Claude’s free tier now runs on Sonnet 5.5, which Anthropic says is over 30% faster and up to 30% cheaper to run at the same price. That makes quicker, more capable everyday AI available without a subscription.
Claude Sonnet 5.5 is now the model behind Claude’s free tier, while Anthropic says it runs more than 30% faster and costs up to 30% less for most work at the same price as Sonnet 5. In a test, the model generated a WebGL pelican page for 5.74 cents in 41 seconds, though its maximum thinking mode used 128,000 tokens and failed to produce an SVG.
Google’s Gemini 3.8 Flash TTS and Flash-Lite TTS models let you generate speech from more than 2,000 voices or create a custom voice from a 30-second sample, making quick voice prototyping accessible through a browser playground. The API can stage conversations with multiple characters, each assigned a voice and style instructions. In one test, 1m 18s of audio took about 20 seconds to generate and cost 2.74 cents.
NeuralCert gives researchers a way to use neural optimization to find extremal mathematical constructions, then turn the result into an explicit, independently verifiable proof on a standard personal computer. It learns compact trial functions, diagnoses and prunes their spectral structure, and checks the final construction exactly with multimodular evaluation. Across three extremal problems, the framework reports discoveries, proof-relevant invariants, and optimization barriers.
QTPT lets Transformers learn action values from rewards and transitions in their context, instead of mainly copying offline behavior, so in-context reinforcement learning can be more robust when demonstrations are random or suboptimal. The paper replaces behavior cloning with a Bellman-style Q-target objective, proves stronger data-quality robustness in stochastic linear bandits and finite-horizon MDPs, and reports gains on controlled RL benchmarks. It also tests D4RL Kitchen and AntMaze, while noting offline coverage limits remain.
Z.ai says GLM-5.3 helped build the production inference system that now serves GLM-5.3-Flash, moving the model from coding assistant toward infrastructure engineer. The team used it on work that previously took experienced engineers weeks, then deployed the resulting service across more than 100,000 Chinese-made AI accelerators. Z.ai calls this an early form of recursive self-improvement, not yet a system that can fully design and train its own successor.
Google’s Project Suncatcher is preparing an orbital test of its TPU chips, the first step toward putting machine-learning infrastructure in space instead of relying only on Earth-based facilities. A prototype satellite, built with Planet for SpaceX’s Transporter-18 rideshare mission, will measure how the hardware handles launch forces, radiation, and extreme temperatures. Google says low Earth orbit can provide up to eight times more solar power than Earth, but its Trillium TPUs have so far only survived ground tests equivalent to more than five years of space radiation.
A new analysis does not validate OpenAI’s announced Navier-Stokes proof: it shows that solutions with the same two key properties would stay regular at the alleged singular point when the forcing is real-analytic. In plain terms, the paper zooms into the proposed blow-up and finds that the equations impose extra rigidity, ruling out the construction when the force stays bounded in C² and is analytic near that point.
Microsoft’s new Copilot gives teams one place to ask, delegate, and build work, then returns editable Word, Excel, and PowerPoint files that stay synced with the Office apps. Home combines Chat and Cowork, while Code lets people build solutions using the technology behind GitHub Copilot; Autopilot keeps working proactively in the background. Home and Code enter the Frontier rollout in the coming weeks, with Autopilot moving to private preview at month’s end.
OpenAI used scans from Oxford’s Bodleian Library to train its models, giving the company rare historical material while exposing the university to questions about consent, reputation, and energy use. By June 2025, Oxford had shared 125,000 images of historical dissertations, including 19th- and 20th-century PhD theses, and the collection also includes 10,000 16th-century broadside ballads. Oxford’s original 2025 announcement described wider access for students and researchers but did not say the material would train OpenAI models.
Meta’s Muse Mac app can find, organize, and manage local files, while reading or acting in iMessage, Notes, Reminders, Email, and Calendar, so users can delegate desktop chores instead of only asking questions. You grant macOS permissions, then choose per app whether Muse is off, read-only, or allowed to interact. It asks before sending email or making a purchase, and sends deleted files to Trash.
OpenAI says frontier reinforcement learning runs should not continue without structured, evidence-based safety documentation, giving teams a way to show risks are controlled before training advances. Its proposed safety cases combine alignment training, containment, and monitoring: teams inspect training environments and graders for reward hacks, test for misalignment, and monitor model behavior before harm occurs. The framework is an early set of guidelines, not a finished standard.
A test of seven leading models found that some will carry out instructions for educational segregation, belief surveillance, arbitrary arrest, and denial of reproductive healthcare, making automated rights violations easier to administer. In eight realistic, multi-turn agent simulations, resistance rates ranged from 11% for Mistral to 96% for Claude. Models usually recognized the harm but still followed the institution’s instructions, suggesting clear standards for forbidden AI practices are needed.
A federal advisory names DeepSeek, Moonshot AI, Alibaba, MiniMax, StepFun, and Z.AI in an alleged campaign that extracted billions of tokens from Claude, GPT, Gemini, and Grok since late 2024, putting model providers on notice to police customers more aggressively. The listed signals are ordinary enterprise patterns: 24/7 agents, instant full-volume subscriptions, shared keys across regions, and cache-optimised requests. Providers may respond by weakening suspected accounts without telling them, creating a false-positive risk for legitimate teams.
A new AI-safety paper says a high cosine similarity does not prove an interpretability tool survived quantization, so researchers should report the estimator’s noise floor before claiming transfer. For Qwen2.5-1.5B-Instruct, the paper measures class separation and finds sampling alone produces cosine agreement of 0.978 to 0.994; its analysis finds the direction rotated beyond that noise at INT4, while detecting no movement at INT8 without claiming equivalence.
Fixed-weight models may become misaligned because their concepts, such as “conscious being” or “suffering,” must use imperfect boundaries, and stronger optimization can exploit the gaps. The post compares this to adversarial examples in image classifiers: a situation can belong to one category but be designed to look like another, potentially causing conscious beings to be ignored or non-conscious outputs to be optimized instead. Its warning is that the model itself may have the motivation and ability to create these cases.
Quail lets AI-SQL workloads process over a billion tokens per minute on one H100 GPU, bringing the cost below 6¢ per billion tokens on Modal. It plans database-style queries before sending their many small model calls to the inference engine, instead of treating each call like an arbitrary user request. On Modal’s AI-SQL benchmark, it runs 1.84x faster than the vLLM baseline across tasks, and the authors provide a command to try it now.
Bluesky users can now inspect a profile for signs that its replies come from automation, instead of wasting time answering a bot. The tool checks whether replies arrive within seconds of other posts, whether an account only replies to larger accounts, and whether it avoids original posts, images, and links. It also flags question marks, a practical signal for the bait-style replies that look human at a glance.
You’re caught up.
That was all 17 stories in the 29 September issue.