DVx Signal/Builder/Issue 012

Gemini 4 Hydrogen, Copilot retirements, Cursor's breach and Inkling weights

A new Gemini default, six Copilot models retiring, and a breach that started in agent traces.

DVx Research4 min read

Gemini 4 Hydrogen replaces 3.8 Flash, with breaking changes

November 12 · generally available · checked November 16Official source ↗
Per 1M tokens3.8 FlashHydrogen
Input$0.75$1.50
Output$3.75$7.50
From Jan 1$1.50 / $7.50No dated change
Google's Gemini API pricing for the paid tier. Hydrogen is also on the free tier. View official source ↗
What changed

Google released Gemini 4 Hydrogen on November 12, a fast model for agent and coding work. Its paid rate is double 3.8 Flash's today, but 3.8 Flash rises to the same price on January 1.

Why it matters

The migration guide lists breaking changes: temperature, top_p, and top_k are gone, explicit caching isn't supported, the default thinking level moves to high, and input and output share one 1M-token window.

What to do

Move one 3.8 Flash workload in December, before both models cost the same. Remove sampling parameters and explicit cache calls, then compare output and cost at the new default thinking level.

GitHub retires six Copilot models on December 14

November 13 · retires December 14 · checked November 16Official source ↗
Copilot model picker with Gemini 4 Hydrogen selected

GitHub's release image shows Gemini 4 Hydrogen, the replacement for Gemini 3.8 Flash, in the Copilot model picker. View official source ↗

What changed

GitHub will retire Claude Fable 5, Claude Haiku 4.5, Gemini 3.8 Flash, Gemini 4 Argon, Grok 4.6, and Kimi K3 across all Copilot experiences. Each has a named replacement except Kimi K3.

Why it matters

Replacements turn on automatically for Business and Enterprise unless an admin turned off the global default. Some cost more at provider list price: Hydrogen is double 3.8 Flash's current rate, and Carbon doubles on December 1.

What to do

Find workflows, agents, and SDK calls pinned to a retiring model. Switch them before December 14, and check usage-based billing for any team that defaults to Gemini models.

Cursor's breach started with a vendor key and secrets in agent traces

November 13 · incident update · checked November 16Official source ↗
DataExposed?
Stored agent tracesSome
Credentials in tracesSome
Privacy Mode dataNo
Production systemsNo
Customer reposNo
Cursor's incident update. Exposed traces were mostly from 2025 sessions outside Privacy Mode; all scope statements are Cursor's. View official source ↗
What changed

Cursor says an attacker used a key Cursor had issued to its inference vendor, Fireworks, which Fireworks exposed. It opened a shared bucket of stored agent traces, and credentials left in those traces led into Cursor's internal tools.

Why it matters

Agent traces record what the agent saw, including tokens and keys pasted into a session. Stored without redaction, they work like a credential store that every vendor with bucket access can reach.

What to do

If you're a Cursor admin, check whether you received a key list and rotate it. For your own agents, redact secrets before storing traces and review which vendors can read where you keep them.

Inkling 2 Small's weights are out under Apache 2.0

November 12 · open weights · checked November 16Official source ↗
SpecInkling 2 Small
Parameters594B total, 17B active
InputText and images
LicenseApache 2.0
Thinking Machines' release notes and the Hugging Face model card. View official source ↗
What changed

Thinking Machines published the weights for Inkling 2 Small on Hugging Face on November 12. It launched on November 5 as a hosted model on Thinking Machines' Tinker API.

Why it matters

Apache 2.0 has no revenue threshold or hosted-service clause, unlike the custom licenses on several recent open-weight models. The model runs on vLLM, SGLang, Unsloth, and TokenSpeed.

What to do

If you self-host, test it on one task where licensing has blocked other models. Measure quality and serving cost against your current API model before committing hardware.

New AgentCore runtimes now start in under 2 seconds by default

November 12 · new default · checked November 16Official source ↗
RuntimeV1V2
Start time29.7 sUnder 2 s
Session sizeNot stated8 vCPU, 32 GB
Idle agentsNot statedScale to zero
AWS's announcement. Start times are AWS's P75 figures for a 2 GB image; V2 is the next-generation runtime. View official source ↗
What changed

New runtimes in Amazon Bedrock AgentCore use the next-generation runtime by default in 19 regions. Each session restores from a snapshot of the initialized agent and still runs in its own Firecracker microVM.

Why it matters

Existing runtimes stay on V1 until you set platformVersion to V2. AWS says multi-turn agents use up to 68% less billed memory per session, but points to its pricing page for rates.

What to do

Move one existing runtime to V2 in a test account. Check that state restored from the snapshot behaves correctly, then compare start time and billed memory before moving the rest.

Also since the last issue

Useful updates

Shorter disclosure deadlines

Anthropic will give vendors 30 days to fix high and critical vulnerabilities it reports from January 1, and 45 days for others. The 7-day window for actively exploited bugs is unchanged.

EU impact assessment template

The European Commission published a draft template for AI Act fundamental rights impact assessments on November 16. Deployers can offer to test it by December 4; the obligation applies from December 2, 2027.

One workflow to try

Pin models before December 14

Search your Copilot settings, agent configs, and SDK code for the six retiring model names. Switch each to its replacement, run one known task, and note any change in output or cost.

Term of the week

Agent trace

An agent trace is the stored record of a session: prompts, tool calls, file contents, and outputs. It can capture any secret the agent saw. Treat stored traces like logs that may hold credentials, and redact them before keeping them.