New AI Models April 2026: Anthropic Won't Ship Its Best. Open Source Will.

Two press releases, one fracture

April 7, 2026. Two announcements, twelve hours apart.

Anthropic confirmed that Claude Mythos exists — the most capable model it has ever built — and said it will not be publicly available. Fifty organizations get gated access under a program called Project Glasswing. Their job: use Mythos defensively to scan their own infrastructure for vulnerabilities before attackers can weaponize the model's capabilities. Preview pricing: $25 per million input tokens, $125 per million output. No public API. No general availability date.

The same day, Zhipu AI released GLM-5.1 under the MIT license. A 744-billion-parameter mixture-of-experts model with 40 billion active per forward pass, 200K context window. On SWE-Bench Pro — expert-level real-world software engineering — it reportedly beat both Claude Opus 4.6 and GPT-5.4. Cost to use: whatever your electricity costs.

That contrast is the story of early April 2026. Not a benchmark race. Not a pricing war. A philosophical split that has been building for months and is now out in the open. The industry's most capable models are being built faster than anyone can agree on who should use them.

The full release list

Everything that shipped between April 1 and April 8, 2026. Not all of these are traditional text LLMs — the line between "language model" and "AI model" stopped being meaningful sometime around February.

Claude Mythos: the model Anthropic built and then locked away

The Mythos story started as a leak. On March 28, security researchers Roy Paz (LayerX Security) and Alexandre Pauwels (University of Cambridge) found roughly 3,000 unpublished assets sitting in an unprotected, publicly searchable Anthropic database. Among them: draft blog posts describing a next-generation model codenamed "Capybara" in one version and "Claude Mythos" in another. Anthropic patched the exposure the same night after Fortune reached out for comment.

Then, on April 7, Anthropic moved Mythos into controlled early access under Project Glasswing. This isn't a beta test. It's a defensive deployment. The partner list reads like a who's-who of critical infrastructure: AWS, Apple, Microsoft, Google, NVIDIA, Cisco, CrowdStrike, JPMorgan, Broadcom, Palo Alto Networks, the Linux Foundation, and roughly 40 others. Their mandate is specific: use Mythos to scan their own systems and open-source codebases for exploitable vulnerabilities before the model reaches a wider audience.

Read this in context. Anthropic spent March in a standoff with the Pentagon after refusing to let Claude be used in autonomous weapons systems. Multiple U.S. agencies began phasing out Claude models over a six-month transition. The company was labeled a "supply-chain risk" — a designation normally reserved for foreign adversaries. Now that same lab is telling the world that its own model is a cybersecurity risk and choosing to gate access on that basis.

Whether you see this as principled caution or competitive positioning depends on your priors. Either way, it sets a precedent. Mythos is the first time a major lab has publicly said: "We built something too capable to release." That framing will shape how every future frontier model gets rolled out.

GLM-5.1: the model that beat the frontier and costs nothing

While Anthropic was locking things down, Zhipu AI went the other direction.

GLM-5.1 dropped on April 7 under the MIT license — the most permissive license commonly used in open source. 744 billion total parameters, 40 billion active per forward pass via MoE routing. 200K context window. And the number that matters most: it reportedly topped both Claude Opus 4.6 and GPT-5.4 on SWE-Bench Pro, the benchmark that measures real-world software engineering with expert-level tasks.

Zhipu AI (Z.ai) has been climbing steadily — GLM-4.5, GLM-4.6, GLM-4.7, GLM-5, and now GLM-5.1. Each iteration has been more capable and more openly licensed. The jump to MIT is significant. Apache 2.0 (Google's choice for Gemma) requires patent grants and attribution. MIT requires almost nothing. Zhipu is saying: take it, use it, build on it, we don't care how.

For developers and companies evaluating coding assistants, the implication is direct. The strongest coding model you can run today — by one credible benchmark — is not behind an API paywall. It's on GitHub. Whether that holds as more independent evaluations come in is an open question, but the signal is hard to ignore.

Gemma 4: Google's best open-source play

Google shipped the Gemma 4 family on April 1 under Apache 2.0. Four variants, each targeting a different deployment tier:

What sets Gemma 4 apart from previous open-weight releases is the unified multimodal design. All four variants handle text, images, and audio natively — not through separate adapters or bolted-on vision encoders, but as part of the base architecture. The 27B variant scores around 0.8 on GPQA, putting it in the range of models two to three times its size from a year ago.

The edge variants (E2B and E4B) are the ones to watch for practical impact. A 2B parameter model that runs on a phone and handles text, images, and audio is not a research project. It's a product component. If you're building mobile apps, embedded systems, or anything that needs local inference without a cloud round-trip, Gemma 4 E2B is the strongest option available under an open license.

Google calls Gemma 4 "our most capable open models to date." Given that they also have Gemini 3.1 Pro at the top of the closed leaderboard, the strategic read is straightforward: Google is competing on both sides of the open/closed divide simultaneously. That's a position no other lab currently holds.

The rest of the field

Three other releases from the first week of April deserve attention, even if they didn't dominate the headlines.

Zooming out: three shifts defining April

The Feb/March sprint produced the frontier: GPT-5.4, Gemini 3.1 Pro, Claude Opus 4.6, Llama 4, GLM-5. The Intelligence Index ceiling hit 57.18 and held. April's first week isn't about raising that ceiling. It's about something more consequential.

Where this is all headed

A useful frame: February and March were about building the frontier. April is about deciding who gets to stand on it.

The Intelligence Index ceiling (57.18) has held since Gemini 3.1 Pro Preview in February. GPT-5.4 matched it in March. Nobody has broken through. Google, OpenAI, and Anthropic all have models in the pipeline — Mythos being the most visible — but Q1 2026 ended without a clear new #1. The frontier is a plateau, at least temporarily.

What's changing fast is everything below and around the frontier.

The ceiling will break eventually. Someone will cross 60 on the Intelligence Index, probably before summer. When they do, the others will follow within weeks. But the more defining question for Q2 isn't "which model scores highest." It's "who controls the deployment." Mythos answers one way. GLM-5.1 answers the other. Both answers are going to coexist, and the tension between them will shape the industry for years.

Practical guidance for early April 2026

The bottom line

Data sourced from developer announcements, LLM Stats, PricePerToken, Fortune, VentureBeat, and WhatLLM.org tracking. Covers April 1–8, 2026. See our interactive model explorer for live pricing, speed, and benchmark data across 280+ models.