Observe ZDNET: Add us as a most popular supply on Google.
AI labs are delivery new fashions nonstop. Apart from being higher and sooner than their predecessors, each new mannequin is not assured to be a significant step change, regardless of how the corporate’s PR might wax poetic about them. Mannequin strengths actually emerge in context: The place are competitor fashions missing or excelling? Which fashions have excellent specialties, and that are simply catching as much as business requirements?
Additionally: How we check AI at ZDNET
Our Mannequin Launch Tracker helps you make sense of the place fashions stand relative to one another, and whether or not they’re value a deeper look. Whereas we do not check each mannequin or mannequin replace on this listing, we’ll at all times embody the important thing parts you have to know, together with our hands-on skilled check, the place relevant. We additionally embody an Skilled Rating for sure fashions. Interested in how we check AI? Try this breakdown of our course of.
Listed below are the largest mannequin releases of 2026 thus far and what to find out about them. We’ll replace this listing at any time when a notable new mannequin arrives.
Kimi K3
Moonshot | July 16, 2026
What it does: At 2.8 trillion parameters, Kimi K3 is the most important open-source mannequin in the marketplace, designed for “long-horizon coding, data work, and reasoning,” in keeping with Moonshot’s announcement. For context, that is greater than DeepSeek V4 Professional, from the corporate that first made US AI labs nervous again in January 2025, at 1.6 trillion.
Most notably, Kimi K3 topped Anthropic’s Fable 5 on the Area benchmark for front-end coding, which measures complicated agentic coding duties. That mentioned, Moonshot itself famous that Kimi K3’s general efficiency does not high Fable 5, although the mannequin does rival it and OpenAI’s GPT-5.6 in a number of particular person benchmarks.
Moonshot plans to launch the mannequin’s weights by July 27.
Additionally: Moonshot AI’s new Kimi K2.6 swarms your complicated duties with 1,000 collaborating brokers
Why it issues: Kimi K3’s aggressive efficiency and record-breaking open-source dimension revive discourse on whether or not proprietary American frontier fashions are value the fee. Fable 5 is $50 per million output tokens, whereas Kimi K3 is simply $15. Nonetheless, open-source fashions lack the protection guardrails that proprietary fashions have; these dangers are amplified by the truth that Moonshot is a Chinese language startup (the US authorities is persistently suspicious of Chinese language tech).
GPT-5.6
OpenAI | July 9, 2026
What it does: Contemporary on the heels of its new voice mannequin, OpenAI launched the GPT-5.6 household, which the corporate mentioned it cleared with the federal government just a few days earlier. The household contains Sol, a brand new flagship mannequin, in addition to Terra, constructed for on a regular basis use, and Luna, the economical selection of the three. OpenAI mentioned Sol ought to optimize value with out sacrificing efficiency, and works with a brand new “extremely” setting that calls on a number of brokers to hurry up job completion.
Sol beat Fable 5, the present darling of the AI race, in adaptive and medium reasoning on UC Berkeley’s Brokers Final Examination benchmark. Terra and Luna additionally “outperform Fable 5 at round one-sixteenth the fee,” OpenAI mentioned within the launch. The mannequin household is about on par with Fable 5 in most main efficiency areas, although, together with coding, work duties, and basic purposes, however continues to be sooner and cheaper, in keeping with OpenAI.
Additionally: OpenAI’s GPT-5.6 and ChatGPT Work purpose to beat Anthropic on value, velocity, and productiveness
Why it issues: That the federal government was concerned within the timing of GPT-5.6’s launch signifies a shift — not in specific regulation, maybe, however round a type of casual partnership between labs and the Heart for AI Requirements and Innovation (CAIS). Set in movement by the back-and-forth round Anthropic’s Mythos and Fable 5, the voluntary testing settlement specified by the Trump administration’s June government order appears to be going swimmingly.
(Disclosure: Ziff Davis, ZDNET’s mother or father firm, filed an April 2025 lawsuit in opposition to OpenAI, alleging it infringed Ziff Davis copyrights in coaching and working its AI programs.)
Muse Spark 1.1
Meta | July 9, 2026
What it does: An improve to Meta’s first mannequin, Muse Spark, which it launched in April, 1.1 scored above Anthropic’s Opus 4.8 and OpenAI’s GPT-5.5 — each spectacular fashions — for job duties people delegate to AI and monetary evaluation. Whereas the remainder of its benchmark outcomes have been extra scattered, these areas are particularly related to AI’s potential affect on white-collar work and the broader job market.
Additionally: Meta’s new AI instrument lets others use your Instagram posts for picture technology – the way to choose out
Why it issues: In contrast to GPT-5.6, Muse Spark 1.1 does not rival Anthropic’s Fable 5, but it surely does seem to give attention to advancing “private superintelligence,” Mark Zuckerberg’s considerably open-ended area of interest within the agentic assistant world. For instance, Meta’s launch mentions utilizing Spark 1.1 for “agentic ceremonial dinner group,” peeling off from Anthropic and OpenAI’s give attention to skilled use circumstances and aligning itself extra with autonomous functionality for private use, just like Gemini Spark. Maintaining tempo with new fashions from different main labs will enable for higher simultaneous testing, which ought to put the race at an fascinating inflection level.
GPT-Reside-1
OpenAI | July 8, 2026
What it does: OpenAI mentioned within the launch that the brand new mannequin, which is accompanied by a lightweight model (GPT-Reside-1 mini), ought to make speaking to ChatGPT in Voice Mode appear”extra like having an actual dialog.” As a result of GPT-Reside can hear and converse concurrently, in keeping with OpenAI, which means it’s best to have the ability to seamlessly interrupt the assistant with questions, pause mid-sentence, and ask ChatGPT to reply extra slowly if you happen to want.
OpenAI added that with GPT-Reside, ChatGPT Voice can delegate extra complicated asks to fashions like GPT-5.5 if Reside cannot deal with them. Everybody can entry GPT-Reside-1 and GPT-Reside-1 mini in ChatGPT for Android, iOS, and within the browser.
Additionally: I examined ChatGPT’s Reside Voice improve, and it virtually felt human – the way to attempt it
Why it issues: Whereas technically a narrower improve than a brand-new, extremely succesful basic mannequin, GPT-Reside-1 ought to enhance the consumer expertise of speaking to AI assistants, an more and more fashionable technique to interact with them on the go. Ultimately, a greater voice mannequin might additionally deal with longer, extra complicated duties, however we’ll have to attend and see.
With the latest rollout of Siri AI, GPT-Reside-1 offers customers one other voice assistant enchancment (hopefully) to match Siri to, although ZDNET’s early testing of Siri AI in opposition to ChatGPT and Gemini already left one thing to be desired.
Replace: Fable 5 Entry
Fable 5 prices extra to make use of, however Anthropic included it as a promotion at no additional value for paid customers, together with these on Professional, Max, Crew, and choose customers on Enterprise plans. Anthropic had initially set Fable 5 to shift to token-based utilization on July 7 for paid customers, however has prolonged paid plan entry to July 12.
“As earlier than, you should utilize as much as 50% of your weekly utilization restrict on Claude Fable 5. After that, you may preserve utilizing Fable 5 with utilization credit, or change to a different mannequin to maintain working inside your remaining limits,” Anthropic defined in an X publish.
After Fable 5 and Mythos 5 have been pulled on advisement from the federal government simply days after their launch, the federal government re-allowed entry to Mythos 5 for sure companions on June 26. On June 30, the Division of Commerce lifted export controls on each Mythos 5 and Fable 5, and Anthropic started restoring international entry to Fable 5 on July 1 (Mythos 5 continues to be solely accessible to particular establishments).
Sonnet 5
Anthropic | June 30, 2026
What it does: Pivoting from its latest give attention to Opus fashions, Anthropic launched Sonnet 5. In line with Anthropic, Sonnet 5 can “make plans, use instruments like browsers and terminals, and run autonomously at a stage that, only a few months in the past, required bigger and costlier fashions.” The corporate mentioned it performs equally to Opus 4.8, launched a month in the past, however prices much less. Sonnet 5 begins at $2 per million enter tokens however will leap to $3 per million in September. It is now the default in Free and Professional plans and accessible to all different plan varieties (Max, Crew, and Enterprise).
Additionally: Why AI tokens will ship your enterprise cloud invoice sky-high once more
In step with the agentic focus of the AI business at massive, Sonnet 5 scores notably excessive on pc use benchmarks and agentic coding — testing discovered it accomplished complicated duties that earlier Sonnet fashions could not. It additionally comes with safeguards carried out mechanically, a nod to each latest PR round Mythos and general enhancements usually intelligence.
Why it issues: With the speed of AI mannequin improvement persevering with to speed up, it is a truthful guess that Sonnet 5 is in truth a major improve over Sonnet 4.6, which was launched again in February (eons in the past in present delivery time). The timing of this launch can be important; Anthropic famous that Sonnet 5 “has a a lot decrease skill to carry out harmful cybersecurity duties than our present Opus fashions,” maybe a stabilizing point out after the complicated rollout of its highly effective Fable 5 and Mythos 5 fashions. Satirically, nevertheless, Sonnet 5 confirmed a better charge of misaligned conduct than Mythos Preview.
Fable 5 and Mythos 5
Anthropic | June 9, 2026
What it does: ZDNET senior contributing editor David Gewirtz known as Fable 5 a “defanged” model of Mythos, made secure for public use. Anthropic launched Mythos 5 to those that already had Mythos Preview entry by way of Mission Glasswing. Fable 5 is restricted from responding to high-risk queries about matters like cybersecurity and organic weapons, however nonetheless “Mythos-class,” which, by Anthropic’s telling, means a step change in functionality. As for Mythos 5, Anthropic mentioned it had plans to broaden entry past preliminary companions by way of a scientific program. However each fashions have been pulled barely 4 days after their launch on orders from the US authorities.
Additionally: Why Anthropic abruptly pulled Fable 5 and Mythos 5 for everybody
Why it issues: The fashions have created loads of commotion, not simply due to their supposed foundation in Mythos and momentary restriction. Fable 5 beguiled security testers who have been unaware that the mannequin was set to downgrade to Opus in response to sure questions, which created belief points between researchers and Anthropic. Regardless of being guardrailed, we now know Amazon researchers jailbroke Fable 5 and communicated with the White Home to alert Anthropic, although Anthropic framed this incident as “slim” of their understanding.
Whereas the main points stay unclear, it is a uncommon occasion of presidency authorities intervening in an American AI product in public, which might imply Mythos-level fashions have prompted a shift within the comparatively hands-off method to AI labs the Trump administration has had to this point.
MAI-Pondering-1
Microsoft AI | June 2, 2026
What it does: At its Construct developer convention, Microsoft mentioned that this new 35-billion-parameter mannequin is, unsurprisingly, designed for multi-step agentic duties. It scored equally on the SWE Bench Professional benchmark check for coding as Anthropic Opus 4.6. The corporate additionally famous that enterprise customers can belief this mannequin for any use as a result of it was skilled solely on clear, commercially secure information — an necessary tidbit given mounting AI copyright lawsuits.
Additionally: Microsoft’s first reasoning mannequin is considered one of 7 AIs simply launched at Construct – what we all know thus far
Why it issues: That is the primary reasoning mannequin from Microsoft AI, a notable milestone for any AI lab, however particularly so this late on this race. Most labs have created a number of superior reasoning fashions at this level, so it is unclear but the place Microsoft’s method stands relative to these, although the corporate is specializing in interesting to enterprise purchasers particularly.
Claude Opus 4.8
Anthropic | Could 28, 2026
What it does: Changing Opus 4.7 beginning Could 28 (on the identical value), Opus 4.8 presents sooner considering modes for one-third the price of the sooner model, in keeping with Anthropic. Like most of Anthropic’s fashions, 4.8 prioritizes coding talents, scoring larger than 4.7 on two coding benchmarks however not totally besting OpenAI’s GPT 5.5. It additionally “reaches new highs on our measures of prosocial traits like supporting consumer autonomy and appearing within the consumer’s greatest curiosity,” the corporate famous within the launch, although definitions for what which means stay murky.
Additionally: I in contrast Claude Opus 4.8 with 4.7 in a 10-round honesty check – and a authorized immediate broke it
Why it issues: Anthropic has at all times prioritized mannequin security and interpretability, however seems to be additional emphasizing that normal with this launch. The corporate mentioned Opus 4.7 had a 92% honesty charge, along with being much less sycophantic and hallucination-prone general. The truth that it claims 4.8 reveals “considerably” decrease charges of misalignment than 4.7 signifies an more and more excessive normal for mannequin security, particularly as a result of Anthropic in contrast 4.8’s alignment to that of Mythos Preview.
Gemini 3.5 Flash
Google | Could 19, 2026
What it does: At Google I/O, the corporate launched its 3.5 mannequin household, designed for constructing brokers (following swimsuit with the place all the AI business is targeted in the intervening time). The primary mannequin accessible now’s Gemini 3.5 Flash, which beat Gemini 3.1 Professional on just a few coding and agentic benchmarks and now undergirds AI Mode in Search and the Gemini App because the default mannequin. Like all Flash fashions, it is optimized for velocity and a less expensive, extra light-weight consumer expertise, however can nonetheless deal with “long-horizon” agentic duties, Google famous. 3.5 Professional ought to roll out in June.
Additionally: Anthropic launches Opus 4.8, with honesty as its killer characteristic
Why it issues: A very powerful place 3.5 will function is in Search. As Google doubles down on an AI-first search expertise, the fashions operating it form the responses individuals will passively devour each day, which means it must be particularly as much as snuff by way of accuracy and hallucinations. Notably, although, the System Card for 3.5 Flash lacks any point out of hallucination charge or sycophancy.
GPT-5.5 Instantaneous
OpenAI | Could 5, 2026
What it does: OpenAI mentioned in its announcement that the lighter model of OpenAI’s just-released GPT-5.5 is much less verbose than its predecessor, GPT-5.3 Instantaneous. It additionally touted fewer hallucinations and improved factuality, saying “GPT‑5.5 Instantaneous produced 52.5% fewer hallucinated claims than GPT‑5.3 Instantaneous on high-stakes prompts overlaying areas like medication, legislation, and finance.”
Additionally: Anthropic’s Mythos is evolving sooner than anticipated, experiences AI security company
Why it issues: GPT-5.5 Instantaneous replaces GPT-5.3 because the default mannequin in ChatGPT. Once more, whereas the expectation is that every new AI mannequin will get extra environment friendly, simpler to make use of, and makes up much less stuff, a major enchancment in hallucinations for a mannequin most individuals use for quick queries might imply much less misinformation spreading among the many lots. That is particularly crucial given how many individuals are utilizing ChatGPT for on a regular basis well being questions, for instance.
(Disclosure: Ziff Davis, ZDNET’s mother or father firm, filed an April 2025 lawsuit in opposition to OpenAI, alleging it infringed Ziff Davis copyrights in coaching and working its AI programs.)
Nemotron 3 Nano Omni
Nvidia | April 28, 2026
What it does: The most recent in Nvidia’s open Nemotron household, this mannequin offers brokers with multimodal enter. Meaning they’ll “understand and motive throughout visible, audio, and textual inputs inside a single shared notion‑to‑motion loop,” in keeping with Nvidia, thereby unifying a number of capabilities right into a single system.
Additionally: AI is an arms race, and the US needs $9 billion in Nvidia superchips to maintain up
Why it issues: Usually, programs of brokers want to make use of separate fashions for speech, imaginative and prescient, and textual content, which means they leap throughout paperwork, video, and audio to finish multi-step duties. That slows down workflows, undermines the context brokers collect, and racks up inference prices. Nvidia’s method, if it really works, would streamline this course of and scale back token use, saving you cash. Attempt it on Hugging Face.
GPT-5.5
OpenAI | April 23, 2026
Skilled Rating: 93/100
What it does: ZDNET tester-in-residence David Gewirtz technically gave GPT-5.5 an A- rating, however mentioned it “will be reductively described as higher and sooner than GPT-5.4,” which is hopefully the bare-minimum expectation for a brand new mannequin. Particularly, although, the mannequin obtained higher at agentic coding, clearly figuring out ideas, scientific analysis, and factual accuracy.
Additionally: I put GPT-5.5 by way of a 10-round check: It scored 93/100, shedding factors just for exuberance
Why it issues: Whereas the mannequin itself might not be leaps and bounds forward of its fast predecessor, the short turnaround from 5.4 to five.4 — lower than two months — signifies how quickly agentic coding is accelerating OpenAI’s mannequin launch cycle. As ZDNET’s David Gewirtz breaks down, the corporate, very like different frontier labs utilizing AI to construct AI, is delivery updates at an exponentially growing charge.
ChatGPT Photos 2
OpenAI | April 23, 2026
What it does: Quickly after sunsetting Sora, its generative video mannequin and social platform, OpenAI considerably confusingly introduced Photos 2. ZDNET mannequin tester David Gewirtz obtained an early take a look at Photos 2 earlier than its launch and was impressed. Whereas he did not give this mannequin a proper Skilled Rating, he mentioned it is enjoyable, an enormous leap, and really helpful for work.
Why it issues: OpenAI appeared to be getting out of the extra consumer-minded AI product sport when it discontinued Sora, having been crushed by Anthropic at securing profitable enterprise contracts. That OpenAI nonetheless got here out with Photos 2 inside that redirection narrative signifies that it sees picture turbines as related sufficient to enterprise AI — particularly on the heels of Anthropic’s Claude Design.
Claude Opus 4.7
Anthropic | April 16, 2026
What it does: Arriving comparatively shortly after Opus 4.6, this mannequin boasts new highs in honesty, decreased sycophancy, and hallucinations. It additionally seems to have a knack for cybersecurity, because it backs the brand new Claude Safety, launched shortly after the mannequin itself — however no, it isn’t Mythos, as many suspected.
Additionally: Anthropic’s new Claude Safety instrument scans your codebase for flaws – and helps you determine what to repair first
Why it issues: Hallucinations and honesty are among the many most troublesome, hard-to-solve points plaguing even the most effective fashions. For Anthropic to assert such important positive factors in these areas isn’t any small feat for an AI lab that takes security critically.
Claude Mythos (Preview)
Anthropic | April 7, 2026
What it does: This can be a powerful one as a result of Mythos is not really accessible to the general public. Anthropic created fairly a media storm when it positioned the brand new general-purpose mannequin as too highly effective to launch as regular. Whereas the mannequin is seemingly a step change from earlier Anthropic fashions, the corporate was particularly alarmed due to the safety risk it posed, stating that “it’s strikingly succesful at pc safety duties.”
In response to that, Anthropic spearheaded Mission Glasswing, a collaborative effort with a number of rival AI labs, together with Google, Nvidia, and Microsoft, in addition to safety authorities like Palo Alto Networks, “to assist safe the world’s most important software program, and to organize the business for the practices all of us might want to undertake to maintain forward of cyberattackers.”
Additionally: Apple, Google, and Microsoft be a part of Anthropic’s Mission Glasswing to defend world’s most important software program
Why it issues: If we’re to imagine Anthropic’s steering that Mythos poses a major risk to the world’s software program — a lot in order that solely a choose few companions can entry it — cybersecurity apparatuses as they stand might not be ready to satisfy the quickly evolving frontier of mannequin capabilities. Mythos might not be the one mannequin of its caliber, however merely the primary of many to return as soon as different labs obtain related breakthroughs.
For now, only a few weeks into its launch, Mythos helps catch software program bugs in droves.
GPT-5.4
OpenAI | March 5, 2026
What it does: OpenAI framed this new mannequin, launched barely three months after GPT-5.2, as particularly designed for skilled work. In line with the corporate’s personal testing (which ought to at all times be taken with a grain of salt till verified by a 3rd get together), GPT-5.4 matches or outperforms human professionals 83% of the time.
Why it issues: As AI firms focus extra on gaining enterprise belief (and contracts) whereas lauding what agentic AI can do, they want fashions that may deal with complicated work-related duties with minimal threat, delay, or prohibitively excessive prices. Any mannequin development that reveals prowess in skilled workflows has a greater likelihood of being taken critically by firms struggling to undertake AI, although nothing ensures seamless integration.
Additionally: OpenAI’s new GPT-5.4 clobbers people on pro-level work in checks – by 83%
Claude Opus 4.6
Anthropic | Feb. 5, 2026
What it does: This mannequin shortly redefined the usual for autonomous agentic work, particularly for coding. That is no shock given Anthropic’s authority in constructing fashions particularly adept at programming duties. Opus 4.6 additionally demonstrated enchancment in complicated, longer-running duties general.
Why it issues: Opus 4.6’s skill to deal with duties higher by itself means you may reliably offload extra of your workflow to it — one thing agentic choices often battle with.
Additionally: Anthropic says its new Claude Opus 4.6 can nail your work deliverables on the primary attempt
GPT-5.3-Codex
OpenAI | Feb. 5, 2026
What it does: This new coding mannequin — which OpenAI mentioned helped construct and debug itself — will be interrupted and redirected mid-task, which, if true, is a big boon for builders utilizing it on complicated or shifting initiatives with tons of trial-and-error. GPT-5.3-Codex additionally boasts run instances of over a day and a greater grasp on consumer intent.
Additionally: OpenAI’s new Spark mannequin codes 15x sooner than GPT-5.3-Codex – however there is a catch
Why it issues: OpenAI is attempting to catch as much as Anthropic’s lead in agentic coding (and, coincidentally or not, launched 5.3 Codex on the identical day as Anthropic launched Opus 4.6). Whereas ZDNET specialists usually want Claude Code to different instruments for vibe coding, OpenAI’s rumored shift towards enterprise purchasers and away from enjoyable shopper instruments might ultimately shut that hole.
