Gemini 4 Carbon is now available
Google’s release of the Gemini 4 Carbon AI model is drawing mixed reactions, with many praising its capabilities—especially for writing, research, and complex coding—while others report severe latency and token usage that make it impractical for interactive work. A major point of frustration is Google’s fragmented and opaque access model: Carbon is largely restricted to high-priced Ultra subscriptions and API use, while cheaper or promotional “Pro” and “Plus” tiers often lack current flagship models or have confusing, inconsistent quotas. Commenters contrast this with Anthropic’s and OpenAI’s more straightforward flagship access, arguing that Google’s slow, convoluted rollouts undermine both hype and adoption despite competitive model quality.
Pricing and Subscription Confusion
- Access to Carbon is mostly gated behind higher tiers (AI Ultra, enterprise, or paid API).
- Many users on cheaper or promo “Pro/Plus” plans only see older models (3.6 Flash, 3.5/3.1, etc.) or Argon with tight limits.
- Regional pricing and bundles (students, phone promos, telco deals) create inconsistent access; some promo “Pro” plans lack Argon entirely.
- Several users feel misled by vague marketing like “access to Flash models” or naming specific old models that never get upgraded.
Availability & Rollout
- Carbon appears in multiple places: API (AI Studio), Antigravity, VS Code, Cursor, Gemini app, and enterprise, but not uniformly.
- Rollout is staggered: some see Carbon as default in the app, others on Ultra only, others not at all.
- Several call Google’s launch and rollout strategy confusing or “Cartmanland”-like: hyped but hard to actually use.
Model Quality & Use Cases
- Some report Carbon is “way better than Argon,” especially for writing, research, technical editing, SVG/3D/design, and complex coding pipelines.
- Others find Gemini models weaker than Claude for code reliability, memory over context, and variable consistency.
- One anecdote: Carbon gave stronger editorial feedback on a paper draft than Argon and top Anthropic/OpenAI models.
Performance, Quotas & Token Costs
- Carbon is reported to use more tokens than Argon; users on Pro/Ultra hit quotas quickly, sometimes within a single project.
- Some find Carbon effectively unusable due to very slow time-to-first-token and low throughput, especially for interactive use; others emphasize it’s intended for long-running tasks.
- There is disagreement on cost-effectiveness: list prices may be lower than competitors, but per-task cost can be higher if the model loops or is slow.
Tooling, Harnesses & Features
- Antigravity is the main recommended harness; features like
/boost,/teamwork, and new/swarmfor coordinating subagents are discussed. - Opinions on Google’s harnesses differ: some prefer competitor IDEs (e.g., Claude Code), others like Antigravity, especially with advanced modes.
- Users share small tooling around account switching and note Carbon’s pelican open-world demo as a showcase of multi-agent + 3D capabilities.
Comparisons with Other Labs
- Many compare Carbon/Argon to Anthropic’s Fable/Opus and DeepSeek:
- Some say Anthropic remains ahead overall and is easier to access via subscriptions.
- Others say Gemini is underrated, particularly for multimodal work and writing.
- On hallucinations, experiences conflict: some see Gemini as worse; others cite benchmarks and heavy real-world use claiming it’s state-of-the-art in factuality.
Product Strategy, Naming & Miscellaneous
- Several argue Google is squandering technical gains with slow, opaque consumer offerings and over-protective quotas, ceding mindshare to rivals.
- Others say gradual, tiered rollout is normal business practice and likely profit-driven.
- Naming (Argon vs Carbon, clashes with “Carbon” the language) and the long-running “pelican” meme are mocked.
- Some users report the Carbon blog page crashing Chrome/Brave and general Gemini lagginess.