Apollo's economist says AI is lifting revenue per worker and margins only at the Mag 7, while the other 493 companies in the S&P stay flat. For builders, the day's research points to cheaper agents: solving repeated decisions once took an open model from 29% to 96% with 93% fewer model calls.
Nanbeige released Nanbeige4.3-33B, open weights built as a teacher model whose skills will later be distilled into smaller ones, such as a planned Nanbeige4.3-3B. On Terminal-Bench 2.1, a test of coding agents working in a terminal, it scores 80.9 against 73.0 for Qwen3.8-27B. It is not ahead everywhere: on SWE-Bench Pro it scores 60.3 to Qwen's 61.7. Context runs up to 524,288 tokens.
fal Research and Black Forest Labs launched FLUX 3 Max, a video model post-trained on top of FLUX 3 for text-to-video and image-to-video. In fal's measurements it generates a 10-second video in 12.45 seconds, against 277.80 seconds for Seedance 2.5. On Artificial Analysis's image-to-video ranking it gains 47 Elo points over the original FLUX 3, which fal says puts it alongside the top models.
Bellman language models treat an agent's recurring choices as a planning problem. The method reads past conversations into a fixed set of states, solves the best move for each once, and runs that as a controller around the model; unseen states go back to the LLM. On a telecom support benchmark, Qwen3.8-27B rose from 29% to 96% with 93% fewer model calls. It helps little where the hard part is writing content, such as fixing SymPy issues.
UserSWE tests coding agents on fixing repository bugs over several turns with a simulated user who reveals details gradually, with tests run after every turn. Across eleven agents on 72 tasks from 42 repositories, the leader changed with the number of turns allowed. Claude Fable 5.1 solved the most tasks in one turn, GPT-6-Astra the most by the eighth, and Claude Opus 5 led at most budgets in between.
In long agent tasks, training usually only learns whether the whole attempt succeeded, which makes it hard to tell which steps helped. SCALE adds a critic that predicts, at each step, the chance the task will still be completed, and turns changes in that chance into credit for individual actions. Across five agent benchmarks and three model backbones, it beat the compared reinforcement learning baselines in most settings, with gains of up to 4.1 percentage points.
Apollo's chief economist says revenue per worker and profit margins are rising for the Mag 7 but flat for the other 493 companies in the S&P, the group it calls the S&P 493. There are still no signs of AI lifting productivity growth. Apollo says mainstream companies are starting to treat token spending as a cost center that weighs on margins, and that the AI story will be decided in the S&P 493.
ElevenLabs launched Eleven Realtime, speech-to-speech models where each model is a complete voice agent, replacing a chain of speech recognition, an LLM and text to speech. v1 Max answered 98.8% of spoken Big Bench Audio questions and started speaking in 0.57 seconds on average; v1 Mini scored 75.9%. On tool-use tasks Max completed 87.1% against 60.3% for Mini. Its API follows OpenAI's Realtime API, so switching means changing an endpoint and model name.
Bitcoin Core now classes some vulnerabilities as easily discoverable, meaning they can be found with minimal effort, such as through a simple LLM prompt. Those get a fixed point release for each maintained major version as soon as possible, with a security advisory, skipping the usual embargo. Other bugs keep the old timelines: low severity issues are disclosed 2 weeks after a major version with the fix ships.
The Association for Human Mathematics, which formed in opposition to AI companies' math work, now lists 1008 members, and 654 of them belong to its AI-free caucus. All members commit not to work with commercial AI companies and not to publish math texts generated by AI models. Caucus members go further: they find proofs, explore examples and pose research questions without prompting AI tools.
Anthropic published a starter kit for moving a router-plus-specialists agent system to one agent loop that loads skills and tools when needed. It works in three steps: measure the current system on an eval, merge subagents that always run together, then turn each remaining one into a skill and tools, one at a time. A subagent only goes away when the numbers say its replacement holds. The kit's sample support agent has a router and six subagents.
You’re caught up.
That was all 10 stories in the 26 October issue.