Gemini 4 Carbon

Original Article ↗ Hacker News Discussion ↗

Google’s new Gemini 4 Carbon model is being touted as state-of-the-art, with top scores on prominent AI benchmarks and claims of major gains in coding accuracy, safety, and resistance to prompt injection. Commenters are split between excitement—especially given its aggressive pricing and potential to rival Anthropic’s Opus and OpenAI’s latest models—and skepticism about benchmark cherry-picking, Artificial Analysis’ shifting metrics, and whether gains translate to real-world agentic coding and long-context work. Many also criticize Google’s pattern of “paper launches” with slow, confusing rollouts, fragmented subscriptions, and opaque naming, which they see as undermining otherwise strong underlying models.

Model Capabilities & Benchmarks

  • Carbon is presented as Google’s new frontier Gemini 4 model, seemingly a further-trained checkpoint of Argon with the same API pricing.
  • Claims: #1 overall on Artificial Analysis (AA) Index v5, strong coding, better long-context performance, improved safety and less “fabrication.”
  • Specific technical claims that impressed people:
    • 0% “cheating” on Terminal-Bench 4.0 with cheating attempts down to 0.4%.
    • 0% success on a prompt-injection benchmark used in the blog.
  • Some commenters highlight remaining weaknesses: on coding-focused benchmarks like FrontierSWE it still trails Opus 5.5 and Astra.

Artificial Analysis Index & Benchmark Skepticism

  • AA updated its index (v5) the day before; Carbon then appears as #1.
  • Several commenters see this timing as coordinated and “sus.”
  • Concerns:
    • Changing weights/methodology whenever certain models (especially open-weight ones) approach the top.
    • Inclusion of private, non-verifiable benchmarks.
    • Scores shifting drastically across index versions, weakening trust.
  • Others still regard AA as the best available aggregate signal, while acknowledging its limitations.

Release Strategy and Availability

  • Widespread frustration that Google announces models (“coming soon”) before broad access:
    • Argon took ~weeks from announcement to general usability; some paid users still can’t access it in all surfaces.
    • Carbon is benchmarked and marketed but not usable yet except for a narrow “cyber defender” preview.
  • Many see this as “vaporware” marketing and earnings-call theater; others argue big labs must now wait on government/“cyber” vetting.

Naming, Versioning, and Product Confusion

  • Strong consensus that Google’s naming is confusing:
    • Argon vs Carbon gives no obvious tier ordering; periodic table order conflicts with perceived capability.
    • Coexists with 3.1 Pro, 3.6/3.8 Flash, Flash-Lite, etc.
  • Multiple people wish for simple numeric versions (e.g., “Gemini 4.1”) or a clear fixed tiering like “Lite / Flash / Pro / Ultra.”

Pricing, Plans, and Usage Limits

  • API pricing ($2/$10 promo → $4/$20 per 1M tokens) is seen as aggressive and attractive, especially if performance matches claims.
  • But consumer subscriptions are viewed as a mess:
    • Plus tier stuck on old 3.1 Pro and 3.6 Flash, no Argon/Carbon; some feel “scammed.”
    • Pro/Ultra tiers differ by account type (personal vs Workspace); some Workspace users can’t even buy Ultra.
    • Argon usage pools are tiny; a few prompts can exhaust weekly quota, pushing users back to Flash-Lite.

Harnesses, Developer UX, and Coding Performance

  • Antigravity (Google’s agentic harness) draws heavy criticism:
    • Hard to onboard, confusing quotas, unstable features, poor defaults; some call it the “worst coding harness” they’ve used.
    • Others say Argon/Carbon perform much better inside well-tuned harnesses (Antigravity with clean configs, or custom setups).
  • Real-world coding impressions are mixed:
    • Some say Argon is “a clear tier below” Opus/Sol, struggles as a local coding agent, and underperforms in one open-source agent duel framework.
    • Others report Argon is excellent for planning, code review, refactoring, and cyber work; “thorough, smart, adheres to goals,” and sometimes preferred over Opus 5.5.
    • Several note Argon’s style: very “goody two shoes,” frequently asking for confirmation, and sometimes worse at executing than planning.

Security, Cyber Rollout, and Gatekeeping

  • Carbon is first going to a “Fairwind” program for governments and critical infrastructure “defenders” before general release.
  • One camp sees this as responsible: frontier cyber capabilities are dangerous and should be gated.
  • Another sees it as counterproductive gatekeeping:
    • Belief that serious attackers already have strong models/weights.
    • Fear that restricting powerful “cyber” models to big vendors and governments leaves open-source and small org defenders at a disadvantage.

Broader Market & Competitive Landscape

  • Many expect Anthropic or OpenAI to retake the #1 spot within days or weeks; pace of frontier releases is seen as “insane.”
  • Some argue that coding evals (FrontierSWE, SWE-Sweep, etc.) matter more economically than AA’s composite “intelligence” index.
  • Recurrent theme: Google has strong research and hardware but is held back by bureaucracy, fragmented products, poor UX, and confusing subscriptions.
  • A few see Google and Meta as long-term winners with distribution to “normies”; others think poor UX and trust issues (account bans, deprecations) will limit developer adoption.