What AI models are actually coming in the next six months? August 20th 2026

Curated by Alexis Dufresne

If you use AI at work, the tools you rely on could change again before February. OpenAI, Google, Meta, Anthropic, several Chinese labs, and a group of world-model startups are all preparing or rumored to be preparing new releases. Some have announced dates. Others have only appeared in testing reports, leaks, or investor comments. This issue sorts those signals into a practical list: what is likely to ship, what will probably slip, and which releases might actually be worth changing your plans for.

In the Wild

The audience has already front-run the release calendar. See the latest full movement in In the Wild.

A Fable 5.1 leak video is pulling 1,678 views an hour. The rumor roundup also promises a Gemini update and an "uncensored" Qwen model, which is a neat snapshot of how model speculation now gets packaged. Watch it.
Canva entered Photo & Video at No. 3. Whatever wins the model race will increasingly arrive inside a product people already use, not as a chatbot tab. View the app.
AI video creation climbed three places. AI Video Generator + Creator moved to No. 9 in Graphics & Design, another signal that better video models will find demand immediately. View the app.
Gauth climbed two places in Education. Camera-first explanations are becoming a default model interface for students. View the app.
Alta entered Lifestyle with an AI closet. The winning consumer model may be the one users never have to name. View the app.

Quick Hits

The Lab Gladiator Era

The four largest Western labs have four different reasons they cannot publish a clean date.

OpenAI's next model is waiting on a security architecture, not another training run. Astra is real, and OpenAI says its latest evaluations show such large gains in agentic coding and cybersecurity that it cannot rule out critical capability. Some internal work is paused until stronger controls are in place. This is the most consequential model in the queue and the least schedulable.
Meta has turned December 31 into a referendum on its AI rebuild. Watermelon, the next Muse Spark generation, is still training with vastly more compute and is supposed to arrive this year. A delay would be more than calendar slip; it would reopen the question of whether Meta's spending and talent raid produced a frontier model.
Google has two flagships in the pipe and one of them is already late. Gemini 3.5 Pro is in testing but reportedly months behind schedule, while Google says Gemini 4 is in pretraining. The likely sequence is a delayed 3.5 Pro release before any true generational jump, not the surprise Gemini 4 launch the rumor accounts want. Read the status.
Anthropic's rumor stack contains a patch, a moonshot, and a model you cannot have. Fable 5.1 has reportedly appeared in some accounts, while SemiAnalysis founder Dylan Patel theorizes that Mythos 2 has been used to train Mythos 3. Separately, Anthropic's own risk report describes a stronger internal Model 2 that it does not plan to release. The sightings and theory are signals, not a roadmap.

DeepSeek's Quiet Takeover

The most credible surprise may not come from a US lab.

MiniMax could put 2.7 trillion open-weight parameters into the market before October. Reuters reports that the Chinese startup is training what may be the world's largest open-weight model, with a release possible in the third quarter. MiniMax declined to comment, so treat the window as informed reporting rather than a promise. If it lands, the immediate story will be inference cost and deployability, not parameter bragging rights. Read the report.
Z.ai says a Fable-class open model will arrive before year-end. Founder Jie Tang has publicly said his company will likely ship an open model that rivals Anthropic's Fable before 2027. Its current GLM-5.2 already approaches leading US models on some agentic and cybersecurity tests at roughly half the cost. The next release could reset the price of frontier capability.

Auto Mode Everything

The physical-AI labs have the clearest dates because their claims eventually have to touch a factory floor.

Nvidia has two physical-AI releases approaching from opposite directions. Cosmos 3 is meant to unify synthetic world generation, physical reasoning, and action simulation; GR00T N2 turns that stack toward robot control. Nvidia says Cosmos 3 is coming soon and GR00T N2 is slated for year-end. The important benchmark will not be video quality. It will be successful action in an unfamiliar room.
Genesis has promised to put its model into customer environments before the year closes. GENE is the reasoning and control system inside Eno, a general-purpose robot designed for long-horizon industrial work. Production and targeted customer deployments are planned by year-end. This is not a fresh checkpoint release, but it may be the cleanest test of whether a world-action model can graduate from a demo reel.

The Six-Month Release Board

These are AI Weekly's editorial odds, based on public commitments, reported testing, training status, and the number of unresolved gates. They are forecasts, not company guidance.

Now through September: the leak window

Fable 5.1, 65%. The account sightings make a point release believable, and Anthropic has every incentive to improve its public model while keeping more dangerous capability behind Mythos access controls. Expect a better agent and coding model, not a new paradigm.

MiniMax's 2.7T open model, 75%. The Reuters window is specific, the model is reportedly in development, and Chinese labs are shipping at a cadence Western observers keep underestimating. The risk is that a Q3 API preview gets mistaken for downloadable weights.

An SSI model, 25%. A connected investor says August. SSI says nothing. A paper, limited research preview, or hand-picked partner demo is more plausible than a public API.

October through December: the real launch window

Watermelon, 80%. This is the cleanest frontier promise. Meta has attached both a year-end date and the credibility of its AI reboot to the release. Expect the pitch to focus on coding, agents, and the advantage of Meta's user data, not simply a benchmark crown.

Z.ai's Fable-class open model, 70%. Tang has said before year-end, and China has already compressed the capability gap faster than US labs expected. If this ships with permissive weights and lower inference cost, it could matter more to developers than whichever closed model tops the leaderboard.

GR00T N2, 85%; Cosmos 3, 65%; Eno deployments, 70%. Physical AI is the most believable Q4 cluster. Nvidia has given GR00T a date, Cosmos a "soon," and Genesis has named the customer window. Slippage will show up as narrower access, fewer robot bodies, or carefully chosen environments rather than a cancelled launch.

Gemini 3.5 Pro, 60%; Astra, 45%. Google needs to clear a performance delay. OpenAI needs to clear a safety and security gate. Both models probably exist in release-capable form before December. That does not mean either company will make them broadly available.

January through February: the rollover pile

This is where missed Q4 promises go, along with the labs that have capacity but no public schedule. Thinking Machines begins drawing on a gigawatt of Vera Rubin compute early next year, but that points to future training, not a February frontier launch. Reflection is already training on SpaceXAI capacity, also without a date. Gemini 4, Grok 5, Mistral's next Large model, the next Marble generation from World Labs, and new flagship video models from Runway or Google all belong below 35% for this six-month window. Not because the work is not happening. Because there is no release evidence strong enough to turn activity into a calendar.

This Is Three Races, Not One Leaderboard

The usual framing asks which lab will have the smartest model by Christmas. That misses the shape of what is coming. The closed-model race is becoming a deployment problem: Astra may be too cyber-capable for ordinary release, Anthropic is separating public and restricted systems, and Google has to decide whether to ship late or wait for a cleaner jump. The open-model race is becoming an economics problem: MiniMax and Z.ai do not need to beat every benchmark if they make near-frontier capability downloadable and much cheaper. The physical-model race is becoming a reliability problem: Cosmos, GR00T, and GENE have to preserve an internal model of the world long enough to complete work outside a controlled demo.

That creates three different winners. OpenAI or Meta can own the most capable agent. A Chinese lab can own the model developers can actually afford and control. Nvidia can own the layer that connects models to machines. The next six months will not produce one "best model." They will reveal which kind of intelligence the market values enough to deploy.

Key Takeaways

Plan for at least three model migrations before January. Fable 5.1, a Chinese open model, and Watermelon are credible enough that teams should keep evaluations and routing portable.
The biggest price shock may come from China. A Fable-class open model does not need to win every benchmark to force closed-model vendors to cut prices or loosen access.
Astra's delay is itself a capability signal. The next frontier bottleneck may be secure deployment, not training compute.
World models finally have falsifiable deadlines. GR00T N2 and Eno have to work in unfamiliar physical environments, where a leaderboard cannot hide failure.

Found First

Primary research our scoop hunter surfaced before the press got there. Full write-up on the linked page.

The next model upgrade may actually be a better harness. Across 8,135 trials, procedural agent skills produced a 65.7% lift while explicit knowledge added only 4.5%. Retrieval precision collapsed from 29.6% with five candidate skills to 3.3% with 100. The paper is a warning against treating every capability jump as new weights: procedure, retrieval, and control flow may decide which "model" feels smartest in practice.

Worth Reading

Frontier AI forecasting has a measurement problem: an audit of 62 systems finds that model forecasts often mix benchmark, compute, release, and expert evidence without keeping the event record straight. (arXiv)
A Functional Taxonomy of World Models: World Labs separates renderers, simulators, and planners, then explains why the real prize is a model that closes the loop between all three. (World Labs)
Yann LeCun's AMI Labs raises $1.03 billion for world models: the startup case for replacing language-first systems with models that learn how the world works. (TechCrunch)

Wait, What?

Eleven words turned the world's most secretive AI lab into an August launch event. In a podcast discussion about watts, wafers, and continual learning, investor Gavin Baker said SSI told him it would release a model in August. The full episode offers no model name, modality, benchmark, access plan, or day. A global rumor cycle has been built on one sentence from someone who does not work there.