Gemini 4 Carbon Release Date & Access : DeepMind Lean Proofs and Ultra Pricing
What Makes Gemini 4 Carbon Different?
Google DeepMind has officially launched Gemini 4 Carbon, a specialized frontier model engineered specifically for autonomous cybersecurity research, codebase defense, and multi-day software engineering runs. Available today for Google AI Ultra and Enterprise workspaces in Google Antigravity, Carbon represents a fundamental architectural pivot toward verifiable, long-horizon software engineering.
Unlike standard conversational assistants that degrade or lose context state after hours of continuous repository editing, Carbon is designed to run autonomously for days without hallucinating phantom imports or dropping architectural constraints.
The Signal Protocol Formal Proof Breakthrough
To verify whether Carbon could handle production-grade formal verification, Google DeepMind connected Carbon prover agents to AlphaProof Code and tasked them with analyzing Signal's compiled protocol implementation.
Over a four-week period, Carbon generated approximately 6.5 million lines of machine-checked Lean proofs, formally proving key cryptographic security properties against human specifications. A Lean proof engine verified every step of reasoning deterministically, confirming that Carbon constructs verifiable mathematical arguments rather than generating plausible guesses.
Audited Cybersecurity Benchmarks (Safeguards Off)
| Benchmark / Capability | Gemini 4 Argon | Gemini 4 Carbon | Operational Impact |
|---|---|---|---|
| SEC-bench Pro (Flaw Detection) | 64.5% | 88.2% | Finds zero-days & memory safety bugs |
| ExploitGym (Exploitability Test) | ~35.0% | 54.8% | Tests whether flaws are practically exploitable |
| CWE-bench (Patch Verification) | 48.0% | 71.0% | Fixes root causes without breaking test suites |
| Gray Swan Prompt Injection | Baseline | 0.0% Success | Immune to repo prompt injection attacks |
Token Economics & The 95% Cache Discount
Google priced Gemini 4 Carbon introductory access at $2.00 per million input tokens and $10.00 per million output tokens (rising to $4.00 / $20.00 post-introductory). Crucially, cached input tokens receive an aggressive 95% discount, dropping prompt re-reads to just $0.10 per million tokens.
Architectural Verdict & Full Review
👉 Read the complete, unedited technical review, benchmark scorecards, and setup guide on SoftReviewed:
Gemini 4 Carbon review, benchmarks, and pricing
Comments
Post a Comment