April 2026 is one of the most consequential months for large language model releases. From GPT-6's anticipated launch to open-source powerhouses like Gemma 4 and GLM-5.1, the landscape is shifting fast. Here's what shipped, what's coming, and what it means for developers and AI teams.
The Big Releases
GPT-6 ("Spud") — OpenAI
OpenAI's GPT-6, codenamed "Spud," was announced for April 14 but the launch window has slipped. Sam Altman confirmed the model is still "a few weeks out." What we know so far:
- Context window: 2 million tokens
- Pricing: ~$2.50/1M input, ~$12/1M output (provisional)
- Key focus: Reasoning depth and agentic capabilities
- Status: Not yet publicly available
Once released, GPT-6 is expected to set a new benchmark for frontier proprietary models, particularly in multi-step reasoning and tool use.
Claude Mythos — Anthropic
Anthropic previewed Claude Mythos to select partners on April 7. This is Anthropic's most expensive model to date:
- Pricing: ~$25/1M input, ~$125/1M output
- Target: Enterprise and research workloads requiring maximum quality
- Status: Limited preview, gated access
Claude Mythos appears positioned as a premium reasoning model rather than a general-purpose tool, competing with the top tier of GPT-6 and Gemini Ultra.
Gemma 4 — Google (Apache 2.0)
Google shipped four Gemma 4 variants on April 2, all under the Apache 2.0 license. This is the most significant open-weight release of the month:
| Variant | Size | Architecture | Best For |
|---|---|---|---|
| Gemma 4 31B | 31B dense | Dense transformer | General purpose, fine-tuning |
| Gemma 4 26B MoE | 26B MoE | Mixture of experts | Efficient inference |
| Gemma 4 E4B | ~4B effective | Lightweight | Edge deployment, mobile |
| Gemma 4 E2B | ~2B effective | Ultra-lightweight | On-device, embedded |
All variants support 256K token context windows. The Apache 2.0 license makes them commercially usable without restrictions.
GLM-5.1 — Zhipu AI (MIT License)
GLM-5.1 is a massive open-weight model from Chinese lab Zhipu AI:
- Parameters: 744B MoE (40B active during inference)
- Context: 200K tokens
- License: MIT — fully permissive
- Why it matters: Delivers frontier-level performance at a fraction of the compute cost by activating only 40B of its 744B parameters per token
The MIT licensing is notable — this is one of the most permissively licensed frontier-scale models available.
Qwen 3.6-Plus — Alibaba
Alibaba's Qwen 3.6-Plus brings a 1-million-token context window to open-source:
- Context: 1M tokens
- License: Open weights
- Key advantage: Long-context capabilities rivaling proprietary models
Llama 4 — Meta
Meta is rolling out Llama 4 in two primary variants:
- Llama 4 Scout: 10M token context window — the largest context of any model, open or proprietary
- Llama 4 Maverick: 400B parameters, 1M token context
Both are available under the Llama License. Scout's 10M context window is a technical milestone, though practical utility at that scale remains to be proven.
Arcee Trinity — Arcee AI (Apache 2.0)
A 400B parameter model from Arcee AI, released under Apache 2.0. Details are still emerging, but the open licensing at this scale is significant.
What This Means
Three trends stand out from April's releases:
-
Open-source is closing the gap. Gemma 4, GLM-5.1, Llama 4, and Arcee Trinity collectively offer frontier performance with permissive licensing. The gap between open and proprietary is narrowing monthly.
-
Context windows are exploding. Llama 4 Scout's 10M tokens and Qwen 3.6-Plus's 1M tokens push the boundaries of what's possible. For developers, this means fewer chunking workarounds and more natural document processing.
-
MoE is the dominant architecture. GLM-5.1, Llama 4 Scout, and Gemma 4 26B all use mixture-of-experts to deliver large model quality at smaller inference costs. Expect this trend to accelerate.
Which Model Should You Use?
- General development: Gemma 4 31B (free, Apache 2.0, strong all-around)
- Long-context tasks: Llama 4 Scout (10M tokens) or Qwen 3.6-Plus (1M tokens)
- Maximum quality: GPT-6 or Claude Mythos (when available, at a price premium)
- Resource-constrained: Gemma 4 E2B or E4B (edge-ready, Apache 2.0)
- Enterprise fine-tuning: GLM-5.1 (MIT license, 40B active params, efficient training)
The full model details, benchmarks, and comparison tools for these models are available on their individual pages in our directory.