Which new AI models and tools from April 2026 (week 15) should developers know about?

From 6 to 12 April 2026, Anthropic announced Claude Mythos Preview, a model it gives only to security partners through Project Glasswing (7 April), and launched two features for developers: Claude Managed Agents (8 April) and the advisor tool (9 April). Meta released Muse Spark (8 April), its first model since Llama 4 and a hosted one. Z.ai's open-weight GLM-5.1 (7 April) reported a higher SWE-bench Pro score than GPT-5.4 and Claude Opus 4.6 in its own table. Safetensors moved to the PyTorch Foundation.

Why does Anthropic keep Claude Mythos Preview behind Project Glasswing?

Anthropic announced Project Glasswing on 7 April 2026. The reason is Claude Mythos Preview, a general-purpose model that according to Anthropic surpasses all but the most skilled humans at finding and exploiting software vulnerabilities. Anthropic says it has already found thousands of high-severity vulnerabilities, including some in every major operating system and web browser. There is no public release. Launch partners such as AWS, Apple, Google, Microsoft, NVIDIA and the Linux Foundation, plus more than 40 other organizations that maintain critical software, get access for defensive work. Anthropic commits up to $100 million in usage credits and $4 million in donations to open-source security organizations. After the preview, participants pay $25 per million input tokens and $125 per million output tokens.

In Anthropic's own table, Mythos Preview scores 83.1% on CyberGym, a benchmark that tests whether a model can reproduce known vulnerabilities (Opus 4.6: 66.6%), 77.8% on SWE-bench Pro (53.4%) and 82.0% on Terminal-Bench 2.0 (65.4%). The Batch compared the approach to OpenAI keeping GPT-2 under wraps in 2019 and said it "has elements of a publicity stunt". Simon Willison expected the caution to be warranted. The post on Project Glasswing covers the initiative in more detail.

What did Meta release with Muse Spark?

Muse Spark (8 April 2026) is the first model from Meta Superintelligence Labs and, as Simon Willison notes, Meta's first model release since Llama 4 a year earlier. Unlike Llama, it is hosted and has no open weights: you use it in meta.ai and the Meta AI app, and the API is a private preview for select users. At launch, meta.ai offered an Instant and a Thinking mode. A Contemplating mode, in which several agents reason in parallel, was rolling out gradually.

Willison found tools wired into the meta.ai chat, among them a Python code interpreter in a sandbox and a visual grounding tool that returns points, bounding boxes or counts for objects in an image. According to The Batch, Artificial Analysis scored Muse Spark at 52 on its Intelligence Index: fourth, behind Gemini 3.1 Pro and GPT-5.4 (both 57) and Claude Opus 4.6 (53). It used about 59 million tokens for the index, against about 158 million for Opus 4.6. Meta itself acknowledges gaps in long-horizon agentic work and coding.

How does the open-weight GLM-5.1 score on coding?

GLM-5.1 from Z.ai (7 April 2026) is a model of 754 billion parameters under the MIT license, according to Simon Willison. You can download the open weights and run the model yourself. In Z.ai's own table it scores 58.4 on SWE-bench Pro, against 57.7 for GPT-5.4 and 57.3 for Claude Opus 4.6. Two caveats apply. Anthropic's own table gives Opus 4.6 a score of 53.4 on the same benchmark, so compare scores within one table only. And the unreleased Mythos Preview scores 77.8 there.

What changed for Safetensors?

Safetensors, Hugging Face's file format for model weights, joined the PyTorch Foundation on 8 April 2026 as a project under the Linux Foundation, next to PyTorch, vLLM, Ray and DeepSpeed. The format stores tensors behind a small JSON header and cannot run code when you load it, which the older pickle-based formats could. According to Hugging Face, it is the default format for sharing models on the Hugging Face Hub. For users nothing changes. The roadmap includes loading tensors straight onto GPUs, loading only the weights each GPU needs in tensor-parallel and pipeline-parallel setups, and work with the PyTorch team so that Safetensors can be used as a serialization format within PyTorch core.

What did Anthropic add for agent builders?

Claude Managed Agents (8 April 2026, public beta) is a set of APIs for running agents on Anthropic's infrastructure. You define the agent's tasks, tools and guardrails. Anthropic's harness decides when to call tools, manages the context and recovers from errors, with sandboxed code execution, checkpointing, credential management, scoped permissions and tracing. Sessions can run on their own for hours. The launch price was the standard token rate plus $0.08 per session-hour of active runtime. Anthropic's headline promises production agents 10 times faster, and in its internal tests on structured file generation, task success rose by up to 10 points over a standard prompting loop.

The advisor tool (9 April 2026, beta) pairs a cheaper executor model, Sonnet or Haiku, with Opus 4.6 as an advisor. The executor does the work and decides itself when to consult Opus, which returns a plan, a correction or a stop signal but never calls tools. In Anthropic's tests, Sonnet with an Opus advisor scored 2.7 percentage points higher on SWE-bench Multilingual than Sonnet alone, at 11.9% lower cost per task. On BrowseComp, a web research benchmark, Haiku with an Opus advisor scored 41.2%, more than double its 19.7% alone. That combination still scores 29% lower than Sonnet alone, but it costs 85% less per task.

Which tool connects agents to business data?

Dinobase, presented on Hacker News on 7 April 2026, is an open-source "agent-first database" under the MIT license. It syncs data from 101 sources, such as Stripe, HubSpot and Postgres, into one SQL layer on DuckDB, so an agent can join data across sources in a single query. A Claude agent annotates the tables and columns in the background, and every change an agent writes back goes through a preview and confirmation step. In the project's own benchmark over 11 LLMs and 75 questions, agents answered 91% correctly with Dinobase SQL, against 35% with separate MCP tools per source, at 16 to 22 times lower cost per correct answer.