Carbon Opens Beyond Cybersecurity
8 stories · ~7 min read

Listen
If You Only Read One Thing
Google's strongest coding model is easier to buy than its strongest security guarantee. Gemini 4 Carbon now has a public paid API; AlphaProof Code's Signal verification required a carefully specified target and machine-checked proofs. The model can attempt more work immediately. Getting a result that somebody else can trust still depends on what the surrounding system is able to check.
Gemini Carbon Opens at a Temporary Discount
Gemini 4 Carbon has moved from a restricted cybersecurity program into ordinary paid API access. That makes Google's stronger model a real alternative for difficult coding and document work. It doesn't make Carbon the economical default for everything.
The November 10 API release notes name the stable endpoint, gemini-4-carbon. The launch price is $2 per million input tokens and $10 per million output tokens through November 30. Google's rate card doubles both on December 1, including the charge for thinking tokens. A workflow that fits November's budget can miss December's without changing its prompts.
This is the access change anticipated in October 29's briefing, when Carbon's strong coding results still came with restricted availability. Repeating those scores would miss the development: paying developers can now test whether that performance survives their own tools and assignments.
There's also fresh evidence beyond coding. Artificial Analysis's Multimodal Knowledge Agent evaluation, developed with Google DeepMind, tests 200 private tasks involving visually complex files and finished documents. Carbon satisfies 54% of the harder weighted criteria, against Fable 5.2's 57%, for roughly one-fifth the reported task cost. These are rubric criteria, not percentages of flawless completed jobs. An automated judge grades the outputs.
That comparison makes Carbon interesting for research that ends in a spreadsheet or presentation. It also argues against picking one universal winner: different models lead different industries. The repeatable practice is to compare the deliverable the job actually requires, keeping the files, tools and grading fixed while changing the model.
The price comparison needs care. Carbon's general benchmark page lists the higher $4/$20 token rates. Its quoted task bill isn't a receipt at Google's temporary launch price. Nor can a document benchmark settle the cost of repairing a private repository.
Google itself describes Argon as the choice for most work and Carbon as the model for harder assignments that merit more investigation. That split is credible: extra reasoning buys something only when it changes an accepted result. After December 1, the comparison worth keeping is cost per accepted deliverable on the same retained jobs, with repair attempts included.
AlphaProof Code Checks Signal's Compiled Security Properties
Google's more consequential agent result is a way to check work without trusting the model's explanation. AlphaProof Code connects security requirements to the compiled code of Signal's encryption library. The agent writes the proof; deterministic software decides whether the proof holds.
DeepMind reports that the project took four weeks. Human experts reviewed the mathematical requirements, and Carbon agents generated the proofs. Earlier work often verified abstract protocol designs or smaller pieces. Checking the existing compiled implementation closes a different gap: a sound design can still acquire bugs when implemented.
The technical report describes about 6.5 million lines of proof resting on roughly 26,000 lines defining security requirements and processor behavior. Lean, a language with a small proof-checking core, checks the reasoning. The reduction in human inspection is substantial, but the trusted requirements still need expert scrutiny.
The most transferable part is how the system keeps the assignment from changing. A mathematically valid proof can answer an easier question than the one approved. AlphaProof Code therefore compares the proved statement, including definitions it depends on, with a separately retained specification. An agent can't silently weaken a definition and count the resulting proof as success.
November 5's exploit pipeline checked whether agents used the assigned vulnerability. AlphaProof Code adds a preserved mathematical target. In an ordinary coding workflow, the analogous arrangement would keep acceptance tests and their fixtures outside the worker's editable files. A passing result would then refer to the original checks. That's an application of the architecture, not a measured productivity result from this study.
The researchers also restored two historical bugs and checked that the specifications could demonstrate the broken versions were insecure. Those failures matter: an acceptance system that approves correct work but also approves known broken work offers little protection.
There are substantial limits. The results concern a particular compiled library and harness, under cryptographic assumptions; they don't certify the whole Signal app. Side channels are outside scope.
The released repository requires Linux for the comparator. Building the default proofs takes several hours and about 15–34 GiB of peak memory, according to the instructions. This is specialist verification, with costs that ordinary tests avoid.
This can move agent engineering toward stronger guarantees where failures are expensive and requirements can be stated precisely. The observable test is whether a changed binary still passes against the preserved requirements, with any amendments separately visible and approved.
The Contrarian Take
Everyone says: A machine-checked security proof means the software is secure.
Here's why that's wrong: The checker establishes the claim people specified, including its assumptions. AlphaProof Code's results concern a particular compiled library and execution environment; they don't certify the whole Signal app. Restoring two historical bugs usefully tests whether the specification rejects known failures, but doesn't show that it covers every attack. The advance is that agents can produce proofs others can check. Deciding which protections the software needs still takes human judgment.
Under the Radar
-
New Relic separates diagnosis from execution. Its November 10 Kubernetes agent is in public preview, using more than 90 diagnostic detectors and generating fixes for engineers to review and execute. That gives an incident investigation operational context without requiring the generated remedy to run automatically. The announcement supplies no controlled reduction in outage duration; detector coverage and safe remediation are different achievements.
-
Commvault's recovery plan reaches beyond the agent. AgentRecover's announced extensions target Microsoft Foundry early access in November and Amazon Bedrock AgentCore in December. The aim is to recover affected data and systems alongside the agent. Restarting an agent won't undo a database write it already made. Cross-system recovery is the consequential promise here, with delivery and recovery effectiveness still to be demonstrated.
Quick Takes
-
Claude brings shared work into the conversation. Sheets and Whiteboards entered beta on paid plans November 10; Dashboards became generally available as Claude Data. Teams can edit formulas and organize ideas with Claude in shared artifacts, carrying context between them. That changes the handoff from generating a file to maintaining shared work. Meetings remains “coming soon,” so the broader launch shouldn't be read as a shipped meeting assistant. (Source)
-
Qwen4-Max opens a bounded preview. The model is available through Token Plan, with a million-token context window and retirement scheduled after November 25 at midnight UTC+8. Its documentation also changes
max_tokensto include thinking and the final answer together. A wrapper that previously reserved that budget for the answer can now truncate useful output. Access is real; a general capability lead hasn't been established. (Source) -
NaiveAI's fast median hides long reasoning runs. Latent Space's 30-question timing experiment found a 3.1-second median for thinking-enabled NaiveAI versus DeepSeek's 10.1 seconds. Their means were nearly equal, around 45 seconds, and answer accuracy wasn't graded. A few long searches consumed the apparent advantage. The experiment publishes code, making response-time distributions a reproducible comparison; a tokens-per-second headline alone misses the waiting time that matters. (Source)
-
Notion identifies the customer consequence of Fireworks' breach. Its advisory says fewer than 1% of workspaces were affected, with Insights chat transcripts in the exposed dataset. For most affected workspaces, transcripts lacked attribution labels and contained no tool-call results. This extends yesterday's provider disclosure with a named downstream product. Removing labels reduces attribution, but secrets inside a conversation can remain sensitive. (Source)
The Thread
Better agents are creating a maintenance obligation for the guarantees around them. Carbon can investigate harder assignments, but a longer investigation still needs a stopping condition. AlphaProof Code makes a stronger promise by preserving a precisely defined target. That shifts some engineering effort from reading every generated line to maintaining the smaller body of requirements that makes those lines trustworthy. The next bottleneck may be deciding which guarantees deserve that investment. A cryptographic library can justify extensive formalization; a disposable dashboard often can't. Cheap generation therefore increases the value of classifying software by the consequences of failure. Otherwise the same elaborate review machinery gets applied everywhere, while the few components that warrant stronger guarantees compete for the attention they need.
Predictions
New predictions:
- I predict: By December 1, Commvault will publish documentation or an access announcement confirming that AgentRecover for Microsoft Foundry has entered early access. (Confidence: medium; Check by: 2026-12-02)
Issue date: November 11, 2026. Generated at 08:36 UTC.
Tomorrow morning in your inbox.
Subscribe for free. 10-minute read, every weekday.