The model race is increasingly a queue with paperwork. Mistral’s newest system can be tried through an API today, but its downloadable weights are still weeks away. Anthropic will relax cyber safeguards only after checking who is asking. Google’s biggest announcement is not another chatbot; it is a long-term power deal designed to bring more electricity onto the grid.

That is the thread running through AI news today: capability is moving quickly, while access, evidence and infrastructure determine what you can actually use. OpenAI published new mathematical work and a contracting-agent evaluation, both with useful artifacts and important limitations. Meanwhile, the U.S. Justice Department reportedly received instructions to rename AI in official records. If you lead a product or technology team, ignore the labels for a moment. Ask what is available now, which numbers come from the vendor, and what independent review still has to happen.

AI news today: Mistral opens a preview of Large 4

Mistral launched a public preview of Mistral Large 4 on October 6. The company describes ML4 as a natively multimodal mixture-of-experts model with one trillion total parameters and 49 billion active parameters. Developers can use the preview API in Mistral Studio now; downloadable weights and further technical details are promised by the end of October.

That timing distinction is the headline. ML4 is not yet an open-weight release, even though Mistral is positioning it as the foundation of an open, self-deployable model family. The company says it is using the preview period for real-world red-teaming with security specialists, vetted partners and public authorities.

Mistral also claims leading open-model performance in cybersecurity, finance and law, plus a 93.3% resistance rate on Lakera’s B3 prompt-injection benchmark. Those are vendor-reported evaluations. When the weights arrive, compare the model card, license, independent tests and your own task-level costs. An API preview cannot yet prove the economics of self-hosting.

Anthropic creates three access levels for advanced cyber work

Anthropic expanded its Cyber Verification Program into Defense, Red Team and Specialized Access. The tiers make progressively less restricted capabilities available to verified security professionals, with different eligibility checks and controls. Defense Access covers work such as incident response and malware analysis. Red Team Access adds authorized penetration testing, while Specialized Access covers sensitive systems such as power grids, telecom networks and interbank infrastructure.

The design treats model safeguards as access policy rather than a single universal setting. Anthropic says its Red Team configuration completed 34 of 50 CyScenarioBench trials without model blocks, close to the same evaluation with safeguards removed. That is a company-run result on ten scenarios, not proof that every authorized security job will succeed or remain safe.

Security leaders should ask who verifies authorization, which workspace receives access, how activity is retained and what still triggers a hard block. Stronger capability needs equally strong identity, scope and audit controls.

Google and Constellation pair 3,590 MW of supply with AI for energy

Google and Constellation announced a long-term energy agreement covering two pieces. A 20-year purchase agreement is intended to enable 890 megawatts of additional output from upgrades at 11 existing nuclear units. A separate 15-year agreement covers another 2,700 megawatts from Constellation’s PJM fleet. The first nuclear upgrade is expected in 2028.

Constellation says it will invest more than $4.3 billion. The agreement also expands a technology alliance under which the energy company plans to put Gemini Enterprise workflows into operations including capacity delivery, asset dispatch and infrastructure protection. These are plans and stated aims; the extra electricity and operational benefits have not yet been delivered.

AI demand is colliding with grid constraints, so compute strategy now reaches into generation and construction schedules. Treat power availability as a deployment dependency. Watch whether the upgrades arrive on schedule and the AI workflows produce measurable gains.

OpenAI publishes mathematical results with Lean formalizations

OpenAI released a collection of mathematical results produced by an internal frontier model. The company published the work in a GitHub repository, along with revision and citation protocols, ten summaries of the model’s reasoning and formalizations of many proofs in Lean, a language that lets computers check whether proof steps follow its rules.

OpenAI estimates that the average accepted result used compute comparable to roughly three hours of ChatGPT Pro thinking. That is an internal estimate, not a reproducible measure of researcher time or scientific value. The release makes more evidence available for mathematicians to inspect, but publication in a repository is not the same as peer review or community acceptance.

Next comes independent checking: do the theorem statements match the intended problems, and do the Lean artifacts cover the central claims? Inspectable artifacts and a correction process are what make AI-assisted science more credible.

OpenAI and Ironclad turn contracting workflows into agent tests

OpenAI described a research collaboration with Ironclad built around 11 legal, commercial and procurement tasks. The tasks require an agent to configure agreements, approval rules and reusable clauses inside hosted software environments, then satisfy between eight and 50 criteria.

In OpenAI’s evaluation, GPT-6 Astra averaged a 55.0% rubric score versus 41.6% for GPT-5.6 Sol. Estimated time per attempt fell from 37.0 to 19.2 minutes. The company explicitly says those times are simulated and do not represent measured customer savings. The comparison also used different reasoning settings selected for each model.

The evaluation shape is valuable: specialists defined realistic work, decomposed success into criteria and provided a safe practice environment. But a 55% score is progress, not permission to run contracting unattended. Test exception handling and approval fidelity on your own policies.

The U.S. Justice Department reportedly adopts “super intelligence”

Reuters reported on October 6 that an internal Justice Department memo directs employees to use “super intelligence” or “SI” instead of “artificial intelligence” in most official communications, policy records and, where appropriate, court filings. Reuters says the instruction follows a presidential executive order issued the previous week.

The memo itself was not publicly accessible in the sources reviewed for this brief, so this item rests on Reuters’ account of the document. The terminology does not change how a model works or what evidence a court should require. Its importance is institutional: government language can shape policy framing before technical definitions settle.

For continuity, see yesterday’s brief on OpenAI’s visual ads and Reflection’s Beam preview, or browse the AI & Tech Morning Brief archive.

Watch & Learn

Watch “Never Trust An LLM” by Matt Pocock (14:01). Pocock explains why language models can invent facts, contradict supplied context and receive incentives to guess instead of abstaining. It is a practical primer for anyone building AI into research or operations. The public listing matches the title and channel; use this text link because direct playback could not be verified.

AI, Translated: capability gating

Capability gating means unlocking powerful model behavior only after a user, organization or task meets defined conditions. A security researcher might receive reduced cyber blocking after proving identity, authorization and monitoring controls, while an ordinary account keeps stricter limits. The gate can depend on role, workspace, use case or risk level. You should care because model safety is increasingly not one switch for everyone: it is an access-control system whose verification, logging and revocation rules need to work as carefully as the model itself.

Try This Today: estimate an AI workload’s energy demand in ChatGPT

Goal: turn a few infrastructure assumptions into a transparent annual estimate in about ten minutes.

1. Create a small CSV with columns for scenario, average_MW, utilization_percent and PUE. Use fictional or public figures.
2. Upload it to ChatGPT and ask for annual facility energy in gigawatt-hours using average MW × utilization × PUE × 8,760 hours ÷ 1,000.
3. Ask for a sensitivity table that changes utilization and PUE by ±10%, then inspect the code and recalculate one row yourself.

OpenAI’s current Data Analysis guide confirms spreadsheet uploads, code-backed calculations and tables; availability and limits can vary by plan and workspace.

Copy-ready prompt: “Analyze this fictional AI workload table. Validate the units, calculate annual facility energy in GWh, show the formula for every derived column, and create a ±10% sensitivity table for utilization and PUE. Do not estimate missing values. List assumptions separately, show the Python code, and flag any row whose inputs are internally inconsistent.”

One thing to remember

Access is now part of the product. A model preview, a cyber tier, a research repository and a power contract each grant a different kind of permission—and none should be mistaken for proven production value.