the technology isn’t good enough, but because running them becomes too expensive. Lower costs mean more companies can build products that were previously out of reach.

5. Open weights available to everyone

MiniMax isn’t keeping everything locked behind a proprietary wall.

And open-source availability means:

• Greater transparency
• Community-driven improvements
• Easier experimentation
• More deployment flexibility
• Reduced vendor lock-in

Open models used to trail the very best closed systems — not anymore.

M3 is part of the new wave of the models that completely disrupts the notion of open-source models being inherently weaker.

Why M3 matters

The biggest takeaway from M3 isn’t that it’s smarter than every other model.

It’s that MiniMax is attacking some of the biggest problems in AI at the same time:

• Cost
• Speed
• Long-term autonomy
• Multimodal understanding
• Accessibility through open source

If these early results hold up in real-world use, M3 could become one of the most important AI launches of the year — not because it’s the biggest model, but because it makes powerful AI far more practical for everyone.

Claude Opus 4.8 is absolutely insane.

This is going to fundamentally transform the AI coding ecosystem and the sheer amount of tasks we entrust agents with. You absolutely cannot ignore this.

Claude Opus 4.8 stays perfectly on track across extremely long complex coding sessions:

Like it’s not just way more powerful than Opus 4.7 with so many abilities massively upgraded —

It’s also a much more HONEST model:

It can actually admit when it has no idea what something means — instead of making up nonsense on the fly.

It’s literally FOUR TIMES more likely to catch any potential bugs and flaws in any code it generates.

All thanks to a major breakthrough in AI self-awareness and reliability.

And don’t get me started on the new Dynamic Workflows that turn any Claude-powered agent into a super-powered beast.

It even easily passed the tricky car wash test that’s frustrated so many great models — which shows a serious jump in real-world understanding and logic — something very important for AGI.

1. A massive leap in AI integrity & reliability

One of the biggest problems with modern AI is that models often don’t know when they’re wrong.

They write code with bugs, reach conclusions on incomplete information, and confidently report success even when they’ve failed.

Opus 4.8 was designed specifically to fix this.

The model proactively identifies gaps in its own reasoning instead of rushing toward conclusions…

The model is far more likely to identify uncertainty, flag gaps in its reasoning, and challenge questionable assumptions before proceeding.

• Admitting uncertainty: The model proactively identifies gaps in its own reasoning instead of rushing toward conclusions.
• Catching bugs: Opus 4.8 is 4× less likely than Opus 4.7 to let flaws in code it generated pass without comment.
• Reporting failures honestly: The model achieved a perfect 0% rate of uncritically reporting flawed results in Anthropic’s internal evaluations.
• Pushing back when necessary: Rather than blindly following instructions into failure, it actively questions weak assumptions and incomplete plans.

As AI systems become increasingly autonomous, reliability starts to matter more than raw intelligence.

A model that catches its own mistakes can end up being much more valuable than one that’s simply smarter.

2. Dynamic Workflows change everything

The second major announcement is Dynamic Workflows.

Instead of tackling large projects as a single agent, Opus 4.8 can act as an orchestrator, spawning and managing hundreds of parallel AI subagents simultaneously.

With Dynamic Workflows, Claude can:

• Write orchestration scripts
• Launch hundreds of specialized AI agents
• Distribute work across large codebases
• Verify outputs against existing test suites
• Aggregate results into a final solution

Anthropic showcased this capability by using Dynamic Workflows to help port the Bun runtime from Zig to Rust — a project involving roughly 750,000 lines of code.

Using a swarm of coordinated agents, Claude completed the migration in 11 days and achieved a remarkable 99.8% pass rate on Bun’s existing test suite.

That’s not a chatbot helping write functions.

That’s an AI system coordinating work at the scale of an engineering organization.

3. Completely ridiculous intelligence gains

Unbelievable gains in practically every major benchmark.

But the most eye-catching result comes from USAMO 2026 — one of the world’s most difficult mathematics competitions.

Opus 4.7 scored 69.3.

Opus 4.8 scored 96.7.

That’s NOT normal.

But it didn’t stop there.

• USAMO 2026: 69.3 → 96.7
• SWE-Bench Pro: 64.3 → 69.2
• GPQA Diamond: 84.1 → 88.4
• Humanity’s Last Exam: 35.8 → 44.7
• AIME 2025: 79.4 → 88.9
• GDPval: 1753 → 1890

The pattern is clear: this wasn’t a model that got better at one thing — it got better at almost everything:

• Reasoning
• Coding
• Mathematics
• Scientific problem-solving
• Knowledge work
• Agentic tasks

Virtually every category moved forward — and many of the jumps are unusually large by frontier-model standards.

4. You can now control how hard Claude thinks

Opus 4.8 also introduces manual effort controls across Claude.ai, Claude Code, and the API.

Users can now decide how much compute the model should spend on a problem:

• Low effort for speed
• Medium effort for balanced performance
• High effort for difficult reasoning and coding tasks

This gives users direct control over one of AI’s most important tradeoffs: speed versus intelligence.

5. Fast Mode is faster — and cheaper

Anthropic also upgraded Fast Mode.

The new mode delivers:

• 2.5× faster token generation
• 3× lower cost than previous Claude fast modes

That matters because many AI applications today aren’t limited by intelligence—they’re limited by latency and cost.

Reducing both simultaneously makes entirely new categories of AI products economically viable.

Claude Opus 4.8 isn’t just more intelligent. It’s much more trustworthy now.

The benchmark gains are impressive. Dynamic Workflows is exciting. Fast Mode and effort controls are useful.

But the most important thing Anthropic may have accomplished is teaching a frontier model to recognize when it might be wrong.

As AI starts automating more and more crucial portions of our workflows, this starts to matter more than any benchmark score.

This is something completely different —

Qwen 3.7 Max is going to make such a massive impact on the workflow of so many developers:

This wasn’t launched with the usual displays of fancy benchmarks and claims of better reasoning — even though there’s a LOT to be said about that — almost as good as top Claude models, yet so much cheaper and faster.

Even better than the new Gemini 3.5 — and GPT-5.5 highest thinking mode!

But Qwen 3.7 Max was made specially to address one of the biggest “holy-grail” challenges facing AI agents today:

How do you keep an AI focused for hours without it losing track of its objective?

And from what I’ve been seeing these past few days — it’s been a game-changing success.

1. The shocking 35-hour test

The most impressive thing about Qwen 3.7 Max isn’t a leaderboard score:

It’s endurance.

One of the biggest weaknesses of modern AI systems is context drift—the tendency to lose track of instructions, goals, or prior decisions during long workflows.

Alibaba tested this directly by assigning Qwen 3.7 Max a highly specialized engineering challenge: optimizing a GPU attention kernel it had never seen before.

Then they left it running.

Over roughly 35 consecutive hours, the model:

• Made 1,158 tool calls
• Ran 432 code evaluations
• Diagnosed its own compilation failures
• Tested multiple optimization strategies
• Achieved a reported 10× performance improvement

This Max model maintained coherence and direction throughout the entire process—a critical capability for autonomous software engineering agents.

2. New 1 million token context window

Qwen 3.7 Max also now supports a 1 million token context window to support this incredible long-running ability.

That allows developers to work with:

• Entire code repositories
• Large documentation libraries
• Research archives
• Long project histories
• Multi-day agent memory

in a single context.

The benefit isn’t just memory. It means less summarization, less context compression, and greater continuity across complex workflows.

3. A smart ecosystem strategy

One of the most underrated features is that Qwen 3.7 Max natively supports the Anthropic API protocol.

This means developers can potentially plug it into existing tools such as:

• Claude Code
• OpenClaw
• Anthropic-based agent frameworks
• Internal Claude-compatible enterprise tooling

without rebuilding their entire stack.

Instead of forcing developers into a new ecosystem, Alibaba is reducing switching costs and making adoption easier.

4. Fast, capable, and cheap

Reasoning models are typically slow.

Qwen 3.7 Max reportedly outputs at around 192 tokens per second, significantly faster than many reasoning-focused models that operate closer to 60–70 tokens per second.

Speed matters because agents make hundreds or thousands of decisions. Small latency improvements compound into major productivity gains.

The model also ranks among the top tier of reasoning systems on independent leaderboards while delivering strong performance on difficult benchmarks such as GPQA Diamond and advanced math evaluations.

Perhaps most disruptive is the pricing:

• $2.50 per million input tokens
• $7.50 per million output tokens
• Native prompt caching that can reduce repeated-context costs by up to 90%

For companies building AI agents, that dramatically improves the economics of long-running workflows.

Many AI coding discussions are still focused on the raw intelligence of the models.

But the next wave of competition is about:

• Autonomous software engineering
• Long-horizon planning
• Tool use
• Workflow execution
• Agent reliability

So Qwen 3.7 Max isn’t just another model release, it’s a superior platform for AI agents.

Its defining characteristics aren’t just intelligence—they’re endurance, context, speed, compatibility, and cost efficiency.

And moving forward, these things are going to be far more important than another benchmark victory.

Google just released Antigravity CLI — an upgraded version of Gemini CLI that’s so much more powerful at software development.

Packed with amazing new features and ecosystem integrations, several times faster than Claude Code, and comes with brand new commands for various essential software development tasks.

The new Antigravity CLI in action:

They completely rebuilt the terminal experience around the new way software engineering works in 2026 — multi-agent, asynchronous, persistent, and increasingly autonomous.

From a regular chatbot sitting in a terminal — to a orchestration system for advanced engineering workflows.

1. Dynamic asynchronous subagents change everything

Antigravity CLI uses sophisticated divide-and-conquer — to radically improve performance and code quality for the most complicated tasks.

Main agent delegating a sub-agent to a task in Antigravity CLI:

Many coding assistants still work like linear chatbots. You ask for a change and wait while one agent handles everything sequentially.

Antigravity CLI changes that completely.

The orchestrator automatically breaks large goals into smaller tasks, spawns specialized subagents dynamically, and executes them simultaneously in the background.

You move from working with one assistant to directing an engineering team.

The context advantage is huge too.

Each subagent operates inside its own isolated context window, keeping compile logs, installs, test output, and intermediate work away from the orchestrator context.

This massively improve larger workflows for us developers.

Imagine upgrading frameworks while simultaneously migrating APIs, updating components, generating tests, and validating documentation.

Large refactors stop feeling like events.

2. Built for pure speed with Gemini 3.5 Flash

Gemini 3.5 Flash is much faster than other top-tier models:

Google optimized Antigravity CLI around Gemini 3.5 Flash, with benchmark figures placing output speed around 289 tokens per second — as much as 4 times faster than major competing tools like Claude Code.

That matters because latency compounds in agent systems.

A slow model creates slow planning, slow subagents, slow testing, slow validation, and slower iteration loops.

Everything stacks.

And Google really appears to have optimized specifically for fast orchestration.

This makes multi-file edits, codebase analysis, documentation generation, and architecture migrations feel dramatically faster.

The shorter the loop becomes, the easier iteration becomes.

And iteration drives quality.

3. Persistent isolated local environments

Antigravity CLI also introduces persistent isolated environments.

Instead of constantly rebuilding state every interaction like before, Antigravity agents can preserve files, execution state, and progress across sessions.

This enables workflows like:

• Long-running builds
• Persistent testing environments
• Multi-stage debugging
• Extended implementation sessions

For developers working on larger systems, this could remove huge amounts of setup overhead.

Authentication migrations, distributed debugging sessions, or long validation workflows no longer need constant reconstruction.

Work continues.

State survives.

4. Brand new slash commands

The command system reveals how Google expects developers to use Antigravity.

Not chat-first —

Workflow-first.

/goal tells the system to autonomously execute large objectives without repeatedly asking permission.

Think:

“Upgrade React version.”

“Refactor auth.”

“Migrate architecture.”

Then there is /grill-me, which forces the agent to interrogate you before implementation begins.

It asks about edge cases, requirements, logic gaps, and architecture decisions.

That could prevent a surprising number of specification failures.

Finally /browser gives the agent browser access for testing interfaces and validating outputs.

Frontend workflows start looking much closer to autonomous QA.

5. Deep integration with Antigravity 2.0 and AI Studio

Antigravity CLI shares the same underlying agent harness as Antigravity 2.0 desktop and Google AI Studio.

Desktop, terminal, SDK, and mobile workflows are becoming interfaces over the same runtime.

Imagine walking somewhere, opening AI Studio, speaking through a feature idea, exporting it, then sitting down later and continuing instantly inside agy.

Same state.

Same memory.

Same agents.

Ideas move directly into execution.

6. Live voice transcription inside the terminal

Antigravity CLI embeds Gemini audio models directly into terminal workflows.

You can literally ramble:

“Build dashboard… PostgreSQL… auth… analytics…”

The system transcribes and refines everything into executable objectives automatically.

This could become surprisingly useful during planning and architecture work.

A lot of design thinking happens faster through speech than typing.

Antigravity turns that directly into execution.

And that may be the bigger story.

Less prompting — more directing.

Wow, Google is so not messing around.

Now they just unleashed a major new competitor to OpenClaw — Gemini Spark is going to totally transform how so many people work and compute from now on.

And always-on AI system — always operating continuously in the background.

Seamlessly connects to every single one of your tools to work on any of your goals.

Intelligently adapts to your workflows and style.

Powered by the most advanced AI models on the planet.

This is definitely going to be huge.

1. Always on: set and forget

One of Spark’s biggest selling points is the always-on background execution.

Most AI products today still behave like sessions (one of the main reasons why OpenClaw blew up so much).

You ask → It responds → The interaction ends

Spark changes that model by allowing workflows to continue running after the conversation is over. Google positions it as an agent capable of monitoring, organizing, and acting in the background instead of waiting for repeated prompts.

That means less time restarting the same processes every day.

You can literally schedule anything at any time you want.

Think about workflows like:

• Monitoring emails and surfacing important updates
• Tracking ongoing projects and deadlines
• Watching dashboards or reports for changes
• Running recurring research and summaries

The value here isn’t raw speed.

It’s continuity — working on complex tasks endlessly, so you can get to more important things.

2. Teachable reusable Skills

Another major idea behind Spark is teachable Skills.

Instead of rebuilding workflows repeatedly, Spark is designed to retain context and adapt over time.

Repeated instructions → Reusable workflows

Over time, systems can begin reflecting how users naturally operate:

• Preferred writing style and tone
• Reporting structures
• Research methods
• Team workflows
• Content frameworks

The goal is simple:

Teach once.

Reuse repeatedly.

As more behavior becomes reusable, the distance between intention and execution keeps shrinking.

3. Native connections instead of screen scraping

Many AI agents today still rely on browser automation and analyzing screen pixels.

They:

• Observe screens
• Click buttons
• Mimic user actions

Spark moves toward native integrations instead:

• Deep Workspace integration
  • Gmail
  • Docs
  • Sheets
  • Calendar
• Third-party integrations through MCP (Model Context Protocol)

This allows workflows to operate closer to the application layer itself.

Instead of manually moving information between tools, Spark can coordinate across connected systems more directly.

The result:

Less tool switching. A more widely capable agent.

4. Control still sits with the user

Persistent AI naturally raises another question:

How much autonomy is too much?

Google is trying approach Spark with strong controls and approval layers.

The idea is not unrestricted automation.

It is controlled execution.

Spark can:

✅ Monitor
✅ Organize
✅ Draft
✅ Prepare workflows

While keeping human oversight where needed.

That balance matters because future agent systems will be judged not only by capability, but by trust.

5. Multi-agent coordination

Perhaps the most interesting part of Spark is its move toward multi-agent coordination.

Instead of one assistant handling everything, Spark points toward specialized systems working together.

Potential roles could include:

• Research agents
• Scheduling agents
• Communication agents
• Documentation agents
• Monitoring agents

Different responsibilities.

Shared objective.

The user gradually shifts from operator to coordinator.

And that changes the relationship between humans and AI entirely.

Gemini Spark represents a larger transition happening across AI:

• Sessions → Persistent workflows
• Prompts → Systems
• Single assistants → Agent coordination
• Manual execution → Continuous operation

Will be very interesting to see how all this evolves.

Google is doing something totally mind blowing.

They are trying to kill the IDE. For good.

Many people were rightly expecting their Antigravity tool to get some upgrades in this year’s Google I/O — but they had absolutely no idea just how drastic the changes were going to be.

They went from this — what we’ve all been used to for several decades:

to this — literally just this:

And then later on when you start doing some work:

Google can see the writing on the wall. They understand the unstoppable direction the AI software development landscape is heading in.

The coding editor is totally gone.

Something not everyone is very happy about:

A complete transformation from AI assistant to full-fledged AI engineering team.

1. From AI assistant → genius software engineering team

The first version of Antigravity still lived in the IDE world.

Antigravity 2.0 does not.

Google completely stripped away the editor-first experience and rebuilt the system around projects, not repositories.

Instead of opening folders and files, you manage long-running initiatives while AI agents execute the work. Google literally calls the desktop app “unabashedly agent-first.”

The mental model changes completely:

Old world:

Open repo → write code → commit

Antigravity world:

Define goal → deploy agents → orchestrate execution

You’re no longer pairing with an assistant.

You’re managing a software building organization.

2. An entire OS from scratch

This was the moment that shocked the world.

Google DeepMind gave Antigravity 2.0 a single high-level prompt and had it build a functional operating system — FROM SCRATCH.

The numbers were absurd:

• 93 subagents
• 2.6 billion tokens processed
• Completed in 12 hours
• Total cost reportedly under $1,000 in API usage

And then came the live demo.

Google loaded the classic game Doom onto the AI-generated operating system and ran it live on stage.

During the demo, Antigravity was even prompted to generate missing keyboard drivers in real time, instantly extending the OS and making gameplay possible.

Think about what that means.

A multi-agent system coordinated dozens of autonomous workers, built an operating system, compiled it, extended it live, and ran software on top of it.

This is no longer:

“Build me a React component.”

This is:

Autonomous systems engineering at massive scale.

3. Dynamic subagents are a game changer

Most AI systems break when tasks become too large.

Throw enough complexity at a single model and eventually it loses context, forgets state, or collapses under memory pressure.

Google solved this by turning the primary agent into a manager.

Instead of doing everything itself, it dynamically spawns specialized subagents in parallel.

Those agents work independently, finish subtasks, then merge results back into the main workflow. Google highlighted Antigravity’s move toward coordinated multi-agent execution and orchestration across environments.

Like just imagine this beautiful teamwork:

Agent A → architecture

Agent B → UI

Agent C → tests

Agent D → browser validation

Agent E → deployment

All running simultaneously.

The result:

Massive projects stop feeling like giant prompts.

They become distributed systems.

4. Scheduled Tasks keeps AI working on autopilot

One of the most underrated announcements was Scheduled Tasks.

This changes how people interact with AI.

Instead of sitting at your laptop prompting continuously or hacking something together, you can now seamlessly assign recurring work:

“Review analytics every morning at 8 AM.”

“Refactor stale code overnight.”

“Scan for regressions while I sleep.”

The agents run asynchronously in the background without interrupting active workflows.

This moves AI from:

tool you use → system that operates

5. The slash commands reveal what Google thinks AI is actually good at

Google introduced several new command patterns that feel surprisingly honest about AI limitations.

/grill-me

Before touching anything, the agent interrogates you.

It keeps asking clarifying questions until intent becomes crystal clear.

Google is effectively admitting:

Bad prompts create bad autonomous systems.

/goal

This is the opposite.

Full autonomous mode.

Give the objective.

Walk away.

The agent executes until completion without constantly asking permission.

It’s essentially “fire-and-forget engineering.”

/browser

This one is fascinating.

Google openly acknowledged that agents still struggle with deciding when they should search externally.

So instead of pretending the problem is solved, they added an explicit switch:

Turn browsing on.

Give the agent web access.

Continue.

That transparency is rare.

6. Gemini 3.5 Flash is the engine underneath all of this

None of this works without speed.

Antigravity 2.0 runs on Gemini 3.5 Flash, which Google positioned as the model built specifically for agent workflows, coding, and long-horizon tasks.

Google says it delivers major coding improvements and operates 4× faster than competing frontier models, while Antigravity pushes even more performance inside its environment.

Speed matters even more now in this multi-agent era because agents multiply work.

One slow model = waiting.

Ninety-three agents = bottleneck disaster.

3.5 Flash changes the equation.

Antigravity 2.0 isn’t trying to beat Cursor.

It isn’t trying to beat Copilot.

Google is aiming at something bigger:

Replace the software workflow itself.

The IDE is disappearing.

Repositories are becoming projects.

Developers are becoming orchestrators.

And AI is turning into a software engineering organization.

Wow this is absolutely huge.

Gemini 3.5 Flash is here and it’s a total MONSTER — this is so much more than just another minor upgrade.

This thing is literally FOUR times faster than literally any other model in existence right now.

So much faster than the last version of Flash — yet somehow so much better than the last Pro version in practically every way.

And it goes way way beyond just speed or coding accuracy improvements.

We have now fast moving past the era of one agent doing all the work on its own.

Gemini 3.5 was built from the ground up for multi-agent coding and massive parallelism.

It even now generates sophisticated interactive UIs and high-resolution graphics and demos from your prompts.

1. Massive upgrades in agentic coding, terminal workflows, and software engineering capabilities

The biggest improvement in Gemini 3.5 is software engineering capability.

Gemini 3.5 comes with major upgrades in coding, long-horizon execution, and agent benchmarks — despite Flash traditionally being the speed tier.

The gap is so large now that Gemini 3.5 Flash is effectively replacing what developers previously expected from Pro-class reasoning models.

What stands out is its strength in agentic software workflows:

• Multi-file repository changes
• Long-running engineering tasks
• Terminal-based execution flows
• Sustained planning and execution loops
• Autonomous code modification workflows

Terminal capability especially matters because modern engineering increasingly happens outside the editor:

Git workflows, infrastructure changes, deployments, Docker environments, CI pipelines, Kubernetes operations — these are where many models still degrade.

Gemini 3.5 is much stronger in those environments, particularly on terminal-oriented benchmarks and software execution tasks.

2. Unbelievable gains in speed — blazing fast

Performance matters.

Latency changes behavior.

Google claims Gemini 3.5 Flash reaches roughly 289 tokens/sec, putting it at around 4× the speed of comparable frontier models while maintaining much stronger coding performance.

This matters because agent systems stack latency:

Planner → Subagent → Tool call → Verification → Retry → Execution

Small delays multiply.

When a model becomes dramatically faster, entire workflows become practical:

• Parallel debugging agents
• Continuous architecture reviews
• Autonomous refactoring pipelines
• Background code validation agents
• Multi-repository execution systems

Fast AI feels good.

Very fast AI changes what can exist.

3. Built for parallel agentic workflows from day one

This may be the most important architectural shift.

Most agent systems today were originally built for:

one model → one user → one response

Parallel agents came later.

Teams hacked orchestration layers on top.

Gemini 3.5 takes the opposite approach.

Google designed it around parallel agent execution, where multiple agents can split work, coordinate tasks, and execute simultaneously.

Just imagine this engineering workflow:

Agent A handles refactoring.

Agent B updates tests.

Agent C validates infrastructure.

Agent D generates migration docs.

All running concurrently.

This is no longer a chatbot pattern.

It is distributed software execution.

And it is probably the clearest signal yet of where AI development tooling is heading.

4. World-class external tool use

Coding ability matters.

Tool usage determines whether agents are actually useful in the real-world.

Gemini 3.5 scores massive improvements on MCP Atlas, one of the strongest evaluations for sophisticated tool orchestration and external system interaction.

It beat every single major model at MCP integration and tool usage.

That matters because modern agents increasingly operate through:

• IDE integrations
• APIs
• MCP servers
• CI pipelines
• Cloud systems
• Terminals
• External execution environments

Tool use is becoming the real operating system for AI.

And Gemini 3.5 is exceptionally strong at it.

5. Rich interactive UIs and graphics generated in real time

Gemini 3.5 is also pushing hard into interface generation.

Google showcased richer interactive outputs including generated interfaces, dynamic graphics, visual experiences, and interactive applications generated on demand.

Instead of generating isolated explanations and descriptions, Gemini increasingly generates:

• Interactive interfaces
• Visual systems
• Animations
• Stateful experiences
• Rich graphics

OpenAI just completely transformed the software development ecosystem with their new Codex Chrome extension.

They just unlocked a whole new world of possibilities for what you can accomplish — now that it has access to context from your real browser.

Now it can see your active session and use all those crucial developer services you’re signed in to — which it could never do before an empty sandboxed browser.

Codex can now operate seamlessly across your codebases, repositories, APIs, browsers, and developer tools to work on the most complex goals that need data from all these different environments.

1. Powerful signed-in state access

One of the biggest advantages of the Codex Chrome extension is its access to your real signed-in browser state.

Unlike traditional web crawlers or sandboxed browsers, Codex operates inside your active Chrome profile using your existing sessions, cookies, and authenticated accounts.

Why does that matter?

Because most real work happens behind login screens.

You could tell Codex:

“@Chrome open Salesforce and update the status of my last three leads.”

And it can do it because it’s operating as you inside the browser.

This allows Codex to handle workflows like:

• Triaging emails in your actual inbox
• Updating Jira or Salesforce records
• Pulling information from internal dashboards
• Accessing private company wikis without public APIs

Most AI systems struggle with authenticated enterprise environments. The Codex extension changes that by turning the browser itself into an execution layer for AI agents.

2. Intelligent three-tier execution model

At the center of the system is dynamic tool switching.

When you ask Codex to do something, it automatically decides which tool is best — using a built-in priority stack:

Tier 1: Plugins and APIs

If a dedicated integration exists — such as GitHub, Jira, or Gmail — Codex uses the API first.

This is the fastest and most stable option because APIs provide structured, deterministic access. Instead of visually navigating interfaces, Codex can directly update tickets, review pull requests, or retrieve information.

OpenAI’s approach makes it clear that browser automation is not the default. It’s the fallback when APIs are unavailable or insufficient.

Tier 2: Chrome extension

When a task requires authenticated browser access, Codex escalates to Chrome.

This allows it to operate inside real logged-in sessions:

• Internal dashboards
• CRMs
• Admin tools
• Enterprise web apps
• Sites without usable APIs

Rather than running in an isolated sandbox, Codex works within your existing browser state, including cookies and SSO sessions. This solves one of the biggest limitations of current AI agents: most enterprise workflows happen inside web interfaces, not clean APIs.

Tier 3: In-app browser

For local development environments like localhost:3000 or static previews, Codex stays inside its own browser sandbox.

This keeps developers’ personal Chrome sessions clean while providing safer isolation for testing and debugging workflows.

Together, these three layers allow Codex to choose the least invasive and most reliable path for completing work.

3. Multi-tab background operations

One of the most interesting parts of the extension is how it handles browser automation.

Unlike traditional automation tools that take over your screen, Codex works in the background using task-specific Chrome tab groups.

This enables real parallelism.

Codex can:

• Open multiple documentation pages simultaneously
• Compare services or APIs
• Validate workflows across tabs
• Research information in parallel

And it does this while users continue working normally in their own tabs.

OpenAI also made the process reviewable. After completing a task, Codex leaves its tab groups open so users can inspect what it did, what pages it accessed, and how it reached conclusions.

That auditability could become important for enterprise adoption.

4. Integrated DevTools access

For developers, the biggest feature may be Chrome DevTools integration.

Codex can now interact directly with browser internals, giving it visibility into:

• DOM structure
• Console errors
• Network requests
• Performance traces
• Lighthouse audits

This dramatically changes debugging workflows.

For example, Codex can:

1. Detect a failing UI interaction
2. Inspect the DOM
3. Identify the underlying JavaScript or CSS issue
4. Trace the related source code
5. Suggest a fix in the IDE

Instead of relying on screenshots or copied console logs, the AI can directly observe the browser state in real time.

This pushes Codex beyond autocomplete into autonomous debugging and testing.

5. Persistent goals With /goal

The extension launched alongside a new persistent /goal system that enables longer-running autonomous tasks. (mindstudio.ai)

Rather than stopping after every action, Codex can continue working toward a high-level objective across multiple steps.

For example:

“Find the bug in the staging login flow and update the Jira ticket.”

Codex can then:

• Navigate the staging app
• Switch tabs
• Inspect DevTools
• Review logs
• Open Jira
• Update the issue
• Continue iterating until the task is complete

This persistent execution model moves Codex closer to a true software agent rather than a step-by-step assistant.

6. Human-in-the-loop security

Giving AI access to authenticated browser sessions creates obvious security risks, so OpenAI implemented permission-on-demand controls.

Codex must explicitly request permission before:

• Accessing new domains
• Reading browser history
• Uploading files
• Downloading files

Users can approve actions once, permanently allow domains, or block access entirely.

OpenAI also warns that websites should be treated as untrusted input because malicious pages could attempt prompt injection attacks against the agent.

The bigger picture

The real significance of the Codex Chrome extension is not that it can click buttons in a browser.

It’s that OpenAI is building a system capable of fluidly moving between APIs, browsers, DevTools, IDEs, and long-running workflows.

The browser is becoming more than a user interface.

It’s becoming an execution environment for AI agents.

Google AI Studio just keeps ascending toward unbelievable heights…

From a simple prompt playground for basic AI testing…

To a practical development environment for building apps, interfaces, and AI-powered products faster.

With incredible features like prompt autocomplete, visual editing, and integrated sophisticated image generation, we developers can now rapidly move from idea to prototype faster than ever.

1. Tab Tab Tab: Prompt autocomplete

The “tab tab tab” prompt autocomplete feature helps developers expand rough ideas into stronger prompts instantly. Instead of writing a detailed prompt from scratch, a developer can start with something like:

“Create a clean SaaS dashboard with analytics cards…”

AI Studio can then suggest layout details, styling direction, responsiveness, components, and user flows.

This turns prompting into something closer to code autocomplete. It speeds up brainstorming, UI generation, front-end scaffolding, and MVP creation. Developers can quickly generate a React-style structure, landing page, dashboard, or app layout, then export and customize the code further.

For solo founders and indie hackers, this is especially useful because it reduces the time spent on boilerplate HTML, CSS, and basic UI structure.

2. Design previews

Design previews let developers choose the visual direction of an app before it’s finished.

In Google AI Studio’s vibe coding experience, Gemini can now generate custom themes while your app is being created.

Within seconds, developers can compare different looks and pick the one that best fits the product: minimal, playful, premium, futuristic, enterprise-ready, or creator-focused.

For SaaS builders this means landing pages, dashboards, and MVPs no longer have to start with a generic “AI-generated” look. You can establish a stronger visual identity from the beginning, then refine the code later.

3. Edit mode and annotation

Edit mode makes transforms AI studio from a chatbot into a full-blown visual development tool.

With annotation, developers can draw directly on the app interface. They can circle a section, mark an area, or point to a component and write notes such as:

“Make this bigger,” “move this to the top,” or “reduce the spacing here.”

The AI interprets the visual instruction and updates the app accordingly.

This is a major improvement because many UI changes are easier to show than explain. Instead of writing long prompts to describe a design problem, developers can communicate visually.

This brings AI Studio closer to tools like Figma, but with code generation and AI assistance built in.

4. Integrated image generation with Nano Banana

Nano Banana integration solves one of the most common developer problems: creating visual assets.

AI Studio can now generate custom images, logos, icons, illustrations, and UI graphics while the app is being built. This removes the need to search for placeholder images, icon packs, or temporary “programmer art.”

Even better, the generated assets can maintain a consistent aesthetic across the project. Colors, style, tone, and visual language can remain aligned from the landing page to icons and illustrations.

For developers building SaaS products this means they can create beautiful marketing pages and more polished MVPs without needing a designer at the earliest stage.

These features compress the product-building workflow. Developers can prompt an idea, preview the design, annotate changes, directly edit components, generate matching assets, and export code.

That makes Google AI Studio increasingly useful for rapid prototyping, MVP development, SaaS landing pages, and front-end experimentation. It helps developers spend less time fighting boilerplate and more time turning ideas into working products.

High-quality AI coding in 2026 has gone way past just picking a good model and calling it a day.

There’s now much more that needs to be properly calibrated and fine-tuned to get the very best results from your agent.

We now have Skills which let you precisely shape AI behavior by packaging instructions, scripts, and context into reusable units.

And that’s what the new, wildly popular “Karpathy Skills” have been able to take advantage of to the fullest extent.

Karpathy Skills is a set of strict rules and guidelines that drastically improve the accuracy and reliability of your agent, once you add them to your CLAUDE.md (or CURSOR.md) file.

Let’s take a look at some of these key rules, so you can better understand why it makes such a massive difference.

1. The surgical strike

Most LLMs try to be helpful. Too helpful.

You’ve probably experienced this:

You ask for a fix or new feature.
They make the changes… but also:

• clean up unrelated code
• reformat files
• rename variables
• refactor “while they’re there”

It looks productive. But it leads to low model trust and messes up your mental model of the codebase.

The rule:

• Only change the exact lines required
• No drive-by edits
• No unrelated improvements

Why it matters:

• Prevents diff bloat
• Makes PRs readable
• Reduces hidden risk

Think about review time.

• 500-line diff → slow, error-prone
• 5-line diff → fast, obvious

This isn’t about style.
It’s about trust.

A good AI agent doesn’t try to improve everything.
It solves exactly the problem.

2. Extreme disambiguation

Most agents are optimized to continue.

If they’re 80% sure, they guess the missing 20% and move forward.

That’s dangerous.

The rule:

• Don’t assume
• Don’t hide confusion
• Surface tradeoffs

In practice:

• Ask clarifying questions
• Present multiple interpretations
• Push back on unclear requests

Why it matters:

• Prevents hallucinated requirements
• Exposes ambiguity early
• Creates tighter feedback loops

Bad agent:

“Sure, I implemented it.”

Good agent:

“Do you want A or B? They have different tradeoffs.”

3. Goal-first thinking (declarative over imperative)

Most developers naturally give instructions.
AI works better with outcomes.

The Karpathy-style rule is simple:

• Define success criteria
• Loop until verified
• Transform imperative tasks into verifiable goals

This pushes the agent into goal-first thinking.

What this changes

Instead of telling the AI what to do step-by-step, you tell it:

• what should fail
• what should pass
• how to know it’s done

That’s declarative thinking.

Imperative vs declarative

Imperative (weak):

“Add validation to this endpoint.”

Declarative (strong):

“Write a test that fails when invalid input is accepted. Update the code until the test passes.”

Notice the difference:

• Imperative → action
• Declarative → outcome

4. Anti–future-proofing (simplicity first)

AI loves to over-engineer.

You ask for something simple.
It builds something “flexible.”

Suddenly you have:

• abstractions
• configuration layers
• unused hooks
• “just in case” logic

The rule:

• If 200 lines could be 50, rewrite it
• No abstractions for single use

Why it matters:

• Prevents AI slop
• Keeps code readable
• Reduces long-term maintenance

Over-engineering compounds.

• One abstraction → pattern
• Pattern → everywhere
• Everywhere → hard to change

Simplicity doesn’t mean naive.
It means appropriate.

If the problem is small, the solution should be small.

The takeaway

No new architecture.
No breakthrough model.

Just constraints:

• Keep diffs small
• Surface ambiguity
• Define success with tests
• Prefer simple code

Claude Skills give structure to these ideas.
Karpathy-style rules give them teeth.

The result:

Not an AI that writes everything — a reliable AI that writes just enough, and just right.

One you can trust in a real codebase.