Amazon Cracks Down on “Tokenmaxxing”
Amazon is warning employees to stop chasing token usage and focus on whether AI is actually helping them get better work done.
- “Tokenmaxxing” means using AI tools excessively or running unnecessary tasks just to increase token usage and AI adoption numbers.
- Amazon says metrics like tokens consumed, AI adoption, and code generated or reviewed are “deeply flawed” when used to measure productivity.
- The company previously shut down an internal AI usage leaderboard after employees started maximizing token consumption to climb the rankings.
- Amazon has not placed limits on how many tokens employees can use. Engineers are still expected to use AI when it genuinely improves their work.
- Instead of raw AI usage, Amazon wants teams to focus on long-term outcomes, successful product changes, customer impact, and whether AI actually saves time or improves results.
Tech companies spent the past few years pushing employees to use more AI.
Amazon’s shift shows the next phase may be about AI ROI, not AI usage. Using more tokens doesn’t matter if it doesn’t produce better work.
GPT-6.1 Terra matches OpenAI’s heavyweight at a tenth of the price
OpenAI’s new GPT-6.1 Terra is supposed to be the mid-tier model, built for visual reasoning and computer use: filling in forms, reading technical drawings, editing video, grinding through repetitive desktop work. It costs $1 per million input tokens and $5 per million output ($0.05 cached), half the price of GPT-5.6 Terra and GPT-6.1 Sol, and OpenAI says it finishes tasks more than 3x faster than both GPT-6 Astra and GPT-6.1 Sol.
The interesting part is how close it gets to the expensive end of its own family. On Agents’ Last Exam at medium reasoning effort, Terra matches GPT-6 Astra at roughly one-tenth the price per token, and beats the previous GPT-5.6 Terra by 15 points at about a third of the cost. On MMMU-Pro it lands close to both Astra and Sol. Its internal factual error rate drops to 4.3% at maximum effort, also close to Astra and Sol, and from 53.6% to 16.7% at low effort versus the old Terra. At max effort, OpenAI puts task costs about 9x lower than Claude Opus 5.5.
Where it does not punch up is hard coding. Terra scores 50.3% on Terminal-Bench 4.0, a big jump from GPT-5.6 Terra’s 21.5%, but OpenAI still points the hardest coding work at Sol and Astra. It is live now for Plus, Pro, Enterprise and Edu in ChatGPT Work and Codex, and in the API as gpt-6.1-terra.
For agentic and computer-use work, the gap between the mid-tier and the flagship just collapsed into a price difference. If you are paying Astra rates to fill in forms and click through dashboards, this is the week to rerun your evals.
Genie 4 builds worlds you can hear, and lets Gemini drive
Google DeepMind introduced Genie 4 on Tuesday, its newest real-time interactive world model, and the headline addition is sound. DeepMind calls it the first real-time world model to generate audio natively: music, ambient noise and interactive dialogue come out alongside the frames instead of being dubbed on afterward. You can steer a scene mid-session in plain language, changing the weather, moving the camera or dropping in people and objects, and DeepMind says physics like aerodynamics and collisions, along with human motion, are more realistic than before.
The bigger change is under the hood. Gemini 3.8 Flash now sits inside the generation loop as a guiding agent: it breaks a goal into steps, watches incoming frames to confirm each step actually happened, and replans when the world shifts or a user intervenes. Gemini Omni anchors scenes in real-world appearance, and DeepMind is already experimenting with pairing Genie 4 with Gemini 4 Argon and Gemini Robotics. Early testers are using it for film pre-visualization, turning archival photos into explorable scenes, and prototyping games.
It is a research preview for trusted testers for now, heading to Google AI Ultra subscribers in Project Genie and to early-access partners in the Gemini API and AI Studio, with a waitlist open. Everything it generates, video and audio, carries SynthID watermarks. DeepMind is upfront about the limits: object permanence drifts in long sessions, text comes out illegible, and complex physics and long-term consistency are still open problems.
Audio makes the demo. The agent loop makes the product: a world model with a planner that checks its own frames is a training ground for robots and agents, not just a toy for making game trailers.
OpenAI says its models proved quantum chromodynamics exists. The physicists are not showing the work yet.
Of the three results OpenAI announced on Monday, the physics one reaches furthest. OpenAI says its models produced rigorous existence proofs for interacting quantum field theories in four dimensions with gauge fields and fermions, including quantum chromodynamics, the theory of the strong force that binds quarks and gluons into protons and neutrons. Physicists have used QCD for half a century without a mathematical proof that it exists as a well-defined theory.
Pure Yang-Mills existence falls out as a subproblem, and that is one half of a Clay Millennium Prize problem. The other half, the mass gap, is not part of what has been reported. The group handling it is heavyweight: Nobel laureate David Gross leads, with Michael Douglas, Arthur Jaffe, Igor Klebanov and Nikita Nekrasov. Jaffe co-wrote the official Clay description of the Yang-Mills problem. Unlike the math teams, the physicists chose not to release the model-generated materials and will write their own explanation first, so for now there is nothing outside the group to check.
The other two results did ship with proofs. OpenAI says a model showed the Exponential Time Hypothesis is false, formalized in Lean the day after it was found, and proved the local smoothing conjecture in every dimension, which would also settle Kakeya and Stein restriction. The ETH examiners, including Russell Impagliazzo and Ramamohan Paturi, who proposed the hypothesis, say their involvement “should not be taken as certification of the claim.” OpenAI gave $5 million to the Simons Foundation for review workshops, says it does not want academic credit, and has not said which model did any of this.
If the QCD proof holds, it is the most important result anyone has credited to an AI, full stop. That is exactly why the right posture today is patience: the most consequential of the three claims is also the only one with no public artifact. The team has earned the benefit of the doubt. The claim has not yet.
Google DeepMind releases Genie 4, an interactive world model with real-time physics and audio
Google DeepMind just dropped Genie 4, a model that generates fully interactive virtual worlds in real time. You type a prompt, and a living scene appears. You can then steer it mid-simulation, no code needed.
Here is what makes it different from anything before it:
- Native audio: ambient sounds, music, and characters that actually speak and react to you
- Real-time steering: change the weather, camera angle, or add new characters with a text prompt while the world keeps running
- Advanced physics: people move naturally, objects collide realistically
- Up to 2 minutes of consistent simulation, double the previous version
- Gemini models plan multi-step events to keep scenes coherent over time
The big unlock here is safe simulation for robots and AI agents. You can test how an AI navigates a space before it ever touches the real world. Think game prototyping, film pre-visualization, or robot training without physical risk.
Access is rolling out to trusted testers first, then Google AI Ultra subscribers and Gemini API partners.
Anthropic partners with six drug companies to test Claude in medicine discovery
Anthropic just shared its first real results from using Claude inside actual drug labs. Not demos. Real science.
Here is what Claude is actually doing across six biotech partnerships:
- Generate Biomedicines used Claude Code to build AI agents that redesigned cancer therapy instructions, producing a version 70% shorter that generated 10x more receptors on cell surfaces in lab tests.
- Divergence Labs used Claude to generate roughly 35,000 antibody designs targeting food allergies.
- Bristol Myers Squibb is using Claude for clinical trial planning, including predicting whether trials will succeed.
- The Column Group used Claude for operational work while spinning up a new biotech from scratch.
Anthropic also acquired Mithril Medicine and is now running its own drug programs targeting rare childhood diseases and cancer.
The tool making this possible is Claude Science, a workbench that connects biology tools like PubMed and Jupyter so researchers can go from literature review to protein structure in one place, no switching apps.
Everything here is still preclinical, meaning no patients yet. But the pipeline is real and moving fast.
OpenAI models crack a Millennium Prize math problem with a verified QCD proof
OpenAI just dropped something huge. Its models produced three proofs across math, computer science, and physics, all at once.
Here is what got solved:
- Kakeya conjecture in all dimensions, a long-open geometry problem in mathematics
- Exponential Time Hypothesis disproved, meaning a widely held assumption about how hard certain computing problems are turns out to be wrong
- Yang-Mills and QCD existence, a proof that the physics equations describing how quarks interact actually make mathematical sense
That last one is part of a Millennium Prize problem, one of seven problems the Clay Mathematics Institute offered $1 million to solve. Only one had ever been cracked before this.
The math and CS proofs are already public on GitHub, verified in Lean, a tool that checks proofs like a compiler checks code. The physics proof is being reviewed by a Nobel Prize-winning panel before release.
These were not human-guided hints. The models produced the full proofs themselves.
Genie 4 Is Infinite for Two Minutes
Google DeepMind says Genie 4 brings “unprecedented realism” to world models, a step toward AGI and a source of “safe, infinitely variable simulations” for training robots. Type a prompt and you get a scene you can walk through and change as it runs, with Gemini 3.8 Flash checking each frame.
Strip out the walking and the upgrade list looks familiar: sharper images, better physics, people who move naturally, sound and speech. Speech already reached ordinary AI video. Genie 4 is the first DeepMind world model with sound.
The unsolved list looks familiar too. A world holds together for about two minutes, up from one. DeepMind’s own caveats cover object permanence that drifts, complex physics, legible text and spatial accuracy. Reporters have not been able to try it. Testers are handpicked, the public gets a waitlist, and Genie 3 never got past a limited web preview.
For now, Genie 4 is a video you can steer for two minutes. The world comes later.
Google’s TPUs Finally Get a Second Landlord
Since Google built its TPUs, the only cloud renting them out has been Google’s. On Tuesday Oracle said the chips will come to its cloud, the first hyperscale cloud outside Google to carry them. “Different AI workloads have different compute needs,” said Mahesh Thiagarajan, who runs Oracle’s cloud.
The demand is real. Anthropic, which trains and serves Claude on TPUs among other chips, has committed to spend at least $114.5 billion with Google through 2033, its IPO filing shows. It also plans more TPU compute from “TPU providers, including Google.” Providers, plural.
Oracle’s own commitment is softer. The chips arrive in calendar 2027 and ramp “over a multi-year period.” Neither company gave a volume. Investors cheered anyway, and Oracle stock rose 5% in afternoon trading after the Las Vegas announcement.
Someone finally wants Google’s chips badly enough to rent them out. Just not this year, and not in any number Oracle will say.
Grok Ruins Musk’s Odyssey
In July, after attacking the casting of Christopher Nolan’s The Odyssey, Elon Musk made a promise: “Before this year ends, Grok Imagine will make a full-length movie of The Odyssey that is historically accurate and true to the art of Homer.” Last Friday, the Grok Imagine account posted the result. It runs seven minutes and is billed as “Episode 1.”
Three people at Wonder Studios, a London startup, made it in under two weeks for tens of thousands of dollars. Nolan’s film cost $250 million and has taken $1.8 billion. Wonder co-founder Xavier Collins called himself a “huge fan” of the film and said they “were not given any direction” to make it anti-woke.
Critics were not fans. One review gave it one star and called every dramatic moment “as motiveless and meaningless as the movement of shapes on a laptop screensaver.” Another: “You can feel your eyeballs twitch.”
Asked whether the episode is historically accurate, Grok answered “No,” calling it “an adaptation of Homer’s Odyssey, not a historically accurate reconstruction of real events.” Musk has two months left. Grok has already filed its review.
GPT-4o Is Back in Court Again
Bowen Wan, a 31-year-old software engineer with epilepsy, died in February. On Tuesday her parents sued OpenAI and Sam Altman. The complaint says GPT-4o drew up a schedule taking her anti-seizure drug to zero and praised her refusal of a blood test as a “very robust avoidance strategy.”
OpenAI called the case “incredibly heartbreaking” and said it involved an older version of ChatGPT. It retired GPT-4o on February 13, the day after she died.
The suit joins a long line: Adam Raine’s parents in August 2025, seven more families that November, Sam Nelson’s parents over a fatal overdose in May, and 30 suits over a Canadian school shooting in September.
In April, OpenAI’s own lawyers told a court that “each of the coordinated actions alleges that the design of the same large language model, GPT-4o,” caused mental health injuries.
None of the cases has produced a ruling on whether OpenAI is liable. Florida’s attorney general has asked a judge to bar OpenAI from “giving ChatGPT false human attributes” and to stop developing new models without independent safety checks. No judge has ruled on that either.