Nvidia、専門的ローカルエージェントAI向け「Nemotron 3.5 Lightning」をオープンウェイトで公開

*Follow ZDNET: *Add us as a preferred source* on Google.*

AI labs are shipping new models nonstop. Besides being better and faster than their predecessors, not every new model is guaranteed to be a major step change, despite how the company's PR may wax poetic about them. Model strengths really emerge in context: Where are competitor models lacking or excelling? Which models have outstanding specialties, and which are just catching up to industry standards?

Also: How we test AI at ZDNET

Our Model Release Tracker helps you make sense of where models stand relative to each other and whether they're worth a deeper look. While we don't test every model or model update on this list, we'll always include the key elements you need to know, along with our hands-on expert test, where applicable. We also include an Expert Score for certain models. Curious about how we test AI? Check out this breakdown of our process.

Here are the biggest model releases of 2026 so far and what to know about them. We'll update this list whenever a notable new model arrives.

Nemotron 3.5 Lightning

Nvidia | Aug. 11, 2026

What it does: Nvidia framed its new model, part of its existing Nemotron 3 family, as a go-to workhorse for "high-volume" agentic tasks. Faster than similar models and customizable, it's geared toward specialized jobs, which it can target more directly alongside NeMo Switchyard, a new routing library Nvidia released with the model. With "the knowledge of Ultra packed into the size of Nano," as generative AI VP Kari Briski said during a briefing, Nemotron 3.5 Lightning can also run locally for better security.

Also: As Meta fades in open-source AI, Nvidia senses its chance to lead

Why it matters: Timing is everything with this release. Nvidia and even proprietary-first AI companies like OpenAI are pushing open-weight models more than ever, amidst recent security incidents. Meta also just released Muse Glimmer, an open agentic model for local use on individual Macs and PCs.

Especially with lauded Chinese startup DeepSeek raising prices on open models, Nemotron 3.5 Lightning is landing in an industry environment that is more amenable to open models than it perhaps ever has been. Nvidia is framing the model's customizability as a security and privacy perk rather than the risk it's been framed as thus far.

Muse Code and Muse Spark 1.2

Meta | Aug. 5, 2026

What it does: Muse Code (powered by the new Spark 1.2) is Meta's first coding agent, finally joining the likes of OpenAI Codex and Claude Code. Based on benchmarks Meta shared, it still lags behind Claude Code and Codex, but occupies the mid-tier of performance. At $1.25 per million input tokens, Muse Spark 1.2 is priced at a fraction of Anthropic's Claude Opus 5 (another strong coding model), which is $5 per million input tokens.

Also: Claude's Record-a-Skill cut my research from hours to 30 minutes - but the magic has limits

Why it matters: Mostly, this is a one-to-watch release. As sophisticated models stay expensive to use and open-weight or good-enough models become more popular, Muse Code may not need to beat Codex or Claude Code to compete. As Meta's first foray into this part of the industry, it's worth watching how developers react to the price trade-offs.

MAI-Cyber-1-Flash

Microsoft | July 27, 2026

What it does: Microsoft's latest model exists inside MDASH, its agentic security hub designed to triage issues. The company said Cyber-1-Flash is built to detect "challenging vulnerabilities in complex codebases" and costs half of what leading models charge. The model scored more than 10 percentage points above Anthropic's Mythos, Google's Gemini, and OpenAI's just-released GPT-5.6 on CyberGym, a security reasoning benchmark.

Also: OpenAI's attack agent did exactly what it was told - just more relentlessly than expected

Why it matters: Given Mythos' storied cyber capabilities -- which drove Project Glasswing and a few government interventions -- Cyber-1-Flash's benchmark performance demonstrates how rapidly AI is evolving overall. The model meets a critical moment in the AI security conversation, where experts' long-held concerns about severe, unpredictable threats are coming to fruition. An OpenAI agent escaped a testing environment and breached Hugging Face, just weeks after an unrelated first-of-its-kind and fully agentic ransomware attack.

Also: Open weights vs. closed: An AI civil war's afoot, and the stakes are existential

Kimi K3

Moonshot | July 16, 2026

What it does: At 2.8 trillion parameters, Kimi K3 is the largest open-source model on the market, designed for "long-horizon coding, knowledge work, and reasoning," according to Moonshot's announcement. For context, that's bigger than DeepSeek V4 Pro, from the company that first made US AI labs nervous back in January 2025, at 1.6 trillion.

Most notably, Kimi K3 topped Anthropic's Fable 5 on the Arena benchmark for front-end coding, which measures complex agentic coding tasks. That said, Moonshot itself noted that Kimi K3's overall performance doesn't top Fable 5, though the model does rival it and OpenAI's GPT-5.6 in several individual benchmarks.

Moonshot released the model's weights on July 27 as promised.

Also: Moonshot AI's new Kimi K2.6 swarms your complex tasks with 1,000 collaborating agents

Why it matters: Kimi K3's competitive performance and record-breaking open-source size reignite debate over whether proprietary American frontier models are worth the cost. Fable 5 is $50 per million output tokens, while Kimi K3 is just $15. Still, open-source models lack the safety guardrails that proprietary models have; those risks are amplified by the fact that Moonshot is a Chinese startup (the US government has been consistently suspicious of Chinese tech).

Inkling

Thinking Machines | July 15, 2026

What it does: Inkling is the first model from Thinking Machines, a startup founded by several ex-OpenAI executives including Mira Murati. It's open-weight and generalist, scoring similarly to other open models on benchmarks for agentic coding and tool use, and slightly better on design and audio. It also used far fewer tokens than major open models like Nemotron 3 Ultra, Thinking Machines said, making it an economical choice.

Why it matters: By starting with an open-weight model, Thinking Machines is characterizing itself somewhat differently than other US-based labs, which prioritize more lucrative proprietary models. The company has also invested in research on designing AI that better keeps humans in the loop during use. While not immediately competitive, Inkling's strong early performance demonstrates a possible open-weight renaissance for the US, as most impressive open models are still China-based.

GPT-5.6

OpenAI | July 9, 2026

What it does: Fresh on the heels of its new voice model, OpenAI released the GPT-5.6 family, which the company said it had cleared with the government a few days earlier. The family includes Sol, a new flagship model, as well as Terra, built for everyday use, and Luna, the economical choice of the three. OpenAI said Sol should optimize cost without sacrificing performance and works with a new "ultra" setting that calls on several agents to speed up task completion.

Sol beat Fable 5, the current darling of the AI race, in adaptive and medium reasoning on UC Berkeley's Agents Last Exam benchmark. Terra and Luna also "outperform Fable 5 at around one-sixteenth the cost," OpenAI said in the release. The model family is about on par with Fable 5 in most major performance areas, including coding, work tasks, and general applications, though it is still faster and cheaper, according to OpenAI.

Also: OpenAI's GPT-5.6 and ChatGPT Work aim to beat Anthropic on price, speed, and productivity

Why it matters: That the government was involved in the timing of GPT-5.6's release signifies a shift -- not in explicit regulation, perhaps, but around a sort of informal partnership between labs and the Center for AI Standards and Innovation (CAIS). Set in motion by the back-and-forth around Anthropic's Mythos and Fable 5, the voluntary testing agreement laid out in the Trump administration's June executive order seems to be going swimmingly.

*(Disclosure: Ziff Davis, ZDNET's parent company, filed an April 2025 lawsuit against OpenAI, alleging it infringed Ziff Davis copyrights in training and operating its AI systems.)*

Muse Spark 1.1

Meta | July 9, 2026

What it does: An upgrade to Meta's first model, Muse Spark, which it released in April, 1.1 scored above Anthropic's Opus 4.8 and OpenAI's GPT-5.5 -- both impressive models -- for job tasks humans delegate to AI and financial analysis. While the rest of its benchmark results were more scattered, those areas are especially relevant to AI's potential impact on white-collar work and the broader job market.

Also: Meta's new AI tool lets others use your Instagram posts for image generation - how to opt out

Why it matters: Unlike GPT-5.6, Muse Spark 1.1 doesn't rival Anthropic's Fable 5, but it does appear to focus on advancing "personal superintelligence," Mark Zuckerberg's somewhat open-ended niche in the agentic assistant world. For example, Meta's release mentions using Spark 1.1 for "agentic dinner party organization," peeling off from Anthropic and OpenAI's focus on professional use cases and aligning itself more with autonomous capability for personal use, similar to Gemini Spark. Keeping pace with new models from other major labs will allow for better simultaneous testing, which should put the race at an interesting inflection point.

GPT-Live-1

OpenAI | July 8, 2026

What it does: OpenAI said in the release that the new model, which is accompanied by a light version (GPT-Live-1 mini), should make talking to ChatGPT in Voice Mode seem "more like having a real conversation." Because GPT-Live can listen and speak simultaneously, according to OpenAI, that means you should be able to seamlessly interrupt the assistant with questions, pause mid-sentence, and ask ChatGPT to respond more slowly if you need.

OpenAI added that with GPT-Live, ChatGPT Voice can delegate more complex asks to models like GPT-5.5 if Live can't handle them. Everyone can access GPT-Live-1 and GPT-Live-1 mini in ChatGPT for Android, iOS, and in the browser.

Also: I tested ChatGPT's Live Voice upgrade, and it almost felt human - how to try it

Why it matters: While technically a narrower upgrade than a brand-new, highly capable general model, GPT-Live-1 should improve the user experience of talking to AI assistants, an increasingly popular way to engage with them on the go. Eventually, a better voice model could also handle longer, more complex tasks, but we'll have to wait and see.

With the recent rollout of Siri AI, GPT-Live-1 gives users another voice assistant improvement (hopefully) to compare to Siri, though ZDNET's early testing of Siri AI against ChatGPT and Gemini already left something to be desired.

Update: Fable 5 Access

Fable 5 costs more to use, but Anthropic included it as a promotion at no extra cost for paid users, including those on Pro, Max, and Team plans, and****select users on Enterprise plans. Anthropic had originally set Fable 5 to shift to token-based usage on July 7 for paid users but extended paid plan access to July 12.

"As before, you can use up to 50% of your weekly usage limit on Claude Fable 5. After that, you can keep using Fable 5 with usage credits, or switch to another model to keep working within your remaining limits," Anthropic explained in an X post.

After Fable 5 and Mythos 5 were pulled on advisement from the government just days after their release, the government re-allowed access to Mythos 5 for certain partners on June 26. On June 30, the Department of Commerce lifted export controls on both Mythos 5 and Fable 5, and Anthropic began restoring global access to Fable 5 on July 1. Mythos 5 is still only available to specific institutions.

Sonnet 5

Anthropic | June 30, 2026

What it does: Pivoting from its recent focus on Opus models, Anthropic released Sonnet 5. According to Anthropic, Sonnet 5 can "make plans, use tools like browsers and terminals, and run autonomously at a level that, just a few months ago, required larger and more expensive models." The company said it performs similarly to Opus 4.8, released a month ago, but costs less. Sonnet 5 starts at $2 per million input tokens but will jump to $3 per million in September. It's now the default in Free and Pro plans and available to all other plan types (Max, Team, and Enterprise).

Also: Why AI tokens will send your enterprise cloud bill sky-high again

In keeping with the agentic focus of the AI industry at large, Sonnet 5 scores notably high on computer use benchmarks and agentic coding -- testing found it completed complex tasks that earlier Sonnet models couldn't. It also comes with safeguards implemented automatically, a nod to both recent PR around Mythos and overall improvements in general intelligence.

Why it matters: With the rate of AI model development continuing to accelerate, it's a fair bet that Sonnet 5 is in fact a significant upgrade over Sonnet 4.6, which was released back in February (eons ago in current shipping time). The timing of this release is also significant; Anthropic noted that Sonnet 5 "has a much lower ability to perform dangerous cybersecurity tasks than our current Opus models," perhaps a stabilizing mention after the complex rollout of its powerful Fable 5 and Mythos 5 models. Ironically, however, Sonnet 5 showed a higher rate of misaligned behavior than Mythos Preview.

Fable 5 and Mythos 5

Anthropic | June 9, 2026

What it does: ZDNET senior contributing editor David Gewirtz called Fable 5 a "defanged" version of Mythos, made safe for public use. Anthropic released Mythos 5 to those who already had Mythos Preview access through Project Glasswing. Fable 5 is restricted from responding to high-risk queries about topics like cybersecurity and biological weapons, but still "Mythos-class," which, by Anthropic's telling, means a step change in capability. As for Mythos 5, Anthropic said it had plans to expand access beyond initial partners through a systematic program. But both models were pulled barely four days after their release on orders from the US government.

Also: Why Anthropic suddenly pulled Fable 5 and Mythos 5 for everyone

Why it matters: The models have created plenty of commotion, not just because of their supposed basis in Mythos and temporary restriction. Fable 5 beguiled safety testers who were unaware that the model was set to downgrade to Opus in response to certain questions, which created trust issues between researchers and Anthropic. Despite being guardrailed, we now know Amazon researchers jailbroke Fable 5 and communicated with the White House to alert Anthropic, though Anthropic framed this incident as "narrow" in its understanding.

While the details remain unclear, it's a rare public instance of government authorities intervening in an American AI product, which could mean Mythos-level models have caused a shift in the Trump administration's relatively hands-off approach to AI labs thus far.

MAI-Thinking-1

Microsoft AI | June 2, 2026

What it does: At its Build developer conference, Microsoft said that this new 35-billion-parameter model is, unsurprisingly, designed for multi-step agentic tasks. It scored similarly on the SWE Bench Pro benchmark test for coding as Anthropic Opus 4.6. The company also noted that enterprise users can trust this model for any use because it was trained only on clean, commercially safe data -- an important tidbit given mounting AI copyright lawsuits.

Also: Microsoft's first reasoning model is one of 7 AIs just released at Build - what we know so far

Why it matters: This is the first reasoning model from Microsoft AI, a notable milestone for any AI lab, but especially so this late in the race. Most labs have created multiple advanced reasoning models at this point, so it's unclear where Microsoft's approach stands relative to those, though the company is focusing on appealing to enterprise clients specifically.

Claude Opus 4.8

Anthropic | May 28, 2026

What it does: Replacing Opus 4.7 starting May 28 (at the same price), Opus 4.8 offers faster thinking modes for one-third the cost of the earlier version, according to Anthropic. Like most of Anthropic's models, 4.8 prioritizes coding abilities, scoring higher than 4.7 on two coding benchmarks but not fully besting OpenAI's GPT-5.5. It also "reaches new highs on our measures of prosocial traits like supporting user autonomy and acting in the user's best interest," the company noted in the release, though definitions for what that means remain murky.

Also: I compared Claude Opus 4.8 with 4.7 in a 10-round honesty test - and a legal prompt broke it

Why it matters: Anthropic has always prioritized model safety and interpretability, but appears to be further emphasizing that standard with this release. The company said Opus 4.7 had a 92% honesty rate, in addition to being less sycophantic and hallucination-prone overall. The fact that it claims 4.8 shows "substantially" lower rates of misalignment than 4.7 indicates an increasingly high standard for model safety, especially because Anthropic compared 4.8's alignment to that of Mythos Preview.

Gemini 3.5 Flash

Google | May 19, 2026

What it does: At Google I/O, the company launched its 3.5 model family, designed for building agents (following suit with where all of the AI industry is focused at the moment). The first model available now is Gemini 3.5 Flash, which beat Gemini 3.1 Pro on a few coding and agentic benchmarks and now undergirds AI Mode in Search and the Gemini App as the default model. Like all Flash models, it's optimized for speed and a cheaper, more lightweight user experience, but can still handle "long-horizon" agentic tasks, Google noted. 3.5 Pro should roll out in June.

Also: Anthropic launches Opus 4.8, with honesty as its killer feature

Why it matters: The most important place 3.5 will operate is in Search. As Google doubles down on an AI-first search experience, the models running it shape the responses people will passively consume every day, meaning it has to be especially up to snuff in terms of accuracy and hallucinations. Notably, though, the System Card for 3.5 Flash lacks any mention of hallucination rate or sycophancy.

GPT-5.5 Instant

OpenAI | May 5, 2026

What it does: OpenAI said in its announcement that the lighter version of OpenAI's just-released GPT-5.5 is less verbose than its predecessor, GPT-5.3 Instant. It also touted fewer hallucinations and improved factuality, saying "GPT‑5.5 Instant produced 52.5% fewer hallucinated claims than GPT‑5.3 Instant on high-stakes prompts covering areas like medicine, law, and finance."

Also: Anthropic's Mythos is evolving faster than expected, reports AI safety agency

Why it matters: GPT-5.5 Instant replaces GPT-5.3 as the default model in ChatGPT. Again, while the expectation is that each new AI model gets more efficient, gets easier to use, and makes up less stuff, a significant improvement in hallucinations for a model most people use for fast queries could mean less misinformation spreading among the masses. That's especially critical given how many people are using ChatGPT for everyday health questions, for example.

*(Disclosure: Ziff Davis, ZDNET's parent company, filed an April 2025 lawsuit against OpenAI, alleging it infringed Ziff Davis copyrights in training and operating its AI systems.)*

Nemotron 3 Nano Omni

Nvidia | April 28, 2026

What it does: The latest in Nvidia's open Nemotron family, this model provides agents with multimodal input. That means they can "perceive and reason across visual, audio, and textual inputs within a single shared perception‑to‑action loop," according to Nvidia, thereby unifying multiple capabilities into a single system.

Also: AI is an arms race, and the US wants $9 billion in Nvidia superchips to keep up

Why it matters: Normally, systems of agents need to use separate models for speech, vision, and text, meaning they jump across documents, video, and audio to complete multi-step tasks. That slows down workflows, undermines the context agents gather, and racks up inference costs. Nvidia's approach, if it works, would streamline this process and reduce token use, saving you money. Try it on Hugging Face.

GPT-5.5

OpenAI | April 23, 2026

Expert Score: 93/100

What it does: ZDNET tester-in-residence David Gewirtz technically gave GPT-5.5 an A- score, but said it "can be reductively described as better and faster than GPT-5.4," which is hopefully the bare-minimum expectation for a new model. Specifically, though, the model got better at agentic coding, clearly identifying concepts, scientific research, and factual accuracy.

Also: I put GPT-5.5 through a 10-round test: It scored 93/100, losing points only for exuberance

Why it matters: While the model itself may not be leaps and bounds ahead of its immediate predecessor, the quick turnaround from 5.4 to 5.5 -- less than two months -- indicates how rapidly agentic coding is accelerating OpenAI's model release cycle. As ZDNET's David Gewirtz breaks down, the company, much like other frontier labs using AI to build AI, is shipping updates at an exponentially increasing rate.

ChatGPT Images 2

OpenAI | April 23, 2026

What it does: Soon after sunsetting Sora, its generative video model and social platform, OpenAI somewhat confusingly announced Images 2. ZDNET model tester David Gewirtz got an early look at Images 2 before its release and was impressed. While he didn't give this model a formal Expert Score, he said it's fun, a huge leap, and actually useful for work.

Why it matters: OpenAI seemed to be getting out of the more consumer-minded AI product game when it discontinued Sora, having been beaten by Anthropic at securing lucrative enterprise contracts. That OpenAI still came out with Images 2 within that redirection narrative indicates that it sees image generators as relevant enough to enterprise AI -- especially on the heels of Anthropic's Claude Design.

Claude Opus 4.7

Anthropic | April 16, 2026

What it does: Arriving relatively quickly after Opus 4.6, this model boasts new highs in honesty and reductions in sycophancy and hallucinations. It also appears to have a knack for cybersecurity, as it backs the new Claude Security, released shortly after the model itself -- but no, it's not Mythos, as many suspected.

Also: Anthropic's new Claude Security tool scans your codebase for flaws - and helps you decide what to fix first

Why it matters: Hallucinations and honesty are among the most difficult issues plaguing even the best models. For Anthropic to claim such significant gains in those areas is no small feat for an AI lab that takes safety seriously.

Claude Mythos (Preview)

Anthropic | April 7, 2026

What it does: This is a tough one because Mythos isn't actually available to the public. Anthropic created quite a media storm when it positioned the new general-purpose model as too powerful to release as usual. While the model is apparently a step change from earlier Anthropic models, the company was especially alarmed because of the security threat it posed, stating that "it is strikingly capable at computer security tasks."

In response to that, Anthropic spearheaded Project Glasswing, a collaborative effort with several rival AI labs, including Google, Nvidia, and Microsoft, as well as security authorities like Palo Alto Networks, "to help secure the world's most critical software, and to prepare the industry for the practices we all will need to adopt to keep ahead of cyberattackers."

Also: Apple, Google, and Microsoft join Anthropic's Project Glasswing to defend world's most critical software

Why it matters: If we're to believe Anthropic's guidance that Mythos poses a significant threat to the world's software -- so much so that only a select few partners can access it -- cybersecurity apparatuses as they stand may not be prepared to meet the rapidly evolving frontier of model capabilities. Mythos may not be the only model of its caliber but may simply be the first of many to come once other labs achieve similar breakthroughs.

For now, just a few weeks into its release, Mythos is helping catch software bugs in droves.

GPT-5.4

OpenAI | March 5, 2026

What it does: OpenAI framed this new model, released barely three months after GPT-5.2, as specifically designed for professional work. According to the company's own testing (which should always be taken with a grain of salt until verified by a third party), GPT-5.4 matches or outperforms human professionals 83% of the time.

Why it matters: As AI companies focus more on gaining enterprise trust (and contracts) while lauding what agentic AI can do, they need models that can handle complex work-related tasks with minimal risk, delay, or prohibitively high costs. Any model advancement that shows prowess in professional workflows has a better chance of being taken seriously by companies struggling to adopt AI, though nothing guarantees seamless integration.

Also: OpenAI's new GPT-5.4 clobbers humans on pro-level work in tests - by 83%

Claude Opus 4.6

Anthropic | Feb. 5, 2026

What it does: This model quickly redefined the standard for autonomous agentic work, especially for coding. That's no surprise given Anthropic's authority in building models especially adept at programming tasks. Opus 4.6 also demonstrated improvement in complex, longer-running tasks overall.

Why it matters: Opus 4.6's ability to handle tasks better on its own means you can reliably offload more of your workflow to it -- something agentic offerings usually struggle with.

Also: Anthropic says its new Claude Opus 4.6 can nail your work deliverables on the first try

GPT-5.3-Codex

OpenAI | Feb. 5, 2026

What it does: This new coding model -- which OpenAI said helped build and debug itself -- can be interrupted and redirected mid-task, which, if true, is a huge boon for developers using it on complex or shifting projects with tons of trial-and-error. GPT-5.3-Codex also boasts run times of over a day and a better grasp on user intent.

Also: OpenAI's new Spark model codes 15x faster than GPT-5.3-Codex - but there's a catch

Why it matters: OpenAI is trying to catch up to Anthropic's lead in agentic coding (and, coincidentally or not, released GPT-5.3-Codex on the same day as Anthropic launched Opus 4.6). While ZDNET experts often prefer Claude Code to other tools for vibe coding, OpenAI's rumored shift toward enterprise clients and away from fun consumer tools could eventually close that gap.