Google Opens Gemini 4 Carbon to All Buyers, Undercutting Anthropic on API Price

What Google Just Opened Up With Gemini 4 Carbon

Google has made Gemini 4 Carbon, its most capable frontier AI model, generally available through the paid Gemini API, Gemini for business, and consumer products including the Gemini app, AI Mode in Search, and Antigravity. The wider release came on Nov. 10, nearly two weeks after Google announced Carbon on Oct. 28 with what were then industry-leading benchmarks. During the interim, access was limited to selected cybersecurity defenders through Google DeepMind's gated Fairwind program, where a specialized cyber version with fewer guardrails remains available.

The commercial pitch is straightforward: Carbon starts at $2 per million uncached input tokens and $10 per million output tokens, which is one-quarter of Claude Fable 5.2's standard rates through Nov. 30. After that date, Google's prices double to $4 and $20, still half of Fable's standard rates. Google also offers a 95% cached-input discount, implying $0.10 per million cached input tokens during the promotion and $0.20 afterward. For consumers, Carbon is rolling out to Google AI Ultra subscribers at $99.99 per month for Ultra 5x and $199.99 per month for Ultra 20x.

The timing is notable: Anthropic launched Claude Fable 5.2 and Mythos 5.2 just one day before Carbon's wider release, and Fable has since taken the lead on Artificial Analysis's general intelligence index at 65, with Carbon at 60 and Claude Opus 5.5 at 59. Google's model still leads on some agent-focused benchmarks, including Zapier's AutomationBench where Carbon high scores 53.88% at an estimated $2.12 per task versus Fable 5.2 max at 51.45% and $4.64. But on code-review tasks like FrontierCode 2.0, Fable's best run reaches 49.1% versus Carbon's 40.6%.

For enterprise buyers, the more consequential question is not the token price but the total cost of completing a job. Artificial Analysis estimates an average of $7.76 per task for Carbon at high reasoning on its index suite, compared with $26.54 for Fable at maximum effort. However, Fable at high effort costs $9.08 and scores 62, two points above Carbon high. At medium effort, Fable costs $3.80 and scores 56, while Carbon costs $6.59 and scores 58. The practical comparison depends on the least expensive configuration that meets a given quality target, with failures and retries included.

Where Carbon Wins, Where Fable Pulls Ahead — and What the Price Tag Really Buys

Google's Pricing Play Targets Anthropic's Enterprise Base

Google's decision to price Carbon at one-quarter of Fable 5.2's standard rates through Nov. 30 is a direct challenge to Anthropic's enterprise position. The promotional window is short — roughly three weeks from general availability — which suggests Google is prioritizing rapid adoption and workload migration over immediate revenue maximization. The 95% cached-input discount is particularly relevant for teams that repeatedly send the same repository or policy documents, though Google has not published cache-hit rate data, so the actual savings depend on usage patterns that buyers will need to measure themselves.

Benchmark Leadership Is Task-Dependent, Not Absolute

The competitive picture is more nuanced than either company's marketing suggests. Fable 5.2 leads the Artificial Analysis Intelligence Index at 65 with maximum effort, while Carbon scores 60 on high. But on Terminal-Bench 4.0, Carbon high reaches 65.2% versus Fable's best listed run at 61.6%. On FrontierCode 2.0, which grades changes to real repositories against merge standards, Fable leads at 49.1% versus Carbon's 40.6%. Anthropic separately reports 69.9% for Fable on FrontierSWE v2, ahead of Carbon at 61.2%. A team choosing an agent for terminal-heavy diagnostics could reasonably reach a different decision from one trying to get more patches through review.

Unverified Claims Leave Procurement Teams With Work to Do

Google claims first place at less than half the cost of the next-best model on CursorBench 4.0 and Perplexity's WANDR research benchmark, but the launch materials omit the underlying score and cost figures. The public CursorBench table shows Fable 5.2 max leading at 60.0% and $32.79 per task, with no Carbon row in the version reviewed. Anthropic's WANDR table reports 82.1% for Fable 5.2 on its relaxed scoring measure and leaves Carbon blank. These gaps mean Google's claims on those two benchmarks remain unverified by public data, and procurement teams should treat them as marketing until independent results appear.

Cyber Capabilities Follow a Separate, Gated Path

Google's strongest cyber capabilities remain behind the Fairwind program, mirroring Anthropic's approach where Mythos requires approval through cyber or life sciences programs. Google reports 88.2% on SEC-bench Pro for finding vulnerabilities and 54.8% on ExploitGym for demonstrating exploitability, but ran those tests without guardrails. Anthropic reports Mythos 5.2 rediscovered 89.6% of known vulnerabilities in internal testing and scored 76.3% on CyScenarioBench, though the table supplies no Carbon result on either measure. Security teams must first determine which version they can access, then measure how often each model produces a reproducible finding and a patch that passes regression tests.

Open-Weight Alternatives Compete on a Different Requirement

For teams that need to run inference on their own infrastructure, Carbon, Fable, and Astra cannot compete because they are proprietary hosted models with no downloadable weights. Moonshot AI's Kimi K3.1 Pro is the highest-scoring open-weight model on Artificial Analysis's broad index, and Meituan released LongCat-2.5's weights under the MIT License on Nov. 9. France's Mistral offers Large 4 under its own open-weight license. These models compete on a procurement requirement — self-hosting — that the U.S. frontier labs' hosted APIs cannot meet, which matters for organizations with data residency or security constraints.

What This Means for Enterprise AI Stacks

Google's own product design points toward mixing models rather than committing to a single vendor. The /swarm preview in Antigravity lets developers put Carbon on architecture, coordination, or audit work while assigning parallel implementation tasks to lighter models in isolated workspaces. Anthropic describes a similar division of labor for Fable as an advisor or orchestrator, with Opus 5.5 as the starting point for most daily work. Google likewise recommends Argon for most software engineering and knowledge work. For a Carbon pilot, the sensible approach is keeping the existing model as a baseline and giving both systems the same jobs that already consume engineering time — a difficult migration judged on whether it builds, passes tests, and survives review, or an operations workflow judged on whether it makes every required change without touching unrelated records.

How Enterprise AI Buyers Should Evaluate Carbon Before the Promotional Price Ends

The headline token price is not the cost of the job — measure total task cost including retries and human corrections before committing workloads to Carbon.

For enterprise AI buyers and engineering leaders evaluating Carbon against Fable 5.2 or existing models:

• Run a controlled pilot before Nov. 30, when Carbon's introductory pricing of $2/$10 per million tokens doubles to $4/$20. Give both Carbon and your current model the same migration or operations workflow, and track model spend, elapsed time, and reviewer time per completed task.
• Calculate what the same run would cost at December rates. Artificial Analysis estimates $7.76 per task for Carbon at high reasoning versus $26.54 for Fable at maximum effort, but Fable at high effort costs $9.08 and scores two points higher on the index — so the cheapest configuration that meets your quality target is the relevant comparison, not the lowest token price.
• Measure cache-hit rates if you repeatedly send the same repository or policy documents. Google's 95% cached-input discount implies $0.10 per million cached input tokens during the promotion, but actual savings depend on how often requests hit the cache.
• Check whether your cloud contract covers Carbon. Google's rollout materials detail paid Gemini API access through AI Studio and Android Studio but leave Vertex AI availability, regions, and support commitments unspecified — distinctions that matter for teams buying through a particular cloud agreement.
• For security workloads, determine which version you can access first. Carbon for cyber remains gated through Fairwind, and Mythos requires approval through Anthropic's cyber or life sciences programs. Then measure how often each model produces a reproducible finding and a patch that passes your own regression tests.
• If self-hosting is a procurement requirement, Carbon, Fable, and Astra cannot meet it. Evaluate open-weight alternatives such as Moonshot's Kimi K3.1 Pro or Meituan's LongCat-2.5, released under the MIT License on Nov. 9.

Metrics worth watching: Google's published cache-hit rates for Carbon, independent benchmark results on CursorBench 4.0 and WANDR, and whether Google extends the introductory pricing beyond Nov. 30.