The open-weights race intensified on both sides of the Pacific. Reflection AI released the Apache 2.0 weights for Beam, its 501B-parameter (23B-active) mixture-of-experts model, which Artificial Analysis now ranks #2 among US open-weight models on its Intelligence Index. Hours later, DeepSeek shipped DeepSeek-V4.1-Flash-1017 as an open-source release with stronger benchmarks and shorter outputs, and opened its Search & Fetch APIs to everyone.
OpenAI upgraded GPT-6 Astra Pro in ChatGPT with faster responses, better tool and plugin use, and parallel sub-agents users can watch and review. Anthropic's STEM Fellows Program showed Claude Mythos 5 designing quantum circuits that beat the best published designs on seven of eight computational primitives — using fewer qubits and fewer gates, sometimes by more than an order of magnitude.
Google DeepMind demonstrated four Gemini 4 Argon agents autonomously beating Minecraft end-to-end in 27 hours and 14 minutes, writing 2,783 programs with no human intervention. NVIDIA's RTX Spark AI laptops hit retail with six-city launch celebrations, while Microsoft published Nature research showing machine-learning weather forecasts improved two to six weeks out.
Reflection's first model is a 501B-parameter sparse mixture-of-experts system with 23B active parameters, built for coding and agents. The weights are now available across platforms, with SGLang serving Beam in production at Reflection, and Artificial Analysis scores it 30 on its Intelligence Index — #2 among US open-weight models — plus a US-open-weights-leading 34% on its Cyber Index.
ChatGPT's top-tier model received faster responses and improved plugin and tool use, and multiple agents can now tackle different parts of a task in parallel, with progress tracking and reviewable completed work. The upgrade is available to Pro users in the Chat tab, and the official announcement drew 9,153 likes and roughly 1.3 million views in under a day.
Anthropic's STEM Fellows Program ran Claude Mythos 5 in a closed loop with an automated evaluator, giving plain-language instructions for eight quantum circuit tasks. Claude's designs beat the best previously published algorithms on seven of eight, using fewer qubits and fewer gates — sometimes by more than 10x — across tasks spanning chemistry, number theory, and numerical computing.
DeepSeek's latest checkpoint posts stronger benchmark performance, faster task completion, and shorter outputs than its predecessor — a Pareto improvement, per the lab. The release uses reinforcement learning with self-feedback to improve end-to-end project development, and DeepSeek simultaneously opened its Search & Fetch APIs, powered by its in-house search engine serving hundreds of millions of daily requests, to everyone.
A Google DeepMind paper reports four Gemini 4 Argon agents completed Minecraft end-to-end, defeating the Ender Dragon after 27 hours and 14 minutes with no human intervention. Operating in a sandbox using only code and text, the agents wrote 2,783 programs and divided work autonomously — a demonstration of long-horizon multi-agent collaboration, though a sandboxed one rather than real-world task evidence.
Selected posts from the edition, with links to the originals.
🔍 Search & Fetch APIs are now open to everyone! Powered by our in-house search engine, serving hundreds of millions of requests daily worldwide.
View post on X ↗What is ChatGPT Space, that is the question. Endearingly awkward at times because it's the actual team working on it. What would you like to see next?
View post on X ↗One week since launch, Step 5 Preview is already #2 on OpenRouter's weekly leaderboard, with 19.5T tokens used. It has also held #2 on the daily leaderboard.
View post on X ↗Exciting news: Muse Spark Open (Max) by @AIatMeta is the #1 US open model in the Code Arena: WebDev! With 1511 pts, it's #13 among open models overall.
View post on X ↗We are open-sourcing V1.0 of our agentic robotics framework, ASPIRE, with full real robot skill discovery and self improvement capabilities! On the da…
View post on X ↗In @Nature this week, our new research shows how machine learning can improve weather forecasts two to six weeks out, a timeframe that's historically…
View post on X ↗MAI is the most accurate & most popular transcription model on OpenRouter across all languages and use cases. Try it out for your agents now! 👇
View post on X ↗Put your dot on call. Ali, an engineer at OpenAI, used his dot to triage an on-call incident at 3am. His dot investigated the issue and started a Cod…
View post on X ↗Turns out le Chonk is into rocket science. Mistral Large 4 scoring well in engineering and astronomy on Ulam's private benchmark, STEMBench.
View post on X ↗We've partnered with @ndstudio to launch American Forge. It's a six-month residency for exceptional engineers to embed inside the government and buil…
View post on X ↗