Python Gains frozendict and Other Python News for March 2026

Python 3.15.0 Alpha 6: Comprehension Unpacking and More

Python 3.15.0a6 shipped on February 11, continuing the alpha series toward the May 5 beta freeze. This release includes several accepted PEPs that are now testable:

• PEP 798: Unpacking in comprehensions using * and **
• PEP 799: A high-frequency, low-overhead statistical sampling profiler
• PEP 686: UTF-8 as the default text encoding
• PEP 728: TypedDict with typed extra items

PEP 798 is the kind of quality-of-life improvement that makes you smile. It lets you flatten or merge collections directly in a comprehension:

>>> lists = [[1, 2], [3, 4], [5]]
>>> [*it for it in lists]
[1, 2, 3, 4, 5]

>>> dicts = [{"a": 1}, {"b": 2}]
>>> {**d for d in dicts}
{'a': 1, 'b': 2}

No more writing explicit loops just to concatenate a list of lists.

The JIT compiler continues to show gains: 3-4% on x86-64 Linux over the standard interpreter and 7-8% on AArch64 macOS over the tail-calling interpreter, matching the numbers from alpha 5.

Python 3.14.3 and 3.13.12: Maintenance Releases

On February 3, the team shipped Python 3.14.3 with around 300 bug fixes and Python 3.13.12 with about 240 fixes. No new features here, but if you’re running either version in production, it’s worth grabbing these patches to stay current.

PEP 814 Accepted: frozendict Joins the Built-Ins

This one has been on many Python developers’ wishlists for over a decade. PEP 814, authored by Victor Stinner and Donghee Na, adds frozendict as a built-in immutable dictionary type in Python 3.15.

The concept is straightforward. Just as frozenset gives you an immutable version of set, frozendict gives you an immutable version of dict:

>>> config = frozendict(host="localhost", port=8080)
>>> config["host"]
'localhost'

>>> config["host"] = "0.0.0.0"
Traceback (most recent call last):
    ...
TypeError: 'frozendict' object does not support item assignment

Because frozendict instances are immutable, they’re hashable when all their keys and values are hashable. That means you can use them as dictionary keys, store them in sets, or pass them to @functools.lru_cache()-decorated functions:

>>> permissions = {
...     frozendict(role="admin"): ["read", "write", "delete"],
...     frozendict(role="viewer"): ["read"],
... }

The new type supports union operators (| and |=), type annotations via frozendict[str, int], and equality comparison with regular dicts. Notably, frozendict does not inherit from dict, which keeps the immutability guarantees clean and avoids any accidental mutation through inherited methods.

Other PEP Updates

Beyond frozendict, the Steering Council made a couple more PEP decisions in February:

• PEP 815 Final – Deprecate RECORD.jws and RECORD.p7s: This cleans up wheel signature formats that no major packaging tool actually supports. Going forward, the recommendation is to rely on index-hosted attestations like those on PyPI.

• PEP 743 Rejected – Py_OMIT_LEGACY_API: The Council turned down a proposal to let C extension authors opt out of deprecated C API functions via a preprocessor macro, citing concerns about granularity and developer experience. They suggested better linter tooling as the more effective path.

Ending 15 Years of subprocess Polling

One of the month’s most satisfying reads was Giampaolo Rodola’s blog post on ending 15 years of subprocess polling. Since Python 3.3, the subprocess module has used busy-loop polling to check whether a child process has finished. It repeatedly calls waitpid(WNOHANG), sleeps, then checks again with exponential backoff from 0.1ms to 40ms.

Modern operating systems have proper event-driven mechanisms for this. Python 3.15 will now use pidfd_open() on Linux and kqueue() on macOS to get notified when processes exit, dropping context switches from 258 to just 2 in benchmarks. Rodola first implemented the approach in psutil, then upstreamed matching improvements to CPython. If you’ve ever wondered why subprocess.run() seemed to add unnecessary latency, this is your answer.

Dispatch From the Inaugural PyPI Support Specialist

Maria Ashna shared a candid look at her year and a half as PyPI’s first Support Specialist. When she started in July 2024, account recovery requests were four months behind, and PEP 541 (package name claims) had a 16-month backlog. She cleared both queues to current status, processing nearly 3,000 account recovery issues and approving over 8,000 PyPI Organizations applications.

It’s a great behind-the-scenes look at the human infrastructure that keeps the Python packaging ecosystem running.

MicroPythonOS: An Android-Like Experience on Microcontrollers

Following up on the MicroPython-in-a-PlayStation-game story from December, MicroPythonOS is a lightweight graphical operating system for microcontrollers like the ESP32. It delivers an Android-like user interface experience, complete with touch input and app-style navigation, all powered by MicroPython. If you’re into embedded development, you’ll want to check it out.

Django Security Releases: Three SQL Injection Fixes

The Django team shipped security releases on February 3 for Django 6.0.2, 5.2.11, and 4.2.28, addressing six vulnerabilities including three high-severity SQL injection bugs affecting:

• Raster lookups on PostGIS (CVE-2026-1207)
• Column aliases with control characters (CVE-2026-1287)
• QuerySet.order_by and FilteredRelation (CVE-2026-1312)

Two moderate denial-of-service issues and a low-severity timing attack in the mod_wsgi authentication handler were also patched. If you run Django in production, you’ll want to upgrade immediately. Three high-severity SQL injections in a single release are unusual and worth taking seriously.

uv 0.10.0: Breaking Changes and Stabilized Features

uv 0.10.0 landed on February 5 as a breaking-changes batch release. The biggest stabilized feature is uv python upgrade for seamless patch-version upgrades in virtual environments, along with uv workspace list and uv add --bounds.

Here are notable breaking changes to watch for:

• uv venv now requires --clear to remove an existing environment.
• PyPy and GraalPy executables install under their implementation name—for example, pypy3.10 instead of python3.10.
• uv format upgraded to Ruff 0.15.0 with the 2026 style guide.
• Python 3.8 Docker images are discontinued.

Anthropic: Structured Outputs GA and MCP Helpers

Anthropic’s Python SDK shipped four releases in February (v0.81.0–v0.84.0). Highlights include structured outputs, now generally available for Claude Sonnet 4.5, Opus 4.5, and Haiku 4.5, dropping the beta header requirement. The web search tool and fine-grained tool streaming are also GA.

The most practical addition for developers building MCP integrations is v0.84.0’s new helpers for converting MCP tools, prompts, and resources directly to Anthropic API types. If you’ve been manually wrangling tool schemas between MCP and Claude, this should save you some boilerplate.

OpenAI: WebSocket Transport and the Assistants Sunset

The openai-python SDK shipped eight releases in February (v2.17.0–v2.24.0). The headline is WebSocket transport for the Responses API (v2.22.0), enabling persistent, low-latency connections across multi-turn agent workflows.

More significantly, OpenAI announced plans to sunset the Assistants API in 2026 after achieving feature parity in the Responses API. If you’ve built applications on the Assistants API, it’s time to start planning your migration. The Agents SDK also dropped Python 3.9 support.

vLLM 0.16.0: Self-Hosted Realtime API

vLLM 0.16.0 is the largest vLLM release to date, with 440 commits from 203 contributors. The standout feature is a WebSocket-based Realtime API that mirrors OpenAI’s interface, making vLLM a drop-in self-hosted alternative for voice-enabled applications. Async scheduling with pipeline parallelism resulted in a 30.8% end-to-end throughput improvement.

Open-Source AI at Wagtail

On the community side, Thibaud Colas from the Wagtail CMS team published a thoughtful post on open-source AI tools they use for development. The post covers their experience with models like Qwen3 Coder and Mistral Small through providers like Ollama and LM Studio.

The central argument frames open-source model adoption as moving from “renting” to “owning” AI capabilities. If you’re evaluating AI coding tools for your own workflow, it’s a practical starting point.