Fresh wave of model releases and previews
Several notable model updates and previews surfaced this week — Anthropic shared details of a powerful unreleased ‘Mythos’ model with select partners, GLM‑5.1 topped open‑source coding benchmarks and was integrated into tools offering zero‑data‑retention, and Google’s Gemma 4 variants were packaged for Mac via a local app. (x.com) These moves show simultaneous momentum in closed preview programs, open‑source competition for coding tasks, and efforts to push large models down onto endpoint devices. (x.com)
# Fresh wave of model releases and previews Three different parts of the artificial intelligence market moved at once this week. Anthropic disclosed a tightly controlled preview of an unreleased model called Mythos, Z.ai pushed its GLM-5.1 model to the top of several open coding benchmarks and into privacy-focused coding tools, and Google’s new Gemma 4 family started showing up in local Mac setups almost immediately after release. Taken together, the updates point in three directions at once. The biggest labs are still keeping their strongest systems behind partner previews, open-weight challengers are getting better at software work fast, and more developers are trying to run useful models on their own machines instead of sending every prompt to a cloud service. (opencode.ai) Anthropic’s announcement was the most striking because it was not a normal product launch. In a technical post published on April 7, 2026, the company described “Claude Mythos Preview” as a new general-purpose language model and said it had already been shared in a limited preview while Anthropic launched “Project Glasswing” to study its cybersecurity impact. The company’s write-up focused less on chat quality than on offensive security capability. Anthropic said Mythos Preview could identify and exploit zero-day vulnerabilities across major operating systems and major web browsers during internal testing, and said more than 99 percent of the vulnerabilities it found were still unpatched, which is why the company withheld most details. That framing matters because it shows how frontier model releases are changing shape. Instead of shipping every improvement as a public chatbot upgrade, labs are increasingly using restricted previews when a model’s new strength touches a sensitive area like cyber offense, where broader access creates immediate risk. The second development came from Z.ai, which published details for GLM-5.1 on April 7, 2026. The company described it as its latest flagship model for “long-horizon” work, meaning the system is meant to keep solving one software task over extended sessions instead of giving a quick first answer and stalling. Z.ai tied that claim to concrete coding results. In its release materials, the company said GLM-5.1 reached state-of-the-art performance on SWE-Bench Pro and outperformed its earlier GLM-5 model on NL2Repo and Terminal-Bench 2.0, all benchmarks aimed at measuring how well a model can handle real software engineering tasks rather than short code snippets. One of the more revealing examples was not a benchmark leaderboard but a long optimization run. Z.ai said GLM-5.1 kept improving a vector database task for more than 600 iterations and over 6,000 tool calls, reaching 21.5 thousand queries per second, compared with a prior best result of 3,547 queries per second in the shorter evaluation setup it cited. That kind of result helps explain why coding agents are becoming the main battleground for open models. If a model can keep reading files, testing code, revising its plan, and trying again for hours, then the useful comparison is no longer “Can it write a function,” but “Can it stay productive through a messy engineering loop.” Distribution also mattered here. OpenCode says its Zen service offers tested model-provider combinations for coding agents and that its providers follow a zero-retention policy, while the broader OpenCode product says it does not store users’ code or context data, making privacy a selling point for teams that want strong coding assistance without broad data exposure. (opencode.ai) The third movement came from Google’s Gemma line, but the important part was not only the model itself. Google announced on April 2, 2026 that Gemma 4 is the first Gemma family released under the Open Source Initiative approved Apache 2.0 license, and said the family spans systems “from edge devices to 31B parameters,” which lowers legal and technical friction for local deployment. Google also said the Gemma community had already passed 400 million downloads and produced more than 100,000 variants before Gemma 4 arrived. That means a new permissively licensed release does not enter an empty market; it lands in an ecosystem that already knows how to quantize models, package them for laptops, and wire them into local runtimes on Apple Silicon Macs. So this week’s story is not one winner beating everyone else. It is three layers of the market moving at the same time: a frontier lab holding back a powerful cyber-capable model for limited preview, an open challenger pushing coding performance upward with longer-running agents, and a major platform company making it easier for developers to run capable models close to the device.
